Datasets:
Citation_context stringlengths 42 5.8k | Cited_content stringlengths 87 10.5k | Label stringclasses 3 values | Twist_category stringclasses 6 values |
|---|---|---|---|
A formal definition of non-differentiability has not been provided by the authors. However, authors in this context use this idea of differentiability to explain that most automated metrics cannot provide quality estimates after a language model generates one token or a parietal output - similar to humans, they can provide quality estimates only after a language model generates a full sequence. This non-differentiability is pointed out as a problem by the authors for the new metrics (i. things like BLUERT or BERTScore) - it is unclear from this paper alone if previous vanilla metrics (i. BLEU) were per token differentiable or not. The authors are looking for ways to use these non-differentiable scores in RL frameworks, potentially as a reward function. | The ultimate aim of language technology is to interact with humans.However, most language models are trained without direct signals of human preference,with supervised target strings serving as (a sometimes crude) proxy.One option to incorporate user feedback is via human-in-the-loop, i.e., a user would be expected to provide feedback for each sample online as the model trains, but this degree of dense supervision is often prohibitive and inefficient.Automated metrics offer a promising compromise: learned models of human preference like BERTScore (Zhang et al., 2019), BLEURT (Sellam et al., 2020), summarization preferences (Wu et al., 2021) have significantly improved correlation with human judgment compared to earlier metrics (BLEU, METEOR, etc.), and are cheap to evaluate. But — these functions are usually not per-token differentiable: like humans, metrics can only offer quality estimates for full generations.Reinforcement Learning (RL) offers a natural path forward for optimizing non-differentiable, scalar objectives for LM-based generation when it is cast as a sequential decision-making problem. However, Goodhart’s Law222Strathern (1997) paraphrases: When a measure becomes a target, it ceases to be a good measure.looms: particularly in the case of imperfect metrics that use neural networks, it is easy to find nonsense samples that achieve high-quality estimates. Recent works have shown promising results in aligning LMs to human preferences via RL by constraining preference-based rewards to incorporate notions of fluency (Wu et al., 2021; Ouyang et al., 2022) but progress in this line of work is heavily hindered by a lack of open-source benchmarks and algorithmic implementations—resulting in perception that RL is a challenging paradigm for NLP (Choshen et al., 2020; Kreutzer et al., 2021). | Correct | null |
A formal definition of non-differentiability has not been provided by the authors. However, authors in this context use this idea of differentiability to explain that most automated metrics can provide quality estimates after a language model generates one token or a parietal output - similar to humans, they can provide quality estimates only after a language model generates a full sequence. This non-differentiability is pointed out as a problem by the authors for the new metrics (i. things like BLUERT or BERTScore) - it is unclear from this paper alone if previous vanilla metrics (i. BLEU) were per token differentiable or not. The authors are looking for ways to use these non-differentiable scores in RL frameworks, potentially as a reward function. | The ultimate aim of language technology is to interact with humans.However, most language models are trained without direct signals of human preference,with supervised target strings serving as (a sometimes crude) proxy.One option to incorporate user feedback is via human-in-the-loop, i.e., a user would be expected to provide feedback for each sample online as the model trains, but this degree of dense supervision is often prohibitive and inefficient.Automated metrics offer a promising compromise: learned models of human preference like BERTScore (Zhang et al., 2019), BLEURT (Sellam et al., 2020), summarization preferences (Wu et al., 2021) have significantly improved correlation with human judgment compared to earlier metrics (BLEU, METEOR, etc.), and are cheap to evaluate. But — these functions are usually not per-token differentiable: like humans, metrics can only offer quality estimates for full generations.Reinforcement Learning (RL) offers a natural path forward for optimizing non-differentiable, scalar objectives for LM-based generation when it is cast as a sequential decision-making problem. However, Goodhart’s Law222Strathern (1997) paraphrases: When a measure becomes a target, it ceases to be a good measure.looms: particularly in the case of imperfect metrics that use neural networks, it is easy to find nonsense samples that achieve high-quality estimates. Recent works have shown promising results in aligning LMs to human preferences via RL by constraining preference-based rewards to incorporate notions of fluency (Wu et al., 2021; Ouyang et al., 2022) but progress in this line of work is heavily hindered by a lack of open-source benchmarks and algorithmic implementations—resulting in perception that RL is a challenging paradigm for NLP (Choshen et al., 2020; Kreutzer et al., 2021). | Incorrect | Opposite |
The action space for language modeling is equal to the vocabulary set of language models. Since the vocabularies are very large (i. tens of thousands of possible tokens), the action space is also very large. In general, locomotion in the real world can be condensed to three quantities - moving across X, Y or Z axes, or a linear combination thereof. The authors mention that typical RL problems have an action space that is an order of magnitude smaller, but do not specifically explain how the action spaces for typical problems is modeled or constructed. | Language generation action spaces are orders of magnitude larger than what most discrete action space RL algorithms are designed for (Ranzato et al., 2015; Ammanabrolu, 2021), e.g., GPT-2/3 and T5 have a vocabulary size of 50K and 32K respectively.We hypothesize that the size of the action space is a core cause of instability when training LMs with existing RL methods.To address this issue, we introduce NLPO (Natural Language Policy Optimization), which is inspired by work on action elimination/invalid-action masking (Zahavy et al., 2018; Huang & Ontañón, 2020; Ammanabrolu & Hausknecht, 2020). NLPO, a parameterized-masked extension of PPO, learns to mask out less relevant tokens in-context as it trains. NLPO accomplishes this via top-p sampling, which restricts tokens to the smallest possible set whose cumulative probability is greater than the probability parameter p (Holtzman et al., 2018).RL for Large Action Spaces. MIXER (Ranzato et al., 2015) combined ideas from schedule sampling and REINFORCE (Williams, 1992).Bahdanau et al. (2016) proposed an actor-critic algorithm to address the variance/large action space problems when using REINFORCE for language generation; follow-up works such asKG-A2C (Ammanabrolu & Hausknecht, 2020), TrufLL (Martin et al., 2022), AE-DQN (Zahavy et al., 2018), and GALAD (Ammanabrolu et al., 2022) addressed similar issues by attempting to eliminate and reduce the action space during exploration.Each environment is an NLP task: we are given a supervised dataset \mathcal{D}=\{({\bm{x}}^{i},{\bm{y}}^{i})\}_{i=1}^{N} of N examples, where {\bm{x}}\in\mathcal{X} is an language input and {\bm{y}}\in\mathcal{Y} is the target string. Generation can be viewed as a Markov Decision Process (MDP) \langle\mathcal{S},\mathcal{A},\mathcal{R},P,\gamma,T\rangle using a finite vocabulary \mathcal{V}.Each episode in the MDP begins by sampling a datapoint ({\bm{x}},{\bm{y}}) from our dataset and ends when the current time step t exceeds the horizon T or an end of sentence (EOS) token is generated.The input {\bm{x}}=(x_{0},\cdots,x_{m}) is a task-specific prompt that is used as our initial state \bm{s}_{0}=(x_{0},\cdots,x_{m}), where \bm{s}_{0}\in\mathcal{S} and \mathcal{S} is the state space with x_{m}\in\mathcal{V}. An action in the environment a_{t}\in\mathcal{A} consists of a token from our vocabulary \mathcal{V}. The transition function P:\mathcal{S}\times\mathcal{A}\rightarrow\mathcal{S} deterministically appends an action a_{t} to the end of the state \bm{s}_{t-1}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{t-1}). This continues until the end of the horizon t\leq T and we obtain a state \bm{s}_{T}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{T}).At the end of an episode a reward \mathcal{R}:\mathcal{S}\times\mathcal{A}\times\mathcal{Y}\rightarrow\mathbb{R}^{1} that depends on the (\bm{s}_{T},{\bm{y}}) (e.g., an automated metric like PARENT Dhingra et al. (2019)) is emitted. RL4LMs provides an OpenAI gym (Brockman et al., 2016) styleAPI for an RL environmentthat simulates this LM-Based MDP formulation.Abstracting the details of the MDP environment structure allows for new tasks to be added quickly with compatibility across all implemented algorithms. | Correct | null |
The action space for language modeling is equal to the vocabulary set of language models. Since the vocabularies are very large (i. tens of thousands of possible tokens), the action space is also very large. In general, locomotion in the real world can be condensed to five quantities - moving across X, Y or Z axes, or a linear combination thereof. The authors mention that typical RL problems have an action space that is an order of magnitude smaller, but do not specifically explain how the action spaces for typical problems is modeled or constructed. | Language generation action spaces are orders of magnitude larger than what most discrete action space RL algorithms are designed for (Ranzato et al., 2015; Ammanabrolu, 2021), e.g., GPT-2/3 and T5 have a vocabulary size of 50K and 32K respectively.We hypothesize that the size of the action space is a core cause of instability when training LMs with existing RL methods.To address this issue, we introduce NLPO (Natural Language Policy Optimization), which is inspired by work on action elimination/invalid-action masking (Zahavy et al., 2018; Huang & Ontañón, 2020; Ammanabrolu & Hausknecht, 2020). NLPO, a parameterized-masked extension of PPO, learns to mask out less relevant tokens in-context as it trains. NLPO accomplishes this via top-p sampling, which restricts tokens to the smallest possible set whose cumulative probability is greater than the probability parameter p (Holtzman et al., 2018).RL for Large Action Spaces. MIXER (Ranzato et al., 2015) combined ideas from schedule sampling and REINFORCE (Williams, 1992).Bahdanau et al. (2016) proposed an actor-critic algorithm to address the variance/large action space problems when using REINFORCE for language generation; follow-up works such asKG-A2C (Ammanabrolu & Hausknecht, 2020), TrufLL (Martin et al., 2022), AE-DQN (Zahavy et al., 2018), and GALAD (Ammanabrolu et al., 2022) addressed similar issues by attempting to eliminate and reduce the action space during exploration.Each environment is an NLP task: we are given a supervised dataset \mathcal{D}=\{({\bm{x}}^{i},{\bm{y}}^{i})\}_{i=1}^{N} of N examples, where {\bm{x}}\in\mathcal{X} is an language input and {\bm{y}}\in\mathcal{Y} is the target string. Generation can be viewed as a Markov Decision Process (MDP) \langle\mathcal{S},\mathcal{A},\mathcal{R},P,\gamma,T\rangle using a finite vocabulary \mathcal{V}.Each episode in the MDP begins by sampling a datapoint ({\bm{x}},{\bm{y}}) from our dataset and ends when the current time step t exceeds the horizon T or an end of sentence (EOS) token is generated.The input {\bm{x}}=(x_{0},\cdots,x_{m}) is a task-specific prompt that is used as our initial state \bm{s}_{0}=(x_{0},\cdots,x_{m}), where \bm{s}_{0}\in\mathcal{S} and \mathcal{S} is the state space with x_{m}\in\mathcal{V}. An action in the environment a_{t}\in\mathcal{A} consists of a token from our vocabulary \mathcal{V}. The transition function P:\mathcal{S}\times\mathcal{A}\rightarrow\mathcal{S} deterministically appends an action a_{t} to the end of the state \bm{s}_{t-1}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{t-1}). This continues until the end of the horizon t\leq T and we obtain a state \bm{s}_{T}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{T}).At the end of an episode a reward \mathcal{R}:\mathcal{S}\times\mathcal{A}\times\mathcal{Y}\rightarrow\mathbb{R}^{1} that depends on the (\bm{s}_{T},{\bm{y}}) (e.g., an automated metric like PARENT Dhingra et al. (2019)) is emitted. RL4LMs provides an OpenAI gym (Brockman et al., 2016) styleAPI for an RL environmentthat simulates this LM-Based MDP formulation.Abstracting the details of the MDP environment structure allows for new tasks to be added quickly with compatibility across all implemented algorithms. | Incorrect | Change Number |
Actor critic models are a class of reinforcement learning algorithms. The paper does not contain information on Q-learning or how actor critic models are different from Q-learning models. | RL4LMs supports fine-tuning and training LMs from scratch via on-policy actor-critic algorithms on language environments.Formally, this class of algorithms allows us to train a parameterized control policy defined as \pi_{\theta}:\mathcal{S}\rightarrow\mathcal{A}, a function that attempts to select an action in a given state so as to maximize long term discounted rewards over a trajectory \mathbb{E}_{\pi}[\sum_{t=0}^{T}\gamma^{t}\mathcal{R}(\bm{s}_{t},a_{t})].Our benchmark experiments focus on fine-tuning a pre-trained LM denoted as \pi_{0} from which we initial our agent’s policy \pi_{\theta}=\pi_{0}.Similarly, the value network V_{\phi} used to estimate the value function is also initialized from \pi_{0} except for the final layer which is randomly initialized to output a single scalar value.As with other deep RL actor-critic algorithms, we define our value and Q-value functions as V_{t}^{\pi}=\mathbb{E}_{a_{t}\sim\pi}[\sum_{\tau=t}^{T}\gamma R(\bm{s}_{\tau},a_{\tau},{\bm{y}})],Q_{t}^{\pi}(\bm{s}_{t},a_{t})=R(\bm{s}_{t},a_{t},{\bm{y}})+\gamma\mathbb{E}_{s_{t+1}\sim P}[V_{t+1}^{\pi}(\bm{s}_{t+1})]leading to a definition of our advantage function as A_{t}^{\pi}(\bm{s},a)=Q_{t}^{\pi}(\bm{s},a)-V_{t}^{\pi}.To increase training stability, advantage is appoximated using Generalized Advantage Estimation (Schulman et al., 2015). | Correct | null |
Actor critic models are a class of reinforcement learning algorithms. The paper contains information on Q-learning and how actor critic models are different from Q-learning models. | RL4LMs supports fine-tuning and training LMs from scratch via on-policy actor-critic algorithms on language environments.Formally, this class of algorithms allows us to train a parameterized control policy defined as \pi_{\theta}:\mathcal{S}\rightarrow\mathcal{A}, a function that attempts to select an action in a given state so as to maximize long term discounted rewards over a trajectory \mathbb{E}_{\pi}[\sum_{t=0}^{T}\gamma^{t}\mathcal{R}(\bm{s}_{t},a_{t})].Our benchmark experiments focus on fine-tuning a pre-trained LM denoted as \pi_{0} from which we initial our agent’s policy \pi_{\theta}=\pi_{0}.Similarly, the value network V_{\phi} used to estimate the value function is also initialized from \pi_{0} except for the final layer which is randomly initialized to output a single scalar value.As with other deep RL actor-critic algorithms, we define our value and Q-value functions as V_{t}^{\pi}=\mathbb{E}_{a_{t}\sim\pi}[\sum_{\tau=t}^{T}\gamma R(\bm{s}_{\tau},a_{\tau},{\bm{y}})],Q_{t}^{\pi}(\bm{s}_{t},a_{t})=R(\bm{s}_{t},a_{t},{\bm{y}})+\gamma\mathbb{E}_{s_{t+1}\sim P}[V_{t+1}^{\pi}(\bm{s}_{t+1})]leading to a definition of our advantage function as A_{t}^{\pi}(\bm{s},a)=Q_{t}^{\pi}(\bm{s},a)-V_{t}^{\pi}.To increase training stability, advantage is appoximated using Generalized Advantage Estimation (Schulman et al., 2015). | Incorrect | Opposite |
The authors hypothesize that their dynamic masking function helps because it adds a new constraint that the RL algorithm has to abide by. Additionally, since this is a dynamic function, being updated oftenly (every mu steps), it is likely that the masking function ensures that the selected top-p tokens are more relevant to the current state the RL algorithm needs to analyse and decide on. | Specifically, NLPOmaintains a masking policy \pi_{\psi}: the masking policy is a copy of the current policy (\pi_{\theta}), but is updated only every \mu steps.A parameterized-invalid-mask is created from \pi_{\psi} by first selecting the top-p tokens from the vocabulary,444\pi_{\psi} could be trained with alternate sampling techniques like top-k or beam search (or even hard-coded via rules by domain experts), though we find top-p sampling to be most effective in practice. and then applying an invalid-mask to the remaining tokens—i.e. setting their probabilities to zero when sampling actions from \pi_{\theta} during training;this periodic updating policy \pi_{\psi} is inspired by off-policy Q-learning algorithms (Andrychowicz et al., 2017),providing the policy \pi_{\theta} with an additional constraint that balances between the benefits of containing more task relevant information than the KL penalty derived from \pi_{0} and the risk of reward hacking.We provide pseudocode in Algorithm 1 (green portions highlight the differences with PPO).PPO vs. NLPO.Figure 2 shows that NLPO generally outperforms PPO and supervised, especially when applied after supervised training.We hypothesize that the primary reason for NLPO’s improved performance and stability is because the masking policy provides an additional constraint for the current policy.This constraint is not based on the initial untuned policy like the KL penalty but of the policy from \mu iterations ago and likely contains more task-relevant information learned during RL training.Table 3 (and Appendix Table 8) shows how performance increases up to a point and then decreases as p in top-p sampling is increased for the masking policy, relaxing the constraint by eliminating less tokens at each step, implying that there is a balance to be found in how much the model should be constrained during RL training. | Correct | null |
The authors hypothesize that their dynamic masking function doesn't help because it adds a new constraint that the RL algorithm has to abide by. Additionally, since this is a dynamic function, being updated oftenly (every mu steps), it is likely that the masking function ensures that the selected top-p tokens are more relevant to the current state the RL algorithm needs to analyse and decide on. | Specifically, NLPOmaintains a masking policy \pi_{\psi}: the masking policy is a copy of the current policy (\pi_{\theta}), but is updated only every \mu steps.A parameterized-invalid-mask is created from \pi_{\psi} by first selecting the top-p tokens from the vocabulary,444\pi_{\psi} could be trained with alternate sampling techniques like top-k or beam search (or even hard-coded via rules by domain experts), though we find top-p sampling to be most effective in practice. and then applying an invalid-mask to the remaining tokens—i.e. setting their probabilities to zero when sampling actions from \pi_{\theta} during training;this periodic updating policy \pi_{\psi} is inspired by off-policy Q-learning algorithms (Andrychowicz et al., 2017),providing the policy \pi_{\theta} with an additional constraint that balances between the benefits of containing more task relevant information than the KL penalty derived from \pi_{0} and the risk of reward hacking.We provide pseudocode in Algorithm 1 (green portions highlight the differences with PPO).PPO vs. NLPO.Figure 2 shows that NLPO generally outperforms PPO and supervised, especially when applied after supervised training.We hypothesize that the primary reason for NLPO’s improved performance and stability is because the masking policy provides an additional constraint for the current policy.This constraint is not based on the initial untuned policy like the KL penalty but of the policy from \mu iterations ago and likely contains more task-relevant information learned during RL training.Table 3 (and Appendix Table 8) shows how performance increases up to a point and then decreases as p in top-p sampling is increased for the masking policy, relaxing the constraint by eliminating less tokens at each step, implying that there is a balance to be found in how much the model should be constrained during RL training. | Incorrect | Opposite |
The authors mention that they update the masking function every "mu" steps, but the main text of the paper itself does not appear to contain the exact value of mu itself - there is a possibility that the author's model could work with mu=1 instead (i. update every step) instead of updating it every couple of steps (i. mu > 1), though the authors do not explain in this paper if this were done. However, the authors mention that one reason why NLPO outperforms PPO is probably because of this dynamic masking function that is updated occasionally - this indicates to us that mu is probably not an unbounded or very large number (if mu were very large, tending to infinity, the function would no longer be dynamic). | Language generation action spaces are orders of magnitude larger than what most discrete action space RL algorithms are designed for (Ranzato et al., 2015; Ammanabrolu, 2021), e.g., GPT-2/3 and T5 have a vocabulary size of 50K and 32K respectively.We hypothesize that the size of the action space is a core cause of instability when training LMs with existing RL methods.To address this issue, we introduce NLPO (Natural Language Policy Optimization), which is inspired by work on action elimination/invalid-action masking (Zahavy et al., 2018; Huang & Ontañón, 2020; Ammanabrolu & Hausknecht, 2020). NLPO, a parameterized-masked extension of PPO, learns to mask out less relevant tokens in-context as it trains. NLPO accomplishes this via top-p sampling, which restricts tokens to the smallest possible set whose cumulative probability is greater than the probability parameter p (Holtzman et al., 2018).Specifically, NLPOmaintains a masking policy \pi_{\psi}: the masking policy is a copy of the current policy (\pi_{\theta}), but is updated only every \mu steps.A parameterized-invalid-mask is created from \pi_{\psi} by first selecting the top-p tokens from the vocabulary,444\pi_{\psi} could be trained with alternate sampling techniques like top-k or beam search (or even hard-coded via rules by domain experts), though we find top-p sampling to be most effective in practice. and then applying an invalid-mask to the remaining tokens—i.e. setting their probabilities to zero when sampling actions from \pi_{\theta} during training;this periodic updating policy \pi_{\psi} is inspired by off-policy Q-learning algorithms (Andrychowicz et al., 2017),providing the policy \pi_{\theta} with an additional constraint that balances between the benefits of containing more task relevant information than the KL penalty derived from \pi_{0} and the risk of reward hacking.We provide pseudocode in Algorithm 1 (green portions highlight the differences with PPO).PPO vs. NLPO.Figure 2 shows that NLPO generally outperforms PPO and supervised, especially when applied after supervised training.We hypothesize that the primary reason for NLPO’s improved performance and stability is because the masking policy provides an additional constraint for the current policy.This constraint is not based on the initial untuned policy like the KL penalty but of the policy from \mu iterations ago and likely contains more task-relevant information learned during RL training.Table 3 (and Appendix Table 8) shows how performance increases up to a point and then decreases as p in top-p sampling is increased for the masking policy, relaxing the constraint by eliminating less tokens at each step, implying that there is a balance to be found in how much the model should be constrained during RL training. | Correct | null |
The authors mention that they update the masking function every "mu" steps, but the main text of the paper itself does not appear to contain the exact value of mu itself - there is a possibility that the author's model could work with mu=5 instead (i. update every step) instead of updating it every couple of steps (i. mu > 5), though the authors do not explain in this paper if this were done. However, the authors mention that one reason why NLPO outperforms PPO is probably because of this dynamic masking function that is updated occasionally - this indicates to us that mu is probably not an unbounded or very large number (if mu were very large, tending to infinity, the function would no longer be dynamic). | Language generation action spaces are orders of magnitude larger than what most discrete action space RL algorithms are designed for (Ranzato et al., 2015; Ammanabrolu, 2021), e.g., GPT-2/3 and T5 have a vocabulary size of 50K and 32K respectively.We hypothesize that the size of the action space is a core cause of instability when training LMs with existing RL methods.To address this issue, we introduce NLPO (Natural Language Policy Optimization), which is inspired by work on action elimination/invalid-action masking (Zahavy et al., 2018; Huang & Ontañón, 2020; Ammanabrolu & Hausknecht, 2020). NLPO, a parameterized-masked extension of PPO, learns to mask out less relevant tokens in-context as it trains. NLPO accomplishes this via top-p sampling, which restricts tokens to the smallest possible set whose cumulative probability is greater than the probability parameter p (Holtzman et al., 2018).Specifically, NLPOmaintains a masking policy \pi_{\psi}: the masking policy is a copy of the current policy (\pi_{\theta}), but is updated only every \mu steps.A parameterized-invalid-mask is created from \pi_{\psi} by first selecting the top-p tokens from the vocabulary,444\pi_{\psi} could be trained with alternate sampling techniques like top-k or beam search (or even hard-coded via rules by domain experts), though we find top-p sampling to be most effective in practice. and then applying an invalid-mask to the remaining tokens—i.e. setting their probabilities to zero when sampling actions from \pi_{\theta} during training;this periodic updating policy \pi_{\psi} is inspired by off-policy Q-learning algorithms (Andrychowicz et al., 2017),providing the policy \pi_{\theta} with an additional constraint that balances between the benefits of containing more task relevant information than the KL penalty derived from \pi_{0} and the risk of reward hacking.We provide pseudocode in Algorithm 1 (green portions highlight the differences with PPO).PPO vs. NLPO.Figure 2 shows that NLPO generally outperforms PPO and supervised, especially when applied after supervised training.We hypothesize that the primary reason for NLPO’s improved performance and stability is because the masking policy provides an additional constraint for the current policy.This constraint is not based on the initial untuned policy like the KL penalty but of the policy from \mu iterations ago and likely contains more task-relevant information learned during RL training.Table 3 (and Appendix Table 8) shows how performance increases up to a point and then decreases as p in top-p sampling is increased for the masking policy, relaxing the constraint by eliminating less tokens at each step, implying that there is a balance to be found in how much the model should be constrained during RL training. | Incorrect | Change Number |
The use of token-masking policy in the proposed method (NLPO) is theorized by the authors to have been a key reason on why it was able to outperform the PPO based model. They hypothesized that their masking function acts as a dynamic constraint added to the algorithm, that is able to more effectively filter and capture relevant information about the state thanks to its' dynamic nature. | Specifically, NLPOmaintains a masking policy \pi_{\psi}: the masking policy is a copy of the current policy (\pi_{\theta}), but is updated only every \mu steps.A parameterized-invalid-mask is created from \pi_{\psi} by first selecting the top-p tokens from the vocabulary,444\pi_{\psi} could be trained with alternate sampling techniques like top-k or beam search (or even hard-coded via rules by domain experts), though we find top-p sampling to be most effective in practice. and then applying an invalid-mask to the remaining tokens—i.e. setting their probabilities to zero when sampling actions from \pi_{\theta} during training;this periodic updating policy \pi_{\psi} is inspired by off-policy Q-learning algorithms (Andrychowicz et al., 2017),providing the policy \pi_{\theta} with an additional constraint that balances between the benefits of containing more task relevant information than the KL penalty derived from \pi_{0} and the risk of reward hacking.We provide pseudocode in Algorithm 1 (green portions highlight the differences with PPO).PPO vs. NLPO.Figure 2 shows that NLPO generally outperforms PPO and supervised, especially when applied after supervised training.We hypothesize that the primary reason for NLPO’s improved performance and stability is because the masking policy provides an additional constraint for the current policy.This constraint is not based on the initial untuned policy like the KL penalty but of the policy from \mu iterations ago and likely contains more task-relevant information learned during RL training.Table 3 (and Appendix Table 8) shows how performance increases up to a point and then decreases as p in top-p sampling is increased for the masking policy, relaxing the constraint by eliminating less tokens at each step, implying that there is a balance to be found in how much the model should be constrained during RL training. | Correct | null |
The use of token-masking policy in the proposed method (TMM) is theorized by the authors to have been a key reason on why it was able to outperform the BERT based model. They hypothesized that their masking function acts as a dynamic constraint added to the algorithm, that is able to more effectively filter and capture relevant information about the state thanks to its' dynamic nature. | Specifically, NLPOmaintains a masking policy \pi_{\psi}: the masking policy is a copy of the current policy (\pi_{\theta}), but is updated only every \mu steps.A parameterized-invalid-mask is created from \pi_{\psi} by first selecting the top-p tokens from the vocabulary,444\pi_{\psi} could be trained with alternate sampling techniques like top-k or beam search (or even hard-coded via rules by domain experts), though we find top-p sampling to be most effective in practice. and then applying an invalid-mask to the remaining tokens—i.e. setting their probabilities to zero when sampling actions from \pi_{\theta} during training;this periodic updating policy \pi_{\psi} is inspired by off-policy Q-learning algorithms (Andrychowicz et al., 2017),providing the policy \pi_{\theta} with an additional constraint that balances between the benefits of containing more task relevant information than the KL penalty derived from \pi_{0} and the risk of reward hacking.We provide pseudocode in Algorithm 1 (green portions highlight the differences with PPO).PPO vs. NLPO.Figure 2 shows that NLPO generally outperforms PPO and supervised, especially when applied after supervised training.We hypothesize that the primary reason for NLPO’s improved performance and stability is because the masking policy provides an additional constraint for the current policy.This constraint is not based on the initial untuned policy like the KL penalty but of the policy from \mu iterations ago and likely contains more task-relevant information learned during RL training.Table 3 (and Appendix Table 8) shows how performance increases up to a point and then decreases as p in top-p sampling is increased for the masking policy, relaxing the constraint by eliminating less tokens at each step, implying that there is a balance to be found in how much the model should be constrained during RL training. | Incorrect | Change Concept |
The combinatorial action space here probably refers to the set of all possible actions that a RL agent for optimizing a language model could possibly take - here, the action set consists of the entire vocabulary of the language model, which can range to tens of thousands for typical GPT/T5 models used today. This is unlike general RL tasks, where the action space is an order of magnitude smaller. | Language generation action spaces are orders of magnitude larger than what most discrete action space RL algorithms are designed for (Ranzato et al., 2015; Ammanabrolu, 2021), e.g., GPT-2/3 and T5 have a vocabulary size of 50K and 32K respectively.We hypothesize that the size of the action space is a core cause of instability when training LMs with existing RL methods.To address this issue, we introduce NLPO (Natural Language Policy Optimization), which is inspired by work on action elimination/invalid-action masking (Zahavy et al., 2018; Huang & Ontañón, 2020; Ammanabrolu & Hausknecht, 2020). NLPO, a parameterized-masked extension of PPO, learns to mask out less relevant tokens in-context as it trains. NLPO accomplishes this via top-p sampling, which restricts tokens to the smallest possible set whose cumulative probability is greater than the probability parameter p (Holtzman et al., 2018).Each environment is an NLP task: we are given a supervised dataset \mathcal{D}=\{({\bm{x}}^{i},{\bm{y}}^{i})\}_{i=1}^{N} of N examples, where {\bm{x}}\in\mathcal{X} is an language input and {\bm{y}}\in\mathcal{Y} is the target string. Generation can be viewed as a Markov Decision Process (MDP) \langle\mathcal{S},\mathcal{A},\mathcal{R},P,\gamma,T\rangle using a finite vocabulary \mathcal{V}.Each episode in the MDP begins by sampling a datapoint ({\bm{x}},{\bm{y}}) from our dataset and ends when the current time step t exceeds the horizon T or an end of sentence (EOS) token is generated.The input {\bm{x}}=(x_{0},\cdots,x_{m}) is a task-specific prompt that is used as our initial state \bm{s}_{0}=(x_{0},\cdots,x_{m}), where \bm{s}_{0}\in\mathcal{S} and \mathcal{S} is the state space with x_{m}\in\mathcal{V}. An action in the environment a_{t}\in\mathcal{A} consists of a token from our vocabulary \mathcal{V}. The transition function P:\mathcal{S}\times\mathcal{A}\rightarrow\mathcal{S} deterministically appends an action a_{t} to the end of the state \bm{s}_{t-1}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{t-1}). This continues until the end of the horizon t\leq T and we obtain a state \bm{s}_{T}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{T}).At the end of an episode a reward \mathcal{R}:\mathcal{S}\times\mathcal{A}\times\mathcal{Y}\rightarrow\mathbb{R}^{1} that depends on the (\bm{s}_{T},{\bm{y}}) (e.g., an automated metric like PARENT Dhingra et al. (2019)) is emitted. RL4LMs provides an OpenAI gym (Brockman et al., 2016) styleAPI for an RL environmentthat simulates this LM-Based MDP formulation.Abstracting the details of the MDP environment structure allows for new tasks to be added quickly with compatibility across all implemented algorithms. | Correct | null |
The combinatorial action space here probably refers to the set of all possible actions that a RL agent for optimizing a language model could possibly take - here, the action set consists of only a small part of vocabulary of the language model, which can range to tens of thousands for typical GPT/T5 models used today. This is unlike general RL tasks, where the action space is an order of magnitude smaller. | Language generation action spaces are orders of magnitude larger than what most discrete action space RL algorithms are designed for (Ranzato et al., 2015; Ammanabrolu, 2021), e.g., GPT-2/3 and T5 have a vocabulary size of 50K and 32K respectively.We hypothesize that the size of the action space is a core cause of instability when training LMs with existing RL methods.To address this issue, we introduce NLPO (Natural Language Policy Optimization), which is inspired by work on action elimination/invalid-action masking (Zahavy et al., 2018; Huang & Ontañón, 2020; Ammanabrolu & Hausknecht, 2020). NLPO, a parameterized-masked extension of PPO, learns to mask out less relevant tokens in-context as it trains. NLPO accomplishes this via top-p sampling, which restricts tokens to the smallest possible set whose cumulative probability is greater than the probability parameter p (Holtzman et al., 2018).Each environment is an NLP task: we are given a supervised dataset \mathcal{D}=\{({\bm{x}}^{i},{\bm{y}}^{i})\}_{i=1}^{N} of N examples, where {\bm{x}}\in\mathcal{X} is an language input and {\bm{y}}\in\mathcal{Y} is the target string. Generation can be viewed as a Markov Decision Process (MDP) \langle\mathcal{S},\mathcal{A},\mathcal{R},P,\gamma,T\rangle using a finite vocabulary \mathcal{V}.Each episode in the MDP begins by sampling a datapoint ({\bm{x}},{\bm{y}}) from our dataset and ends when the current time step t exceeds the horizon T or an end of sentence (EOS) token is generated.The input {\bm{x}}=(x_{0},\cdots,x_{m}) is a task-specific prompt that is used as our initial state \bm{s}_{0}=(x_{0},\cdots,x_{m}), where \bm{s}_{0}\in\mathcal{S} and \mathcal{S} is the state space with x_{m}\in\mathcal{V}. An action in the environment a_{t}\in\mathcal{A} consists of a token from our vocabulary \mathcal{V}. The transition function P:\mathcal{S}\times\mathcal{A}\rightarrow\mathcal{S} deterministically appends an action a_{t} to the end of the state \bm{s}_{t-1}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{t-1}). This continues until the end of the horizon t\leq T and we obtain a state \bm{s}_{T}=(x_{0},\cdots,x_{m},a_{0},\cdots,a_{T}).At the end of an episode a reward \mathcal{R}:\mathcal{S}\times\mathcal{A}\times\mathcal{Y}\rightarrow\mathbb{R}^{1} that depends on the (\bm{s}_{T},{\bm{y}}) (e.g., an automated metric like PARENT Dhingra et al. (2019)) is emitted. RL4LMs provides an OpenAI gym (Brockman et al., 2016) styleAPI for an RL environmentthat simulates this LM-Based MDP formulation.Abstracting the details of the MDP environment structure allows for new tasks to be added quickly with compatibility across all implemented algorithms. | Incorrect | Change Concept |
Authors best practices were to combine effects of different ways of regularization to produce interpretable images. They first search randomly through 300 different combinations of hyperparameters, then they pick the best four sets of hyperparameters that are compliments to each other and then these sets would be used to visualize preferred images for different classes. | If the above regularization methods are applied individually, they aresomewhat effective at producing more interpretable images; Figure 3 shows the effects of each individual hyperparameter.However, preliminary experiments uncovered that their combinedeffect produces better visualizations. To pick a reasonable set ofhyperparameters for all methods at once, we ran a randomhyperparameter search of 300 possible combinations and settled on fourthat complement each other well. The four selected combinations arelisted in Table 1 and optimized images using each are shown for the “Gorilla” class output unit in Figure 4. Of the four, some show highfrequency information, others low frequency; some contain densepixel data, and others contain only sparse outlines of importantregions.We found the version in the lower-left quadrant to be the best single set of hyperparameters, but often greater intuition canbe gleaned by considering all four at once.Figure 5 shows the optimization results computed for a selection of units on all layers. A single image for every filter of all five convolutional layers is shown in Supplementary Figure S1. Nine images for each filter of all layers, including each of the 1000 ImageNet output classes, can be viewed at http://yosinski.com/deepvis. | Correct | null |
Authors best practices were to combine effects of different ways of regularization to produce interpretable images. They first search randomly through 500 different combinations of hyperparameters, then they pick the best four sets of hyperparameters that are compliments to each other and then these sets would be used to visualize preferred images for different classes. | If the above regularization methods are applied individually, they aresomewhat effective at producing more interpretable images; Figure 3 shows the effects of each individual hyperparameter.However, preliminary experiments uncovered that their combinedeffect produces better visualizations. To pick a reasonable set ofhyperparameters for all methods at once, we ran a randomhyperparameter search of 300 possible combinations and settled on fourthat complement each other well. The four selected combinations arelisted in Table 1 and optimized images using each are shown for the “Gorilla” class output unit in Figure 4. Of the four, some show highfrequency information, others low frequency; some contain densepixel data, and others contain only sparse outlines of importantregions.We found the version in the lower-left quadrant to be the best single set of hyperparameters, but often greater intuition canbe gleaned by considering all four at once.Figure 5 shows the optimization results computed for a selection of units on all layers. A single image for every filter of all five convolutional layers is shown in Supplementary Figure S1. Nine images for each filter of all layers, including each of the 1000 ImageNet output classes, can be viewed at http://yosinski.com/deepvis. | Incorrect | Change Number |
The paper's DNN uses multiple smaller Gaussian kernels iteratively as a way of regularization during the optimization process as seen in equation 2 . | Gaussian blur: Producing images via gradient ascent tends to produce examples with high frequency information (see Supplementary Section S1 for a possible reason). While these images cause high activations, they are neither realistic nor interpretable (Nguyen et al., 2014). A useful regularization is thus to penalize high frequency information. We implement this as a Gaussian blur step rθ (x) = GaussianBlur(x, θb width). Convolving with a blur kernel is more computationally expensive than the other regularization methods, so we added another hyperparameter θb every to allow, for example, blurring every several optimization steps instead of every step. Blurring an image multiple times with a small width Gaussian kernel is equivalent to blurring once with a larger width kernel, and the effect will be similar even if the image changes slightly during the optimization process. This technique thus lowers computational costs without limiting the expressiveness of the regularization. Mahendran & Vedaldi (2014) used a penalty with a similar effect to blurring, called total variation, in their work reconstructing images from layer codes. | Correct | null |
The paper's DNN uses multiple smaller Bayesian kernels iteratively as a way of regularization during the optimization process as seen in equation 2 . | Gaussian blur: Producing images via gradient ascent tends to produce examples with high frequency information (see Supplementary Section S1 for a possible reason). While these images cause high activations, they are neither realistic nor interpretable (Nguyen et al., 2014). A useful regularization is thus to penalize high frequency information. We implement this as a Gaussian blur step rθ (x) = GaussianBlur(x, θb width). Convolving with a blur kernel is more computationally expensive than the other regularization methods, so we added another hyperparameter θb every to allow, for example, blurring every several optimization steps instead of every step. Blurring an image multiple times with a small width Gaussian kernel is equivalent to blurring once with a larger width kernel, and the effect will be similar even if the image changes slightly during the optimization process. This technique thus lowers computational costs without limiting the expressiveness of the regularization. Mahendran & Vedaldi (2014) used a penalty with a similar effect to blurring, called total variation, in their work reconstructing images from layer codes. | Incorrect | Change Concept |
Authors mainly introduce four different and newly used regularizations that would help researchers in visualizing responses from different layers. These regularizations are designed to overcome different pathologies commonly encountered by gradient descent without regularization : L2 decay to penalize large pixel values which do not naturally occur, Gaussian blur:a useful regularization to iteratively penalize high frequency information associated with generated images via gradient ascent through each optimization step, Clipping pixels with small norm or Clipping pixels with small contribution. | We investigated the following four regularizations. All are designed to overcome different pathologies commonly encountered by gradient descent without regularization.Gaussian blur: Producing images via gradient ascent tends to produce examples with high frequency information (see Supplementary Section S1 for a possible reason). While these images cause high activations, they are neither realistic nor interpretable (Nguyen et al., 2014). A useful regularization is thus to penalize high frequency information. We implement this as a Gaussian blur step r_{\theta}(\mathbf{x})=\mathrm{GaussianBlur}(\mathbf{x},\theta_{\mathrm{b\_width}}). Convolving with a blur kernel is more computationally expensive than the other regularization methods, so we added another hyperparameter \theta_{\mathrm{b\_every}} to allow, for example, blurring every several optimization steps instead of every step. Blurring an image multiple times with a small width Gaussian kernel is equivalent to blurring once with a larger width kernel, and the effect will be similar even if the image changes slightly during the optimization process. This technique thus lowers computational costs without limiting the expressiveness of the regularization. Mahendran & Vedaldi (2014) used a penalty with a similar effect to blurring, called total variation, in their work reconstructing images from layer codes.The second tool — new regularizations that enable improved, interpretable, optimized visualizations of learned features — will help researchers and practitioners understand, debug, and improve their models. The visualizations also reveal a new twist in an ongoing story. Previous studieshave shown that discriminative networks can easily be fooled or hacked by the addition of certain structurednoise in image space (Szegedy et al., 2013; Nguyen et al., 2014).An oft-cited reason for this property is that discriminative training leads networksto ignore non-discriminative information in their input, e.g. learning to detect jaguars by matching the unique spots on their fur while ignoring the fact that they have four legs. For this reason it has been seen as a hopeless endeavor to create a generative model in which one randomly samples an x from a broad distribution on the space of all possible imagesand then iteratively transforms x into a recognizable image by moving it to a region that satisfies both a prior p(x) and posterior p(y|x) for some class label y.Past attempts have largely supported this view by producing unrealistic images using this method (Nguyen et al., 2014; Simonyan et al., 2013).Paragraph 10 : L2 decay: A common regularization, L2 decay penalizes large values and is implemented as rθ (x) = (1−θdecay)·x. L2 decay tends to prevent a small number of extreme pixel values from dominating the example image. Such extreme single-pixel values neither occur naturally with great frequency nor are useful for visualization.Paragraph 11 :Clipping pixels with small norm: The first two regularizations suppress high amplitude and high frequency information, so after applying both, we are left with an x∗ that contains somewhat small, somewhat smooth values. However, x∗ will still tend to contain non-zero pixel values everywhere. Even if some pixels in x∗ show the primary object or type of input causing the unit under consideration to activate, the gradient with respect to all other pixels in x∗ will still generally be non-zero, so these pixels will also shift to show some pattern as well, contributing in whatever small way they can to ultimately raise the chosen unit’s activation. We wish to bias the search away from such behavior and instead show only the main object, letting other regions be exactly zero if they are not needed.Paragraph 12 : Clipping pixels with small contribution: Instead of clipping pixels with small norms, we can try something slightly smarter and clip pixels with small contributions to the activation. | Correct | null |
Authors mainly introduce three different and newly used regularizations that would help researchers in visualizing responses from different layers. These regularizations are designed to overcome different pathologies commonly encountered by gradient descent without regularization : L2 decay to penalize large pixel values which do not naturally occur, Gaussian blur:a useful regularization to iteratively penalize high frequency information associated with generated images via gradient ascent through each optimization step, Clipping pixels with small norm or Clipping pixels with small contribution. | We investigated the following four regularizations. All are designed to overcome different pathologies commonly encountered by gradient descent without regularization.Gaussian blur: Producing images via gradient ascent tends to produce examples with high frequency information (see Supplementary Section S1 for a possible reason). While these images cause high activations, they are neither realistic nor interpretable (Nguyen et al., 2014). A useful regularization is thus to penalize high frequency information. We implement this as a Gaussian blur step r_{\theta}(\mathbf{x})=\mathrm{GaussianBlur}(\mathbf{x},\theta_{\mathrm{b\_width}}). Convolving with a blur kernel is more computationally expensive than the other regularization methods, so we added another hyperparameter \theta_{\mathrm{b\_every}} to allow, for example, blurring every several optimization steps instead of every step. Blurring an image multiple times with a small width Gaussian kernel is equivalent to blurring once with a larger width kernel, and the effect will be similar even if the image changes slightly during the optimization process. This technique thus lowers computational costs without limiting the expressiveness of the regularization. Mahendran & Vedaldi (2014) used a penalty with a similar effect to blurring, called total variation, in their work reconstructing images from layer codes.The second tool — new regularizations that enable improved, interpretable, optimized visualizations of learned features — will help researchers and practitioners understand, debug, and improve their models. The visualizations also reveal a new twist in an ongoing story. Previous studieshave shown that discriminative networks can easily be fooled or hacked by the addition of certain structurednoise in image space (Szegedy et al., 2013; Nguyen et al., 2014).An oft-cited reason for this property is that discriminative training leads networksto ignore non-discriminative information in their input, e.g. learning to detect jaguars by matching the unique spots on their fur while ignoring the fact that they have four legs. For this reason it has been seen as a hopeless endeavor to create a generative model in which one randomly samples an x from a broad distribution on the space of all possible imagesand then iteratively transforms x into a recognizable image by moving it to a region that satisfies both a prior p(x) and posterior p(y|x) for some class label y.Past attempts have largely supported this view by producing unrealistic images using this method (Nguyen et al., 2014; Simonyan et al., 2013).Paragraph 10 : L2 decay: A common regularization, L2 decay penalizes large values and is implemented as rθ (x) = (1−θdecay)·x. L2 decay tends to prevent a small number of extreme pixel values from dominating the example image. Such extreme single-pixel values neither occur naturally with great frequency nor are useful for visualization.Paragraph 11 :Clipping pixels with small norm: The first two regularizations suppress high amplitude and high frequency information, so after applying both, we are left with an x∗ that contains somewhat small, somewhat smooth values. However, x∗ will still tend to contain non-zero pixel values everywhere. Even if some pixels in x∗ show the primary object or type of input causing the unit under consideration to activate, the gradient with respect to all other pixels in x∗ will still generally be non-zero, so these pixels will also shift to show some pattern as well, contributing in whatever small way they can to ultimately raise the chosen unit’s activation. We wish to bias the search away from such behavior and instead show only the main object, letting other regions be exactly zero if they are not needed.Paragraph 12 : Clipping pixels with small contribution: Instead of clipping pixels with small norms, we can try something slightly smarter and clip pixels with small contributions to the activation. | Incorrect | Change Number |
linear sweep can be seen as a regular increment in the values of some regularization hyperparameter (from leftmost where there is no regularization to rightmost where strong regularization occur ) to see the variation of their effects on the corresponding activations. | Figure 3. The effects of each regularization method from Section 3 when used individually. Each of the four rows shows a linear sweep in hyperparameter space from no regularization (left) to strong regularization (right). When applied too strongly, some regularizations cause the optimization to fail (e.g. L2 decay, top row) or the images to be less interpretable (small norm and small contribution clipping, bottom two rows). For this reason, a random hyperparameter search was useful for finding joint hyperparameter settings that worked well together (see Figure 4). Best viewed electronically, zoomed in. | Correct | null |
linear sweep cannot be seen as a regular increment in the values of some regularization hyperparameter (from leftmost where there is no regularization to rightmost where strong regularization occur ) to see the variation of their effects on the corresponding activations. | Figure 3. The effects of each regularization method from Section 3 when used individually. Each of the four rows shows a linear sweep in hyperparameter space from no regularization (left) to strong regularization (right). When applied too strongly, some regularizations cause the optimization to fail (e.g. L2 decay, top row) or the images to be less interpretable (small norm and small contribution clipping, bottom two rows). For this reason, a random hyperparameter search was useful for finding joint hyperparameter settings that worked well together (see Figure 4). Best viewed electronically, zoomed in. | Incorrect | Opposite |
Calculating absolute difference between some neuron activation of an input and the activation for same input without certain pixel can be considered a way of measuring the contribution of that pixel in the total response of the neuron. To ensure faster computation, we can estimate activation near the input with 1st order approximation (linear) and hence this leads to total contribution estimated as the element wise product of the activation gradient and the input x (each element in this product shows how this pixel affects the total response ), we repeat this product for all different channels ,sum them all, and take absolute value to find pixels with small contribution in either direction, positive or negative to get rid of. | Clipping pixels with small contribution: Instead of clipping pixels with small norms, we can try something slightly smarter and clip pixels with small contributions to the activation. One way of computing a pixel’s contribution to an activation is to measure how much the activation increases or decreases when the pixel is set to zero; that is, to compute the contribution as |a_{i}(\mathbf{x})-a_{i}(\mathbf{x}_{-j})|, where \mathbf{x}_{-j} is \mathbf{x} but with the j^{th} pixel set to zero. This approach is straightforward but prohibitively slow, requiring a forward pass for every pixel. Instead, we approximate this process by linearizing a_{i}(\mathbf{x}) around \mathbf{x}, in which case the contribution of each dimension of \mathbf{x} can be estimated as the elementwise product of \mathbf{x} and the gradient. We then sum over all three channels and take the absolute value, computing \left|\sum_{c}\mathbf{x}\circ\nabla_{\mathbf{x}}a_{i}(\mathbf{x})\right|. We use the absolute value to find pixels with small contribution in either direction, positive or negative. While we could choose to keep the pixel transitions where setting the pixel to zero would result in a large activation increase, these shifts are already handled by gradient ascent, and here we prefer to clip only the pixels that are deemed not to matter, not to take large gradient steps outside the region where the linear approximation is most valid. We define this r_{\theta}(\mathbf{x}) as the operation that sets pixels with contribution under the \theta_{\mathrm{c\_pct}} percentile to zero. | Correct | null |
Calculating absolute difference between some neuron activation of an input and the activation for same input without certain pixel can be considered a way of measuring the contribution of that pixel in the partial response of the neuron. To ensure faster computation, we can estimate activation near the input with 1st order approximation (linear) and hence this leads to total contribution estimated as the element wise product of the activation gradient and the input x (each element in this product shows how this pixel affects the total response ), we repeat this product for all different channels ,sum them all, and take absolute value to find pixels with small contribution in either direction, positive or negative to get rid of. | Clipping pixels with small contribution: Instead of clipping pixels with small norms, we can try something slightly smarter and clip pixels with small contributions to the activation. One way of computing a pixel’s contribution to an activation is to measure how much the activation increases or decreases when the pixel is set to zero; that is, to compute the contribution as |a_{i}(\mathbf{x})-a_{i}(\mathbf{x}_{-j})|, where \mathbf{x}_{-j} is \mathbf{x} but with the j^{th} pixel set to zero. This approach is straightforward but prohibitively slow, requiring a forward pass for every pixel. Instead, we approximate this process by linearizing a_{i}(\mathbf{x}) around \mathbf{x}, in which case the contribution of each dimension of \mathbf{x} can be estimated as the elementwise product of \mathbf{x} and the gradient. We then sum over all three channels and take the absolute value, computing \left|\sum_{c}\mathbf{x}\circ\nabla_{\mathbf{x}}a_{i}(\mathbf{x})\right|. We use the absolute value to find pixels with small contribution in either direction, positive or negative. While we could choose to keep the pixel transitions where setting the pixel to zero would result in a large activation increase, these shifts are already handled by gradient ascent, and here we prefer to clip only the pixels that are deemed not to matter, not to take large gradient steps outside the region where the linear approximation is most valid. We define this r_{\theta}(\mathbf{x}) as the operation that sets pixels with contribution under the \theta_{\mathrm{c\_pct}} percentile to zero. | Incorrect | Change Concept |
Dataset-centric approach requires the trained network together with some dataset to run through the network showing high or low responses of different units while interacting with most significant images of such dataset. This approach can also use deconvolution layers and upsampling to map and highlight the regions of an image that were responsible of the firing of the different units. "Network-centric approach" deals only with network without the need to any dataset. You can start with some initial input, compute activations through the forward path and then compute gradients while backprop. You can then ascent or descent the input towards gradient until you reach a preferred input stimulus x* for the unit under consideration. Working with input images, you can visualize that x* if you want. | Another approach is to try to interpret the function computed by each individual neuron. Past studies in this vein roughly divide into two different camps: dataset-centric and network-centric. The former requires both a trained DNN and running data through that network; the latter requires only the trained network itself. One dataset-centric approach is to display images from the training or test set that cause high or low activations for individual units. Another is the deconvolution method of Zeiler & Fergus (2013), which highlights the portions of a particular image that are responsible for the firing of each neural unit.Network-centric approaches investigate a network directly without any data from a dataset. For example, Erhan et al. (2009) synthesized images that cause high activations for particular units. Starting with some initial input \mathbf{x}=\mathbf{x_{0}}, the activation a_{i}(\mathbf{x}) caused at some unit i by this input is computed, and then steps are taken in input spacealong the gradient \partial a_{i}(\mathbf{x})/\partial\mathbf{x} to synthesize inputs that cause higher and higher activations of unit i, eventually terminating at some \mathbf{x^{*}} which is deemed to be a preferred input stimulus for the unit in question. In the case where the input space is an image, \mathbf{x^{*}} can be displayed directly for interpretation. Others have followed suit, using the gradient to find images that cause higher activations (Simonyan et al., 2013; Nguyen et al., 2014) or lower activations (Szegedy et al., 2013) for output units. | Correct | null |
Network-centric approachDataset-centric approach requires the trained network together with some dataset to run through the network showing high or low responses of different units while interacting with most significant images of such dataset. This approach can also use deconvolution layers and upsampling to map and highlight the regions of an image that were responsible of the firing of the different units. "Dataset-centric approach" deals only with network without the need to any dataset. You can start with some initial input, compute activations through the forward path and then compute gradients while backprop. You can then ascent or descent the input towards gradient until you reach a preferred input stimulus x* for the unit under consideration. Working with input images, you can visualize that x* if you want. | Another approach is to try to interpret the function computed by each individual neuron. Past studies in this vein roughly divide into two different camps: dataset-centric and network-centric. The former requires both a trained DNN and running data through that network; the latter requires only the trained network itself. One dataset-centric approach is to display images from the training or test set that cause high or low activations for individual units. Another is the deconvolution method of Zeiler & Fergus (2013), which highlights the portions of a particular image that are responsible for the firing of each neural unit.Network-centric approaches investigate a network directly without any data from a dataset. For example, Erhan et al. (2009) synthesized images that cause high activations for particular units. Starting with some initial input \mathbf{x}=\mathbf{x_{0}}, the activation a_{i}(\mathbf{x}) caused at some unit i by this input is computed, and then steps are taken in input spacealong the gradient \partial a_{i}(\mathbf{x})/\partial\mathbf{x} to synthesize inputs that cause higher and higher activations of unit i, eventually terminating at some \mathbf{x^{*}} which is deemed to be a preferred input stimulus for the unit in question. In the case where the input space is an image, \mathbf{x^{*}} can be displayed directly for interpretation. Others have followed suit, using the gradient to find images that cause higher activations (Simonyan et al., 2013; Nguyen et al., 2014) or lower activations (Szegedy et al., 2013) for output units. | Incorrect | Change Concept |
An example of "dataset-centric" approach can be deconvolution method which is used to highlight certain regions of some image that has the highest effects in the response of different units. | Another approach is to try to interpret the function computed by each individual neuron. Past studies in this vein roughly divide into two different camps: dataset-centric and network-centric. The former requires both a trained DNN and running data through that network; the latter requires only the trained network itself. One dataset-centric approach is to display images from the training or test set that cause high or low activations for individual units. Another is the deconvolution method of Zeiler & Fergus (2013), which highlights the portions of a particular image that are responsible for the firing of each neural unit. | Correct | null |
An example of "dataset-centric" approach can be deconvolution method which is used to highlight all regions of some image that has the highest effects in the response of different units. | Another approach is to try to interpret the function computed by each individual neuron. Past studies in this vein roughly divide into two different camps: dataset-centric and network-centric. The former requires both a trained DNN and running data through that network; the latter requires only the trained network itself. One dataset-centric approach is to display images from the training or test set that cause high or low activations for individual units. Another is the deconvolution method of Zeiler & Fergus (2013), which highlights the portions of a particular image that are responsible for the firing of each neural unit. | Incorrect | Change Concept |
An example of such approach would be to consider a trained network, start with some initial input and compute the forward path activations. compute gradients through backprop and then move this input towards or against the gradient direction until you have some interesting input that is of much significance in the responses of considered neurons. | Network-centric approaches investigate a network directly without any data from a dataset. For example, Erhan et al. (2009) synthesized images that cause high activations for particular units. Starting with some initial input \mathbf{x}=\mathbf{x_{0}}, the activation a_{i}(\mathbf{x}) caused at some unit i by this input is computed, and then steps are taken in input spacealong the gradient \partial a_{i}(\mathbf{x})/\partial\mathbf{x} to synthesize inputs that cause higher and higher activations of unit i, eventually terminating at some \mathbf{x^{*}} which is deemed to be a preferred input stimulus for the unit in question. In the case where the input space is an image, \mathbf{x^{*}} can be displayed directly for interpretation. Others have followed suit, using the gradient to find images that cause higher activations (Simonyan et al., 2013; Nguyen et al., 2014) or lower activations (Szegedy et al., 2013) for output units. | Correct | null |
An example of such approach would be to consider a trained network, start without any initial input and compute the forward path activations. compute gradients through backprop and then move this input towards or against the gradient direction until you have some interesting input that is of much significance in the responses of considered neurons. | Network-centric approaches investigate a network directly without any data from a dataset. For example, Erhan et al. (2009) synthesized images that cause high activations for particular units. Starting with some initial input \mathbf{x}=\mathbf{x_{0}}, the activation a_{i}(\mathbf{x}) caused at some unit i by this input is computed, and then steps are taken in input spacealong the gradient \partial a_{i}(\mathbf{x})/\partial\mathbf{x} to synthesize inputs that cause higher and higher activations of unit i, eventually terminating at some \mathbf{x^{*}} which is deemed to be a preferred input stimulus for the unit in question. In the case where the input space is an image, \mathbf{x^{*}} can be displayed directly for interpretation. Others have followed suit, using the gradient to find images that cause higher activations (Simonyan et al., 2013; Nguyen et al., 2014) or lower activations (Szegedy et al., 2013) for output units. | Incorrect | Change Concept |
'Hacks' means that they are not likely to naturally exist (non-natural looking images). However they may even cause harmful changes in the response of the network. Adversarial points for instance are examples of such hacks where slight increments in pixels of even correctly classified images can make them fool the network and tend to go beyond there original part of space and hence be misclassified. | These gradient-based approaches are attractive in their simplicity, but the optimization process tends to produce images that do not greatly resemble natural images. Instead, they are composed of a collection of “hacks” that happen to cause high (or low) activations: extreme pixel values, structured high frequency patterns, and copies of common motifs without global structure (Simonyan et al., 2013; Nguyen et al., 2014; Szegedy et al., 2013; Goodfellow et al., 2014). The fact that activations may be effected by such hacks is better understood thanks to several recent studies. Specifically, it has been shown that such hacks may be applied to correctly classified images to cause them to be misclassified even via imperceptibly small changes (Szegedy et al., 2013), that such hacks can be found even without the gradient information to produce unrecognizable “fooling examples” (Nguyen et al., 2014), and that the abundance of non-natural looking images that cause extreme activations can be explained by the locally linear behavior of neural nets(Goodfellow et al., 2014). | Correct | null |
'Hacks' means that they are not likely to naturally exist (non-natural looking images). However they may even cause harmful changes in the response of the network. Adversarial points for instance are examples of such hacks where slight increments in pixels of incorrectly classified images can make them fool the network and tend to go beyond there original part of space and hence be misclassified. | These gradient-based approaches are attractive in their simplicity, but the optimization process tends to produce images that do not greatly resemble natural images. Instead, they are composed of a collection of “hacks” that happen to cause high (or low) activations: extreme pixel values, structured high frequency patterns, and copies of common motifs without global structure (Simonyan et al., 2013; Nguyen et al., 2014; Szegedy et al., 2013; Goodfellow et al., 2014). The fact that activations may be effected by such hacks is better understood thanks to several recent studies. Specifically, it has been shown that such hacks may be applied to correctly classified images to cause them to be misclassified even via imperceptibly small changes (Szegedy et al., 2013), that such hacks can be found even without the gradient information to produce unrecognizable “fooling examples” (Nguyen et al., 2014), and that the abundance of non-natural looking images that cause extreme activations can be explained by the locally linear behavior of neural nets(Goodfellow et al., 2014). | Incorrect | Change Concept |
row-major means that consecutive small grayscale images of each row reside next to each other unlike "column-major" and both are methods of storing elements in memory. | Figure 1 shows examples of this type of plot for the \mathsf{conv5} layer.The \mathsf{conv5} layer has size 256\times13\times13, which we depict as 256 separate 13\times13 grayscale images. Each of the 256 small images contains activations in the same spatial x-y spatial layout as the input data, and the 256 images are simply and arbitrarily tiled into a 16\times16 grid in row-major order.Figure 2 shows a zoomed in view of one particular channel, \mathsf{conv5_{151}}, that responds to human and animal faces. All layers can be viewed in the software tool, including pooling and normalization layers. Visualizing these layers provides intuitions about their effects and functions. | Correct | null |
row-major means that consecutive small RGB images of each row reside next to each other unlike "column-major" and both are methods of storing elements in memory. | Figure 1 shows examples of this type of plot for the \mathsf{conv5} layer.The \mathsf{conv5} layer has size 256\times13\times13, which we depict as 256 separate 13\times13 grayscale images. Each of the 256 small images contains activations in the same spatial x-y spatial layout as the input data, and the 256 images are simply and arbitrarily tiled into a 16\times16 grid in row-major order.Figure 2 shows a zoomed in view of one particular channel, \mathsf{conv5_{151}}, that responds to human and animal faces. All layers can be viewed in the software tool, including pooling and normalization layers. Visualizing these layers provides intuitions about their effects and functions. | Incorrect | Change Concept |
Zero mean input data and Standardization in general improve the convergence properties of BP training, so it can help to reach desired solution fast. Also, Authors may intend to have centered inputs so that network reduces its biasing towards certain classes or certain large or tiny response values, hence we can have reasonable values for activations and more visualizable responses from different neurons. | Our network was trained on ImageNet by first subtracting the per-pixel mean of examples in ImageNet before inputting training examples to the network. Thus, the direct input to the network, \mathbf{x}, can be thought of as a zero-centered input. We may pose the optimization problem as finding an image \mathbf{x^{*}} where | Correct | null |
Zero mean input data and Standardization in general improve the convergence properties of BP training and few-shot tunning so it can help to reach desired solution fast. Also, Authors may intend to have centered inputs so that network reduces its biasing towards certain classes or certain large or tiny response values, hence we can have reasonable values for activations and more visualizable responses from different neurons. | Our network was trained on ImageNet by first subtracting the per-pixel mean of examples in ImageNet before inputting training examples to the network. Thus, the direct input to the network, \mathbf{x}, can be thought of as a zero-centered input. We may pose the optimization problem as finding an image \mathbf{x^{*}} where | Incorrect | Add Unmentioned Information |
The reason is that convolution layers learn parameters that can extract useful information and relations from the feature map that can help it afterwards to judge and give suitable responses of what this category is. Responses from learned detectors can resemble among some set of categories and can also differ among other set of categories. Input -not being in the training classes- still has a feature map that different layers would respond to according to those different detectors which the network has already learned and would still give a probability vector which may not be accurate. Hence, having noise in the input can stimulate different detectors to respond and fire different activations that would lead to changes in the probability output vector. | •One of the most interesting conclusions so far has been that representations on some layers seem to be surprisingly local. Instead of finding distributed representations on all layers, we see, for example, detectors for text, flowers, fruit, and faces on \mathsf{conv4} and \mathsf{conv5}. These conclusions can be drawn either from the live visualization or the optimized images (or, best, by using both in concert) and suggest several directions for future research (discussed in Section 4).•When using direct file input to classify photos from Flickr or Google Images, classifications are often correct and highly confident (softmax probability for correct class near 1). On the other hand, when using input from a webcam, predictions often cannot be correct because no items from the training set are shown in the image. The training set’s 1000 classes, though numerous, do not cover most common household objects. Thus, when shown a typical webcam view of a person with no ImageNet classes present, the output has no single high probability, as is expected. Surprisingly, however, this probability vector is noisy and varies significantly in response to tiny changes in the input, often changing merely in response to the noise from the webcam. We might have instead expected unchanging and low confidence predictions for a given scene when no object the network has been trained to classify is present. Plotting the fully connected layers (\mathsf{fc6} and \mathsf{fc7}) also reveals a similar sensitivity to small input changes.•Although the last three layers are sensitive to small input changes, much of the lower layer computation is more robust. For example, when visualizing the \mathsf{conv5} layer, one can find many invariant detectors for faces, shoulders, text, etc. by moving oneself or objects in front of the camera. Even though the 1000 classes contain no explicitly labeled faces or text, the network learns to identify these concepts simply because they represent useful partial information for making a later classification decision. One face detector, denoted \mathsf{conv5_{151}} (channel number 151 on \mathsf{conv5}), is shown in Figure 2 activating for human and lion faces and in Figure 1 activating for a cat face. Zhou et al. (2014) recently observed a similar effect where convnets trained only to recognize different scene types — playgrounds, restaurant patios, living rooms, etc. — learn object detectors (e.g. for chairs, books, and sofas) on intermediate layers. | Correct | null |
The reason is that convolution layers learn parameters that can extract useful information and relations from the feature map that can help it afterwards to judge and give suitable responses of what this category is. Responses from learned detectors can resemble among some set of categories and can also differ among other set of categories. Input -not being in the training classes- still has a feature map that different layers would respond to according to those unlearned detectors and would still give a probability vector which may not be accurate. Hence, having noise in the input can stimulate different detectors to respond and fire different activations that would lead to changes in the probability output vector. | •One of the most interesting conclusions so far has been that representations on some layers seem to be surprisingly local. Instead of finding distributed representations on all layers, we see, for example, detectors for text, flowers, fruit, and faces on \mathsf{conv4} and \mathsf{conv5}. These conclusions can be drawn either from the live visualization or the optimized images (or, best, by using both in concert) and suggest several directions for future research (discussed in Section 4).•When using direct file input to classify photos from Flickr or Google Images, classifications are often correct and highly confident (softmax probability for correct class near 1). On the other hand, when using input from a webcam, predictions often cannot be correct because no items from the training set are shown in the image. The training set’s 1000 classes, though numerous, do not cover most common household objects. Thus, when shown a typical webcam view of a person with no ImageNet classes present, the output has no single high probability, as is expected. Surprisingly, however, this probability vector is noisy and varies significantly in response to tiny changes in the input, often changing merely in response to the noise from the webcam. We might have instead expected unchanging and low confidence predictions for a given scene when no object the network has been trained to classify is present. Plotting the fully connected layers (\mathsf{fc6} and \mathsf{fc7}) also reveals a similar sensitivity to small input changes.•Although the last three layers are sensitive to small input changes, much of the lower layer computation is more robust. For example, when visualizing the \mathsf{conv5} layer, one can find many invariant detectors for faces, shoulders, text, etc. by moving oneself or objects in front of the camera. Even though the 1000 classes contain no explicitly labeled faces or text, the network learns to identify these concepts simply because they represent useful partial information for making a later classification decision. One face detector, denoted \mathsf{conv5_{151}} (channel number 151 on \mathsf{conv5}), is shown in Figure 2 activating for human and lion faces and in Figure 1 activating for a cat face. Zhou et al. (2014) recently observed a similar effect where convnets trained only to recognize different scene types — playgrounds, restaurant patios, living rooms, etc. — learn object detectors (e.g. for chairs, books, and sofas) on intermediate layers. | Incorrect | Change Concept |
The paper reaches this goal by calculating each pixel norm over the 3 colour channels and zeroing out small-norm pixels according to some threshold (the percentile of all pixel norms in x). | Clipping pixels with small norm: The first two regularizations suppress high amplitude and high frequency information, so after applying both, we are left with an \mathbf{x^{*}} that contains somewhat small, somewhat smooth values. However, \mathbf{x^{*}} will still tend to contain non-zero pixel values everywhere. Even if some pixels in \mathbf{x^{*}} show the primary object or type of input causing the unit under consideration to activate, the gradient with respect to all other pixels in \mathbf{x^{*}} will still generally be non-zero, so these pixels will also shift to show some pattern as well, contributing in whatever small way they can to ultimately raise the chosen unit’s activation. We wish to bias the search away from such behavior and instead show only the main object, letting other regions be exactly zero if they are not needed. We implement this bias using an r_{\theta}(\mathbf{x}) that computes the norm of each pixel (over red, green, and blue channels) and then sets any pixels with small norm to zero. The threshold for the norm, \theta_{\mathrm{n\_pct}}, is specified as a percentile of all pixel norms in \mathbf{x}. | Correct | null |
The paper reaches this goal by calculating each pixel norm over the 3 colour channels and zeroing out small-norm pixels randomly. | Clipping pixels with small norm: The first two regularizations suppress high amplitude and high frequency information, so after applying both, we are left with an \mathbf{x^{*}} that contains somewhat small, somewhat smooth values. However, \mathbf{x^{*}} will still tend to contain non-zero pixel values everywhere. Even if some pixels in \mathbf{x^{*}} show the primary object or type of input causing the unit under consideration to activate, the gradient with respect to all other pixels in \mathbf{x^{*}} will still generally be non-zero, so these pixels will also shift to show some pattern as well, contributing in whatever small way they can to ultimately raise the chosen unit’s activation. We wish to bias the search away from such behavior and instead show only the main object, letting other regions be exactly zero if they are not needed. We implement this bias using an r_{\theta}(\mathbf{x}) that computes the norm of each pixel (over red, green, and blue channels) and then sets any pixels with small norm to zero. The threshold for the norm, \theta_{\mathrm{n\_pct}}, is specified as a percentile of all pixel norms in \mathbf{x}. | Incorrect | Change Concept |
The authors use 300 sets of possible hyperparameter combinations for the random hyperparameter search, then choose four of them that complement each other well. | If the above regularization methods are applied individually, they aresomewhat effective at producing more interpretable images; Figure 3 shows the effects of each individual hyperparameter.However, preliminary experiments uncovered that their combinedeffect produces better visualizations. To pick a reasonable set ofhyperparameters for all methods at once, we ran a randomhyperparameter search of 300 possible combinations and settled on fourthat complement each other well. The four selected combinations arelisted in Table 1 and optimized images using each are shown for the “Gorilla” class output unit in Figure 4. Of the four, some show highfrequency information, others low frequency; some contain densepixel data, and others contain only sparse outlines of importantregions.We found the version in the lower-left quadrant to be the best single set of hyperparameters, but often greater intuition canbe gleaned by considering all four at once.Figure 5 shows the optimization results computed for a selection of units on all layers. A single image for every filter of all five convolutional layers is shown in Supplementary Figure S1. Nine images for each filter of all layers, including each of the 1000 ImageNet output classes, can be viewed at http://yosinski.com/deepvis. | Correct | null |
The authors use 100 sets of possible hyperparameter combinations for the random hyperparameter search, then choose four of them that complement each other well. | If the above regularization methods are applied individually, they aresomewhat effective at producing more interpretable images; Figure 3 shows the effects of each individual hyperparameter.However, preliminary experiments uncovered that their combinedeffect produces better visualizations. To pick a reasonable set ofhyperparameters for all methods at once, we ran a randomhyperparameter search of 300 possible combinations and settled on fourthat complement each other well. The four selected combinations arelisted in Table 1 and optimized images using each are shown for the “Gorilla” class output unit in Figure 4. Of the four, some show highfrequency information, others low frequency; some contain densepixel data, and others contain only sparse outlines of importantregions.We found the version in the lower-left quadrant to be the best single set of hyperparameters, but often greater intuition canbe gleaned by considering all four at once.Figure 5 shows the optimization results computed for a selection of units on all layers. A single image for every filter of all five convolutional layers is shown in Supplementary Figure S1. Nine images for each filter of all layers, including each of the 1000 ImageNet output classes, can be viewed at http://yosinski.com/deepvis. | Incorrect | Change Number |
Generative structure is how the data is distributed inside the space where it lives, for example when learning to detect jaguar class, parameters encode not only the jaguar’s spots(Only to distinguish it through a rare property), but to some extent also its four legs(to learn the pattern with which the whole creature can be found). So, discriminative parameters also contain significant “generative” structure. | However, the results presented here suggest an alternate possibility: the previously used priors may simply have been too weak (see Section S1 for one hypothesis of why a strong p(x) model is needed). With the careful design or learning of a p(x) model that biases toward realism,one may be able to harnessthe large number of parameters present in a discriminately learned p(y|x) modelto generate realistic images by enforcing probability under both models simultaneously.Even with the simple, hand-coded p(x) models we use in this paper as regularizers, complex dependencies between distant pixels already arise (cf. the beetles with structure spanning over 100 pixels in Figure 4). This implies that the discriminative parameters also contain significant “generative” structure from thetraining dataset; that is, the parameters encodenot only the jaguar’s spots, but to some extent also its four legs.With better, learned probabilistic models over the input and activations of higher layers, much more structure may be apparent. Work by Dai et al. (2015) shows some interesting results in this direction.While the images generated in this paper are far from being photo-realistic, they do suggest thattransferring discriminatively trained parameters to generative models — opposite the direction of the usual unsupervised pretraining approach — may be a fruitful area for further investigation. | Correct | null |
Generative structure is how the data is distributed inside the space where it lives, for example when learning to detect jaguar class, parameters encode only the jaguar’s spots(Only to distinguish it through a rare property), without encoding its legs. So, discriminative parameters also contain significant “generative” structure. | However, the results presented here suggest an alternate possibility: the previously used priors may simply have been too weak (see Section S1 for one hypothesis of why a strong p(x) model is needed). With the careful design or learning of a p(x) model that biases toward realism,one may be able to harnessthe large number of parameters present in a discriminately learned p(y|x) modelto generate realistic images by enforcing probability under both models simultaneously.Even with the simple, hand-coded p(x) models we use in this paper as regularizers, complex dependencies between distant pixels already arise (cf. the beetles with structure spanning over 100 pixels in Figure 4). This implies that the discriminative parameters also contain significant “generative” structure from thetraining dataset; that is, the parameters encodenot only the jaguar’s spots, but to some extent also its four legs.With better, learned probabilistic models over the input and activations of higher layers, much more structure may be apparent. Work by Dai et al. (2015) shows some interesting results in this direction.While the images generated in this paper are far from being photo-realistic, they do suggest thattransferring discriminatively trained parameters to generative models — opposite the direction of the usual unsupervised pretraining approach — may be a fruitful area for further investigation. | Incorrect | Change Concept |
It is said that when evaluating a retriever trained on a source domain in an out-of-domain setting, the performance is obtained lower than BM25. Also, dense retrievers are said to be sensitive to domain shift and models that perform well on MS MARCO do not perform well on COVID-19 data. There have been many studies on unsupervised sentence embedding learning, but it is said that they do not work well in unsupervised dense retrieval. Therefore, the performance of the retriever in out-of-domain may be worse. | However, as shown in Thakur et al. (2021b), dense retrieval methods require large amounts of training data to work well.333For reference, the popular MS MARCO dataset (Nguyen et al., 2016) has about 500k training instances; the Natural Questions dataset (Kwiatkowski et al., 2019) has more than 100k training instances. Most importantly, dense retrieval methods are extremely sensitive to domain shifts: Models trained on MS MARCO perform rather poorly for questions for COVID-19 scientific literature (Wang et al., 2020; Voorhees et al., 2021). The MS MARCO dataset was created before COVID-19, hence, it does not include any COVID-19 related topics and models did not learn how to represent this topic well in a vector space.We use the MS MARCO passage ranking dataset Nguyen et al. (2016) as the data from the source domain. It has 8.8M passages and 532.8K query-passage pairs labeled as relevant in the training set. As Table 1 shows, a state-of-the-art dense retrieval model, achieving an MRR@10 of 33.2 points on the MS MARCO passage ranking dataset, performs poorly on the six selected domain-specific retrieval datasets when compared to simple BM25 lexical search.So far, ICT and CD have only been studied on in-domain performance, i.e. a large in-domain labeled dataset is available which is used for subsequent supervised fine-tuning. SimCSE, CT, and TSDAE have been only studied for unsupervised sentence embedding learning. As our results show in Appendix E, they do not work at all for purely unsupervised dense retrieval.If these pre-training approaches can be used for unsupervised domain adaptation for dense retrieval was so far unclear. In this work, we transfer the setup from Wang et al. (2021) to dense retrieval and first pre-train on the target corpus, followed by supervised training on labeled data from MS MARCO Nguyen et al. (2016). Performance is then measured on the target corpus. | Correct | null |
It is said that when evaluating a retriever trained on a source domain in an out-of-domain setting, the performance is obtained lower than BM25. Also, dense retrievers are said to be sensitive to domain shift and models that perform well on MS MARCO do not perform well on Sentimental-analysis data. There have been many studies on unsupervised sentence embedding learning, but it is said that they do not work well in unsupervised dense retrieval. Therefore, the performance of the retriever in out-of-domain may be worse. | However, as shown in Thakur et al. (2021b), dense retrieval methods require large amounts of training data to work well.333For reference, the popular MS MARCO dataset (Nguyen et al., 2016) has about 500k training instances; the Natural Questions dataset (Kwiatkowski et al., 2019) has more than 100k training instances. Most importantly, dense retrieval methods are extremely sensitive to domain shifts: Models trained on MS MARCO perform rather poorly for questions for COVID-19 scientific literature (Wang et al., 2020; Voorhees et al., 2021). The MS MARCO dataset was created before COVID-19, hence, it does not include any COVID-19 related topics and models did not learn how to represent this topic well in a vector space.We use the MS MARCO passage ranking dataset Nguyen et al. (2016) as the data from the source domain. It has 8.8M passages and 532.8K query-passage pairs labeled as relevant in the training set. As Table 1 shows, a state-of-the-art dense retrieval model, achieving an MRR@10 of 33.2 points on the MS MARCO passage ranking dataset, performs poorly on the six selected domain-specific retrieval datasets when compared to simple BM25 lexical search.So far, ICT and CD have only been studied on in-domain performance, i.e. a large in-domain labeled dataset is available which is used for subsequent supervised fine-tuning. SimCSE, CT, and TSDAE have been only studied for unsupervised sentence embedding learning. As our results show in Appendix E, they do not work at all for purely unsupervised dense retrieval.If these pre-training approaches can be used for unsupervised domain adaptation for dense retrieval was so far unclear. In this work, we transfer the setup from Wang et al. (2021) to dense retrieval and first pre-train on the target corpus, followed by supervised training on labeled data from MS MARCO Nguyen et al. (2016). Performance is then measured on the target corpus. | Incorrect | Change Concept |
Using lexical matching makes it difficult to identify synonyms or to distinguish between ambiguous words. | Information Retrieval (IR) is a central component of many natural language applications. Traditionally, lexical methods (Robertson et al., 1994) have been used to search through text content. However, these methods suffer from the lexical gap (Berger et al., 2000) and are not able to recognize synonyms and distinguish between ambiguous words. | Correct | null |
Using lexical matching makes it difficult to identify noun phrases or to distinguish between entities. | Information Retrieval (IR) is a central component of many natural language applications. Traditionally, lexical methods (Robertson et al., 1994) have been used to search through text content. However, these methods suffer from the lexical gap (Berger et al., 2000) and are not able to recognize synonyms and distinguish between ambiguous words. | Incorrect | Change Concept |
During indexing, we use another server with the same CPU and system memory specifications but which has four Titan V GPUs attached, each with 12 GiBs of memory. Across all experiments, only one GPU is dedicated per query for retrieval (i. , for methods with neural computations) but we use up to all four GPUs during indexing. | To evaluate the latency of neural re-ranking models in §4.2, we use a single Tesla V100 GPU that has 32 GiBs of memory on a server with two Intel Xeon Gold 6132 CPUs, each with 14 physical cores (24 hyperthreads), and 469 GiBs of RAM. For the mostly CPU-based retrieval experiments in §4.3 and the indexing experiments in §4.5, we use another server with the same CPU and system memory specifications but which has four Titan V GPUs attached, each with 12 GiBs of memory. Across all experiments, only one GPU is dedicated per query for retrieval (i.e., for methods with neural computations) but we use up to all four GPUs during indexing. | Correct | null |
During indexing, we use another server with the same CPU and system memory specifications but which has two Titan V GPUs attached, each with 8 GiBs of memory. Across all experiments, only one GPU is dedicated per query for retrieval (i. , for methods with neural computations) but we use up to all four GPUs during indexing. | To evaluate the latency of neural re-ranking models in §4.2, we use a single Tesla V100 GPU that has 32 GiBs of memory on a server with two Intel Xeon Gold 6132 CPUs, each with 14 physical cores (24 hyperthreads), and 469 GiBs of RAM. For the mostly CPU-based retrieval experiments in §4.3 and the indexing experiments in §4.5, we use another server with the same CPU and system memory specifications but which has four Titan V GPUs attached, each with 12 GiBs of memory. Across all experiments, only one GPU is dedicated per query for retrieval (i.e., for methods with neural computations) but we use up to all four GPUs during indexing. | Incorrect | Change Number |
They used (MRR@10) for measuring efficiency and effectiveness . | Diving deeper into the quality–cost tradeoff between BERT and ColBERT, Figure 4 demonstrates the relationships between FLOPs and effectiveness (MRR@10) as a function of the re-ranking depth k when re-ranking the top-k results by BM25, comparing ColBERT and BERT{}_{\textnormal{base}} (our training). We conduct this experiment on MS MARCO (Dev). We note here that as the official top-1000 ranking does not provide the BM25 order (and also lacks documents beyond the top-1000 per query), the models in this experiment re-rank the Anserini (Yanget al., 2018) toolkit’s BM25 output. Consequently, both MRR@10 values at k=1000 are slightly higher from those reported in Table 1. | Correct | null |
They used F1-score for measuring efficiency and effectiveness . | Diving deeper into the quality–cost tradeoff between BERT and ColBERT, Figure 4 demonstrates the relationships between FLOPs and effectiveness (MRR@10) as a function of the re-ranking depth k when re-ranking the top-k results by BM25, comparing ColBERT and BERT{}_{\textnormal{base}} (our training). We conduct this experiment on MS MARCO (Dev). We note here that as the official top-1000 ranking does not provide the BM25 order (and also lacks documents beyond the top-1000 per query), the models in this experiment re-rank the Anserini (Yanget al., 2018) toolkit’s BM25 output. Consequently, both MRR@10 values at k=1000 are slightly higher from those reported in Table 1. | Incorrect | Change Concept |
In fact, using ColBERT in the end-to-end setup is superior in terms of MRR@10 to re-ranking with the same model due to the improved recall. Moving beyond MRR@10, large gains in Recall@k for k equals to 50, 200, and 1000. For instance, its Recall@50 actually exceeds the official BM25’s Recall@1000 and even all but docTTTTTquery’s Recall@200, emphasizing the value of end-to-end retrieval (instead of just re-ranking) with ColBERT. | Shifting our attention to ColBERT’s end-to-end retrieval effectiveness, we see its major gains in MRR@10 over all of these end-to-end models. In fact, using ColBERT in the end-to-end setup is superior in terms of MRR@10 to re-ranking with the same model due to the improved recall. Moving beyond MRR@10, we also see large gains in Recall@k for k equals to 50, 200, and 1000. For instance, its Recall@50 actually exceeds the official BM25’s Recall@1000 and even all but docTTTTTquery’s Recall@200, emphasizing the value of end-to-end retrieval (instead of just re-ranking) with ColBERT. | Correct | null |
In fact, using ColBERT in the end-to-end setup is not superior in terms of MRR@10 to re-ranking with the same model due to the decreased recall. Moving beyond MRR@10, large gains in Recall@k for k equals to 50, 200, and 1000. For instance, its Recall@50 actually exceeds the official BM25’s Recall@1000 and even all but docTTTTTquery’s Recall@200, emphasizing the value of re-ranking (instead of just end-to-end retrieval) with ColBERT. | Shifting our attention to ColBERT’s end-to-end retrieval effectiveness, we see its major gains in MRR@10 over all of these end-to-end models. In fact, using ColBERT in the end-to-end setup is superior in terms of MRR@10 to re-ranking with the same model due to the improved recall. Moving beyond MRR@10, we also see large gains in Recall@k for k equals to 50, 200, and 1000. For instance, its Recall@50 actually exceeds the official BM25’s Recall@1000 and even all but docTTTTTquery’s Recall@200, emphasizing the value of end-to-end retrieval (instead of just re-ranking) with ColBERT. | Incorrect | Opposite |
In contrast with this trend, ColBERT (which employs late interaction over BERT performs no worse than the original adaptation of BERT for ranking and is only marginally less effective than BERT and our training of BERT. While highly competitive in effectiveness, ColBERT is orders of magnitude cheaper than BERT, in particular, by over 170\times in latency and 13,900\times in FLOPs. | In contrast with this trend, ColBERT (which employs late interaction over BERT{}_{\textnormal{base}}) performs no worse than the original adaptation of BERT{}_{\textnormal{base}} for ranking by Nogueira and Cho (Nogueira and Cho, 2019; Nogueiraet al., 2019b) and is only marginally less effective than BERT{}_{\textnormal{large}} and our training of BERT{}_{\textnormal{base}} (described above). While highly competitive in effectiveness, ColBERT is orders of magnitude cheaper than BERT{}_{\textnormal{base}}, in particular, by over 170\times in latency and 13,900\times in FLOPs. This highlights the expressiveness of our proposed late interaction mechanism, particularly when coupled with a powerful pre-trained LM like BERT. While ColBERT’s re-ranking latency is slightly higher than the non-BERT re-ranking models shown (i.e., by 10s of milliseconds), this difference is explained by the time it takes to gather, stack, and transfer the document embeddings to the GPU. In particular, the query encoding and interaction in ColBERT consume only 13 milliseconds of its total execution time. We note that ColBERT’s latency and FLOPs can be considerably reduced by padding queries to a shorter length, using smaller vector dimensions (the MRR@10 of which is tested in §4.5), employing quantization of the document vectors, and storing the embeddings on GPU if sufficient memory exists. We leave these directions for future work. | Correct | null |
In contrast with this trend, ColBERT (which employs late interaction over BERT performs no worse than the original adaptation of BERT for ranking and is only marginally less effective than BERT and our training of BERT. While highly competitive in effectiveness, ColBERT is orders of magnitude more expensive than BERT, in particular, by over 170\times in latency and 13,900\times in FLOPs. | In contrast with this trend, ColBERT (which employs late interaction over BERT{}_{\textnormal{base}}) performs no worse than the original adaptation of BERT{}_{\textnormal{base}} for ranking by Nogueira and Cho (Nogueira and Cho, 2019; Nogueiraet al., 2019b) and is only marginally less effective than BERT{}_{\textnormal{large}} and our training of BERT{}_{\textnormal{base}} (described above). While highly competitive in effectiveness, ColBERT is orders of magnitude cheaper than BERT{}_{\textnormal{base}}, in particular, by over 170\times in latency and 13,900\times in FLOPs. This highlights the expressiveness of our proposed late interaction mechanism, particularly when coupled with a powerful pre-trained LM like BERT. While ColBERT’s re-ranking latency is slightly higher than the non-BERT re-ranking models shown (i.e., by 10s of milliseconds), this difference is explained by the time it takes to gather, stack, and transfer the document embeddings to the GPU. In particular, the query encoding and interaction in ColBERT consume only 13 milliseconds of its total execution time. We note that ColBERT’s latency and FLOPs can be considerably reduced by padding queries to a shorter length, using smaller vector dimensions (the MRR@10 of which is tested in §4.5), employing quantization of the document vectors, and storing the embeddings on GPU if sufficient memory exists. We leave these directions for future work. | Incorrect | Change Concept |
creating a large training corpus is often time-consuming and expensive and hence many retrieval systems are applied in a zero-shot setup, with no available training data to train the system. Hence, a zero-shot scenario in this context refer to cases where relevance annotations are not available and does not refer to unavailability of query set. | However, creating a large training corpus is often time-consuming and expensive and hence many retrieval systems are applied in a zero-shot setup, with no available training data to train the system. So far, it is unclear how well existing trained neural models will perform for other text domains or textual retrieval tasks. Even more important, it is unclear how well different approaches, like sparse embeddings vs. dense embeddings, generalize to out-of-distribution data. | Correct | null |
creating a large training corpus is often time-consuming and expensive and hence many retrieval systems are applied in a zero-shot setup, with no available training data to train the system. Hence, a zero-shot scenario in this context refer to cases where query set are not available. | However, creating a large training corpus is often time-consuming and expensive and hence many retrieval systems are applied in a zero-shot setup, with no available training data to train the system. So far, it is unclear how well existing trained neural models will perform for other text domains or textual retrieval tasks. Even more important, it is unclear how well different approaches, like sparse embeddings vs. dense embeddings, generalize to out-of-distribution data. | Incorrect | Change Concept |
Lexical approaches like BM25 and docT5query have a rather low Hole@10 value of 6. 4% and 2. 8%, indicating that the annotation pool contained the top-hits from lexical retrieval systems. In contrast, dense retrieval systems like ANCE and TAS-B have a much higher Hole@10 of 14. 4% and 31. | The results reveal large differences between approaches: Lexical approaches like BM25 and docT5query have a rather low Hole@10 value of 6.4% and 2.8%, indicating that the annotation pool contained the top-hits from lexical retrieval systems. In contrast, dense retrieval systems like ANCE and TAS-B have a much higher Hole@10 of 14.4% and 31.8%, indicating that a large fraction of hits found by these systems have not been judged by annotators.Next, we manually added for all systems, the missing annotation (or holes) following the original annotation guidelines. During annotation, we were unaware of the system who retrieved the missing annotation to avoid a preference bias. In total, we annotated 980 query-document pairs in TREC-COVID. We then re-computed nDCG@10 for all systems with this additional annotations. | Correct | null |
Lexical approaches like BM25 and docT5query have a rather low Hole@10 value of 64% and 28%, indicating that the annotation pool contained the top-hits from lexical retrieval systems. In contrast, dense retrieval systems like ANCE and TAS-B have a much higher Hole@10 of 14. 4% and 31. | The results reveal large differences between approaches: Lexical approaches like BM25 and docT5query have a rather low Hole@10 value of 6.4% and 2.8%, indicating that the annotation pool contained the top-hits from lexical retrieval systems. In contrast, dense retrieval systems like ANCE and TAS-B have a much higher Hole@10 of 14.4% and 31.8%, indicating that a large fraction of hits found by these systems have not been judged by annotators.Next, we manually added for all systems, the missing annotation (or holes) following the original annotation guidelines. During annotation, we were unaware of the system who retrieved the missing annotation to avoid a preference bias. In total, we annotated 980 query-document pairs in TREC-COVID. We then re-computed nDCG@10 for all systems with this additional annotations. | Incorrect | Change Number |
According to the authors, K as a hyperparameter is only best believed as the number of user intents and does not necessarily equal the actual number of user intents. | The main goal of next item prediction task is to optimizeEq. (1).Assume that there are also K different user intents (e.g., purchasing holiday gifts, preparing for fishing activity, etc.)in a recommender system that formsthe intent variable c=\left\{c_{i}\right\}_{i=1}^{K}, thenthe probability of a user interacting with a certainitem can be rewritten as follows:(7)\begin{split}P_{\theta}(s^{u})=\mathbb{E}_{(c)}\left[P_{\theta}(s^{u},c)\right].\end{split}However, users intents are latent by definition.Because of the missing observation of variable c,we are in a ‘chicken-and-eggs’ situation thatwithout c, we cannot estimate parameter \theta,and without \theta we cannot inferwhat the value of c might be.The larger of the intent class number K means users can havemore diverseintentions.The larger value of the strength of SeqCL objective \betameans the ICL task contributes more tothe final model.The results on Yelp is shown in Figure 5.We find that: (1)ICLRec reaches itsbest performance when increasing K to 512,and then it starts to deteriorateas K become larger.When K is very small,the number of users undereach intent prototype can potentially be large.As a result, false-positive samples(i.e., users that actually have different intentsare considered as having the same intent erroneously)are introduced to the contrastive SSL,thus affecting learning.On the other hand, when K is too large,the number of users undereach intent prototype is small,the introduced false-negative sampleswill also impair contrastive SSL.In Yelp, 512 user intents summarizeusers’ distinct behaviors best.(2) A ‘sweet-spot’ of \lambda=0.5 canalso be found.It indicatesthat the ICL task can benefitthe recommendation predictionas an auxiliary task.The impact of the batch size and \beta are provided in Appendix D.Recently, many approaches have been proposed to studyusers’ intents forimproving recommendations (Wanget al., 2019b; Cenet al., 2020; Li et al., 2019; Liet al., 2021b).MCPRN (Wanget al., 2019b)designs mixture-channel purposerouting networks to adaptivelylearnusers’ different purchase purposesof each itemunder different channels (sub-sequences) for session-based recommendation.MITGNN(Liuet al., 2020a)proposes amulti-intenttranslation graph neural networkto mine users’ multiple intentsby considering the correlations of the intents.ICM-SR (Panet al., 2020)designs anintent-guided neighbor detectorto retrieve correctneighbor sessionsfor neighbor representation.Different from session-based recommendation,another line of worksfocus on modeling the sequentialdynamics of users’ interaction behaviorsin a longer time span.DSSRec (Maet al., 2020)proposes a seq2seq trainingstrategy using multiple future interactions as supervision and introducing an intent variable from her historical and future behavior sequences.The intent variable is used to capture mutual information between an individual user’s historical and future behavior sequences.Two users of similar intentsmight be far away in representation space.Unlike this work, our intent variable is learned over all users’ sequences and is used to maximize mutual information across different users with similar learned intents.ASLI (Tanjim et al., 2020)captures intentvia a temporal convolutionalnetwork with side information (e.g., user action types such asclick, add-to-favorite, etc.),and then use the learnedintents toguide SR model to predictthe next item.Instead, our methodcan learn users’ intentsbased on user interactiondata only. | Correct | null |
According to the authors, K as a hyperparameter actually equals the actual number of user intents. | The main goal of next item prediction task is to optimizeEq. (1).Assume that there are also K different user intents (e.g., purchasing holiday gifts, preparing for fishing activity, etc.)in a recommender system that formsthe intent variable c=\left\{c_{i}\right\}_{i=1}^{K}, thenthe probability of a user interacting with a certainitem can be rewritten as follows:(7)\begin{split}P_{\theta}(s^{u})=\mathbb{E}_{(c)}\left[P_{\theta}(s^{u},c)\right].\end{split}However, users intents are latent by definition.Because of the missing observation of variable c,we are in a ‘chicken-and-eggs’ situation thatwithout c, we cannot estimate parameter \theta,and without \theta we cannot inferwhat the value of c might be.The larger of the intent class number K means users can havemore diverseintentions.The larger value of the strength of SeqCL objective \betameans the ICL task contributes more tothe final model.The results on Yelp is shown in Figure 5.We find that: (1)ICLRec reaches itsbest performance when increasing K to 512,and then it starts to deteriorateas K become larger.When K is very small,the number of users undereach intent prototype can potentially be large.As a result, false-positive samples(i.e., users that actually have different intentsare considered as having the same intent erroneously)are introduced to the contrastive SSL,thus affecting learning.On the other hand, when K is too large,the number of users undereach intent prototype is small,the introduced false-negative sampleswill also impair contrastive SSL.In Yelp, 512 user intents summarizeusers’ distinct behaviors best.(2) A ‘sweet-spot’ of \lambda=0.5 canalso be found.It indicatesthat the ICL task can benefitthe recommendation predictionas an auxiliary task.The impact of the batch size and \beta are provided in Appendix D.Recently, many approaches have been proposed to studyusers’ intents forimproving recommendations (Wanget al., 2019b; Cenet al., 2020; Li et al., 2019; Liet al., 2021b).MCPRN (Wanget al., 2019b)designs mixture-channel purposerouting networks to adaptivelylearnusers’ different purchase purposesof each itemunder different channels (sub-sequences) for session-based recommendation.MITGNN(Liuet al., 2020a)proposes amulti-intenttranslation graph neural networkto mine users’ multiple intentsby considering the correlations of the intents.ICM-SR (Panet al., 2020)designs anintent-guided neighbor detectorto retrieve correctneighbor sessionsfor neighbor representation.Different from session-based recommendation,another line of worksfocus on modeling the sequentialdynamics of users’ interaction behaviorsin a longer time span.DSSRec (Maet al., 2020)proposes a seq2seq trainingstrategy using multiple future interactions as supervision and introducing an intent variable from her historical and future behavior sequences.The intent variable is used to capture mutual information between an individual user’s historical and future behavior sequences.Two users of similar intentsmight be far away in representation space.Unlike this work, our intent variable is learned over all users’ sequences and is used to maximize mutual information across different users with similar learned intents.ASLI (Tanjim et al., 2020)captures intentvia a temporal convolutionalnetwork with side information (e.g., user action types such asclick, add-to-favorite, etc.),and then use the learnedintents toguide SR model to predictthe next item.Instead, our methodcan learn users’ intentsbased on user interactiondata only. | Incorrect | Change Concept |
The authors propose the EM framework as it guarantees convergence | Later, we will show that a generalized Expectation-Maximization frameworkprovides a direction to address above problemwith a convergence guarantee.The basic idea of optimizing Eq. (7) via EMis to start with an initial guessof the model parameter \thetaand estimate the expected valuesof the missing variable c, i.e., the E-step.And once we have the values of c,we can maximize the Eq. (7) w.r.t theparameter \theta, i.e., the M step.We can repeat this iterative process until the likelihood cannot increase anymore.To discover the benefits oflatent intentsand address challenges,we propose theIntent Contrastive Learning (ICL),a general learning paradigm thatleverages the latent intentfactor into SR.It learns users’ intentdistributionsfrom all user behavior sequencesvia clustering.And it leveragesthe learnt intentsinto the SR modelvia a new contrastive SSL,whichmaximizes the agreementbetween a view of sequenceand its corresponding intent.The intent representation learning moduleand the contrastive SSL module are mutually reinforcedto train a more expressivesequence encoder.We tackle the challenge of intentmining problem byintroducing alatent variable to represent users’ intentsand learn them alternatelyalong with the SR model optimization throughan expectation-maximization (EM) frameworkto ensure convergence.We suggest fusing learnt intent informationinto SR via the proposed contrastive SSL,as it can improve model’s performance as wellas robustness.Extensive experiments conducted on four real-world datasetsfurther verify the effectiveness of the proposed learning paradigm,which improves performance and robustness,even when recommender systemsfaceheavy data sparsity issues.In this work,we propose a new learning paradigm ICL that canmodel latent intent factorsfrom user interactionsand fuse them into a sequential recommendationmodel via a newcontrastive SSL objective.ICL is formulated withinan EM framework, which guaranteesconvergence. Detailed analyses showthe superiority of ICL andexperiments conducted onfour datasets further demonstratethe effectiveness of the proposed method. | Correct | null |
The authors didn't propose the EM framework as it causes convergence | Later, we will show that a generalized Expectation-Maximization frameworkprovides a direction to address above problemwith a convergence guarantee.The basic idea of optimizing Eq. (7) via EMis to start with an initial guessof the model parameter \thetaand estimate the expected valuesof the missing variable c, i.e., the E-step.And once we have the values of c,we can maximize the Eq. (7) w.r.t theparameter \theta, i.e., the M step.We can repeat this iterative process until the likelihood cannot increase anymore.To discover the benefits oflatent intentsand address challenges,we propose theIntent Contrastive Learning (ICL),a general learning paradigm thatleverages the latent intentfactor into SR.It learns users’ intentdistributionsfrom all user behavior sequencesvia clustering.And it leveragesthe learnt intentsinto the SR modelvia a new contrastive SSL,whichmaximizes the agreementbetween a view of sequenceand its corresponding intent.The intent representation learning moduleand the contrastive SSL module are mutually reinforcedto train a more expressivesequence encoder.We tackle the challenge of intentmining problem byintroducing alatent variable to represent users’ intentsand learn them alternatelyalong with the SR model optimization throughan expectation-maximization (EM) frameworkto ensure convergence.We suggest fusing learnt intent informationinto SR via the proposed contrastive SSL,as it can improve model’s performance as wellas robustness.Extensive experiments conducted on four real-world datasetsfurther verify the effectiveness of the proposed learning paradigm,which improves performance and robustness,even when recommender systemsfaceheavy data sparsity issues.In this work,we propose a new learning paradigm ICL that canmodel latent intent factorsfrom user interactionsand fuse them into a sequential recommendationmodel via a newcontrastive SSL objective.ICL is formulated withinan EM framework, which guaranteesconvergence. Detailed analyses showthe superiority of ICL andexperiments conducted onfour datasets further demonstratethe effectiveness of the proposed method. | Incorrect | Opposite |
BUIR requires positive user-item pairs instead of negative sampling for training. | For all the datasets, BUIRid shows the substantially higher performance than the discriminative methods taking only user-id/item-id (i.e., BPR, NeuMF, CML, and SML).In particular, the sparser the training set becomes, the larger the performance improvement of BUIRid is achieved over the best baseline (denoted by Improvid).It is obvious that BUIRid is more robust to the extreme sparsity compared to the other baselines that are more likely to explicitly use “positive but unobserved” interactions as negative interactions when positive user-item interactions are more rarely observed.BUIRid is not affected by such inconsistent supervision from uncertain negative interactions because it directly optimizes the representations of users and items by using only positive interactions.As a solution to the aforementioned limitations, this paper proposes a novel OCCF framework, named as BUIR, which does not require the negative sampling at all for training the model.The main idea is, given a positive user-item interaction (u, v), to make representations for u and v similar to each other, in order to encode the preference information into the representations.However, a naive end-to-end learning framework that guides positive user-item pairs to be similar to each other without any negative supervision can easily converge to a collapsed solution – the encoder network outputs the same representations for all the users and items.First of all, the BPR framework that optimizes the cross-prediction score, q\left(f(u)\right)^{\top}f(v)+f(u)^{\top}q\left(f(v)\right), is not as effective as ours;it is even worse compared to the conventional BPR, which optimizes the inner-product score f(u)^{\top}f(v).This implies that the performance improvement of BUIR is mainly caused by our learning framework rather than its score modeling based on the predictor.In addition, even without the stochastic augmentation, the neighbor-based encoder (i.e., LGCN) based on the BUIR framework beats LGCN based on the BPR framework, which demonstrates that BUIR successfully addresses the issue of incorrect negative sampling.Lastly, our framework with the stochastic neighbor augmentation further improves the performance by taking benefits from various views of the positive user-item interactions for the optimization.This paper proposes a novel framework for learning the representations of users and items, termed as BUIR, to address the main challenges of the OCCF problem: the implicit assumption about negative interactions, and high sparsity of observed (positively-labeled) interactions.First, BUIR directly bootstraps the representations of users and items by minimizing their cross-prediction error.This makes BUIR use only partially-observed positive interactions for training the model, and accordingly, it can eliminate the need for negative sampling.In addition, BUIR is able to learn the augmented views of each positive interaction obtained from the neighborhood information, which further relieves the data sparsity issue of the OCCF problem.Through the extensive comparison with a wide range of OCCF methods, we demonstrate that BUIR consistently outperforms all the other baselines in terms of top-K recommendation.In particular, the effectiveness of BUIR becomes more significant for much sparse datasets in which the positively-labeled interactions are not enough to optimize the model as well as the assumption about negative interactions becomes less valid.Based on its great compatibility with existing user/item encoder networks, we expect that our BUIR framework can be a major solution for the OCCF problem, replacing the conventional BPR framework.Our extensive evaluation on real-world implicit feedback datasets shows that BUIR consistently performs the best for top-K recommendation among a wide range of OCCF methods.In particular, the performance improvement becomes more significant in sparser datasets, with the help of utilizing augmented views of positive interactions as well as eliminating the effect of uncertain negative interactions.In addition, comparison results on a downstream task, which classifies the items into their category, support that BUIR learns more effective representations than other OCCF baselines. | Correct | null |
BUIR requires negative sampling instead of positive user-item pairs for training. | For all the datasets, BUIRid shows the substantially higher performance than the discriminative methods taking only user-id/item-id (i.e., BPR, NeuMF, CML, and SML).In particular, the sparser the training set becomes, the larger the performance improvement of BUIRid is achieved over the best baseline (denoted by Improvid).It is obvious that BUIRid is more robust to the extreme sparsity compared to the other baselines that are more likely to explicitly use “positive but unobserved” interactions as negative interactions when positive user-item interactions are more rarely observed.BUIRid is not affected by such inconsistent supervision from uncertain negative interactions because it directly optimizes the representations of users and items by using only positive interactions.As a solution to the aforementioned limitations, this paper proposes a novel OCCF framework, named as BUIR, which does not require the negative sampling at all for training the model.The main idea is, given a positive user-item interaction (u, v), to make representations for u and v similar to each other, in order to encode the preference information into the representations.However, a naive end-to-end learning framework that guides positive user-item pairs to be similar to each other without any negative supervision can easily converge to a collapsed solution – the encoder network outputs the same representations for all the users and items.First of all, the BPR framework that optimizes the cross-prediction score, q\left(f(u)\right)^{\top}f(v)+f(u)^{\top}q\left(f(v)\right), is not as effective as ours;it is even worse compared to the conventional BPR, which optimizes the inner-product score f(u)^{\top}f(v).This implies that the performance improvement of BUIR is mainly caused by our learning framework rather than its score modeling based on the predictor.In addition, even without the stochastic augmentation, the neighbor-based encoder (i.e., LGCN) based on the BUIR framework beats LGCN based on the BPR framework, which demonstrates that BUIR successfully addresses the issue of incorrect negative sampling.Lastly, our framework with the stochastic neighbor augmentation further improves the performance by taking benefits from various views of the positive user-item interactions for the optimization.This paper proposes a novel framework for learning the representations of users and items, termed as BUIR, to address the main challenges of the OCCF problem: the implicit assumption about negative interactions, and high sparsity of observed (positively-labeled) interactions.First, BUIR directly bootstraps the representations of users and items by minimizing their cross-prediction error.This makes BUIR use only partially-observed positive interactions for training the model, and accordingly, it can eliminate the need for negative sampling.In addition, BUIR is able to learn the augmented views of each positive interaction obtained from the neighborhood information, which further relieves the data sparsity issue of the OCCF problem.Through the extensive comparison with a wide range of OCCF methods, we demonstrate that BUIR consistently outperforms all the other baselines in terms of top-K recommendation.In particular, the effectiveness of BUIR becomes more significant for much sparse datasets in which the positively-labeled interactions are not enough to optimize the model as well as the assumption about negative interactions becomes less valid.Based on its great compatibility with existing user/item encoder networks, we expect that our BUIR framework can be a major solution for the OCCF problem, replacing the conventional BPR framework.Our extensive evaluation on real-world implicit feedback datasets shows that BUIR consistently performs the best for top-K recommendation among a wide range of OCCF methods.In particular, the performance improvement becomes more significant in sparser datasets, with the help of utilizing augmented views of positive interactions as well as eliminating the effect of uncertain negative interactions.In addition, comparison results on a downstream task, which classifies the items into their category, support that BUIR learns more effective representations than other OCCF baselines. | Incorrect | Change Concept |
Approximating the online encoder keep the target encoder from converging to the collapsed solution. | Bypassing the collapsed solution. It is obvious that the loss in Equation (3) admits the collapsed solution with respect to \theta and \xi, which means both the encoders generate the same representations for all users and items.For this reason, the conventional end-to-end learning strategy, which optimizes both f_{\theta} and f_{\xi} to minimize the loss (i.e., cross-prediction error), may easily lead to such collapsed solution.In contrast, our proposed framework updates each of the encoders in different ways.From Equation (4), the online encoder is optimized to minimize the loss, while the target encoder is updated to slowly approximate the online encoder.That is, the direction of updating the target encoder (\theta-\xi) totally differs from that of updating the online encoder (-\nabla_{\theta}\mathcal{L}_{\theta,\xi}),and this effectively keeps both the encoders from converging to the collapsed solution.Several recent work on bootstrapping-based representation learning (Grill et al., 2020; Chen and He, 2021) empirically demonstrated that this kind of dynamics (i.e., updating two networks differently) allows to avoid the collapsed solution without any explicit term to prevent it. | Correct | null |
The online encoder doesn't keep the target encoder from converging to the collapsed solution. | Bypassing the collapsed solution. It is obvious that the loss in Equation (3) admits the collapsed solution with respect to \theta and \xi, which means both the encoders generate the same representations for all users and items.For this reason, the conventional end-to-end learning strategy, which optimizes both f_{\theta} and f_{\xi} to minimize the loss (i.e., cross-prediction error), may easily lead to such collapsed solution.In contrast, our proposed framework updates each of the encoders in different ways.From Equation (4), the online encoder is optimized to minimize the loss, while the target encoder is updated to slowly approximate the online encoder.That is, the direction of updating the target encoder (\theta-\xi) totally differs from that of updating the online encoder (-\nabla_{\theta}\mathcal{L}_{\theta,\xi}),and this effectively keeps both the encoders from converging to the collapsed solution.Several recent work on bootstrapping-based representation learning (Grill et al., 2020; Chen and He, 2021) empirically demonstrated that this kind of dynamics (i.e., updating two networks differently) allows to avoid the collapsed solution without any explicit term to prevent it. | Incorrect | Opposite |
Online encoders prevent models from collapsing into trivial solutions without explicitly using negative interactions for optimization. | We argue that the above collapsed solution is incurred by the si\x02multaneous optimization of 𝑢 and 𝑣 within the end-to-end learning framework of a single encoder. Hence, we instead adopt the student\x02teacher-like network [6, 29] in which only the student’s output 𝑢 (and 𝑣) is optimized to predict the target 𝑣 (and 𝑢) presented by the teacher. Specifically, BUIR directly bootstraps1 the representations of users and items by employing two distinct encoder networks, referred to as online encoder and target encoder. The high-level idea is training only the online encoder for the prediction task between 𝑢 and 𝑣, where the target for its prediction is provided by the target encoder. That is, the online encoder is optimized so that its user (and item) vectors get closer to the item (and user) vectors com\x02puted by the target encoder. At the same time, the target encoder is updated based on momentum-based moving average [6, 8, 29] to slowly approximate the online encoder, which encourages to pro\x02vide enhanced representations as the target for the online encoder. By doing so, the online encoder can capture the positive relation\x02ship between 𝑢 and 𝑣 into the representations, while preventing the model from collapsing to the trivial solution without explicitly using any negative interactions for the optimization. | Correct | null |
Online encoders allow models collapsing into trivial solutions without explicitly using negative interactions for optimization. | We argue that the above collapsed solution is incurred by the si\x02multaneous optimization of 𝑢 and 𝑣 within the end-to-end learning framework of a single encoder. Hence, we instead adopt the student\x02teacher-like network [6, 29] in which only the student’s output 𝑢 (and 𝑣) is optimized to predict the target 𝑣 (and 𝑢) presented by the teacher. Specifically, BUIR directly bootstraps1 the representations of users and items by employing two distinct encoder networks, referred to as online encoder and target encoder. The high-level idea is training only the online encoder for the prediction task between 𝑢 and 𝑣, where the target for its prediction is provided by the target encoder. That is, the online encoder is optimized so that its user (and item) vectors get closer to the item (and user) vectors com\x02puted by the target encoder. At the same time, the target encoder is updated based on momentum-based moving average [6, 8, 29] to slowly approximate the online encoder, which encourages to pro\x02vide enhanced representations as the target for the online encoder. By doing so, the online encoder can capture the positive relation\x02ship between 𝑢 and 𝑣 into the representations, while preventing the model from collapsing to the trivial solution without explicitly using any negative interactions for the optimization. | Incorrect | Opposite |
Using predictor can optimize the representation without any negative sample. | Existing discriminative OCCF methods (Rendle et al., 2009; Hsieh et al., 2017) have tried to optimize the latent space where the user-item interactions are directly encoded into their inner product (or Euclidean distance).On the contrary, BUIR additionally uses the predictor to model their interaction, which results in the capability of encoding the high-level relationship between users and items into the representations.In conclusion, with the help of the predictor, BUIR accurately computes the user-item interaction scores as well as optimizes the representation without explicitly using negative samples. | Correct | null |
Using predictor can optimize the representation with only a few negative samples. | Existing discriminative OCCF methods (Rendle et al., 2009; Hsieh et al., 2017) have tried to optimize the latent space where the user-item interactions are directly encoded into their inner product (or Euclidean distance).On the contrary, BUIR additionally uses the predictor to model their interaction, which results in the capability of encoding the high-level relationship between users and items into the representations.In conclusion, with the help of the predictor, BUIR accurately computes the user-item interaction scores as well as optimizes the representation without explicitly using negative samples. | Incorrect | Change Concept |
Stochastic means it use random neighborhood information of each user and item during data augmentation. | In Figure 4, our stochastic data augmentation (i.e., P>0) brings a significant improvement compared to the case of using the fixed neighborhood information (i.e., P=0) as encoder inputs.This result shows that the augmented views of positive interactions encourage BUIR to effectively learn users’ preference on items even in much sparse dataset.Interestingly, in case of the Ciao dataset which is less sparse than CiteULike, the benefit of our augmentation linearly increases with the maximum drop probability.This is because there is room for producing more various views (i.e., larger perturbation) based on a relatively more number of neighbors, and it eventually helps to boost the recommendation performance.To sum up, our framework that adopts the neighbor augmentation function successfully relieves the data sparsity issue of the OCCF problem, by leveraging the augmented views of few positive interactions.Furthermore, we introduce a stochastic data augmentation technique to relieve the data sparsity problem in our framework.Motivated by the recent success of self-supervised learning in various domains (Chenet al., 2020; Devlinet al., 2019), we exploit augmented views of an input interaction, which are generated based on the neighborhood information of each user and item (i.e., the set of the items interacted with a user, and the users interacted with an item).The stochastic augmentation is applied to positive user-item pairs when they are passed to the encoder, so as to produce the different views of the pairs.To be precise, by making our encoder use a random subset of a user’s (and item’s) neighbors for the input features, it produces a similar effect to increasing the number of positive pairs from the data itself without any human intervention.In the end, BUIR is allowed to learn various views of each positive user-item pair. | Correct | null |
Stochastic means it use pre-selected neighborhood information of each user and item during data augmentation. | In Figure 4, our stochastic data augmentation (i.e., P>0) brings a significant improvement compared to the case of using the fixed neighborhood information (i.e., P=0) as encoder inputs.This result shows that the augmented views of positive interactions encourage BUIR to effectively learn users’ preference on items even in much sparse dataset.Interestingly, in case of the Ciao dataset which is less sparse than CiteULike, the benefit of our augmentation linearly increases with the maximum drop probability.This is because there is room for producing more various views (i.e., larger perturbation) based on a relatively more number of neighbors, and it eventually helps to boost the recommendation performance.To sum up, our framework that adopts the neighbor augmentation function successfully relieves the data sparsity issue of the OCCF problem, by leveraging the augmented views of few positive interactions.Furthermore, we introduce a stochastic data augmentation technique to relieve the data sparsity problem in our framework.Motivated by the recent success of self-supervised learning in various domains (Chenet al., 2020; Devlinet al., 2019), we exploit augmented views of an input interaction, which are generated based on the neighborhood information of each user and item (i.e., the set of the items interacted with a user, and the users interacted with an item).The stochastic augmentation is applied to positive user-item pairs when they are passed to the encoder, so as to produce the different views of the pairs.To be precise, by making our encoder use a random subset of a user’s (and item’s) neighbors for the input features, it produces a similar effect to increasing the number of positive pairs from the data itself without any human intervention.In the end, BUIR is allowed to learn various views of each positive user-item pair. | Incorrect | Change Concept |
Model gets best performance when the value of parameter tau is larger or equal than 0. 9 and smaller than 1. | Implementation Details. We implement the proposed framework and all the baselines by using PyTorch, and use the Adam optimizer to train them.For BUIR, we fix the momentum coefficient \tau to 0.995, and adopt a single linear layer for the predictor q_{\theta}.666We empirically found that these hyperparameters hardly affect the final performance of BUIR, and the sensitivity analysis on the parameters is provided in Section 4.6.The augmentation function \psi simply uses a uniform distribution for drawing a drop probability p\sim\mathcal{U}(0,1), where each user’s (item’s) neighbor is independently deleted with the probability p.Figure 6 clearly shows that the performance is hardly affected by \tau in the range of [0.9, 1.0).In other words, any values of \tau larger than 0.9 allow the target encoder to successfully provide the target representations to the online encoder, by slowly approximating the online encoder;on the contrary, BUIR cannot learn the effective representations at all in case that the target encoder is fixed (i.e., \tau=1).This observation is consistent with previous work on momentum-based moving average (Tarvainen andValpola, 2017; Heet al., 2020b; Grill et al., 2020) that showed all values of \tau between 0.9 and 0.999 can yield the best performance.Furthermore, BUIR performs the best with a single-layer predictor, because a multi-layer predictor makes it difficult to optimize the relationship between outputs of the two encoder networks.In conclusion, BUIR is more powerful even with fewer hyperparameters, compared to existing OCCF methods that include a variety of regularization terms or modeling components. | Correct | null |
Model gets best performance when the value of parameter tau is larger or equal than 0. 5 and smaller than 0.9. | Implementation Details. We implement the proposed framework and all the baselines by using PyTorch, and use the Adam optimizer to train them.For BUIR, we fix the momentum coefficient \tau to 0.995, and adopt a single linear layer for the predictor q_{\theta}.666We empirically found that these hyperparameters hardly affect the final performance of BUIR, and the sensitivity analysis on the parameters is provided in Section 4.6.The augmentation function \psi simply uses a uniform distribution for drawing a drop probability p\sim\mathcal{U}(0,1), where each user’s (item’s) neighbor is independently deleted with the probability p.Figure 6 clearly shows that the performance is hardly affected by \tau in the range of [0.9, 1.0).In other words, any values of \tau larger than 0.9 allow the target encoder to successfully provide the target representations to the online encoder, by slowly approximating the online encoder;on the contrary, BUIR cannot learn the effective representations at all in case that the target encoder is fixed (i.e., \tau=1).This observation is consistent with previous work on momentum-based moving average (Tarvainen andValpola, 2017; Heet al., 2020b; Grill et al., 2020) that showed all values of \tau between 0.9 and 0.999 can yield the best performance.Furthermore, BUIR performs the best with a single-layer predictor, because a multi-layer predictor makes it difficult to optimize the relationship between outputs of the two encoder networks.In conclusion, BUIR is more powerful even with fewer hyperparameters, compared to existing OCCF methods that include a variety of regularization terms or modeling components. | Incorrect | Change Number |
The online encoder is updated to minimize the error between the output and the target and updated by the gradients back-propagated from the loss, but target network is updated based on the momentum update and updated as the moving average of the online encoder . | BUIR makes use of two distinct encoder networks that have the same structure: online encoder f_{\theta} and target encoder f_{\xi}.They are parameterized by \theta and \xi, respectively.The key idea of BUIR is to train the online encoder by using outputs of the target encoder as its target, while gradually improving the target encoder as well.The main difference of BUIR from existing end-to-end learning frameworks is that f_{\theta} and f_{\xi} are updated in different ways.The online encoder is trained to minimize the error between its output and the target, whereas the target network is slowly updated based on the momentum update (Heet al., 2020b) so as to keep its output consistent.To sum up, the parameters of the online encoder and target encoder are optimized by(4)\begin{split}\theta&\leftarrow\theta-\eta\cdot\nabla_{\theta}\mathcal{L}_{\theta,\xi}\\\xi&\leftarrow\tau\cdot\xi+(1-\tau)\cdot\theta.\end{split}\eta is the learning rate for stochastic optimization, and \tau\in[0,1] is a momentum coefficient (also called as target decay) for momentum-based moving average.The online encoder f_{\theta} (and the predictor q_{\theta}) is effectively optimized by the gradients back-propagated from the loss (Equation (3)), while the target encoder f_{\xi} is updated as the moving average of the online encoder.By taking a large value of \tau, the target encoder slowly approximates the online encoder.This momentum-based update makes \xi evolve more slowly than \theta, which enables to bootstrap the representations by providing enhanced but consistent targets to the online encoders (Heet al., 2020b; Grill et al., 2020).Figure 1 illustrates the overall framework of BUIR with the simple one-hot encoders.Bypassing the collapsed solution. It is obvious that the loss in Equation (3) admits the collapsed solution with respect to \theta and \xi, which means both the encoders generate the same representations for all users and items.For this reason, the conventional end-to-end learning strategy, which optimizes both f_{\theta} and f_{\xi} to minimize the loss (i.e., cross-prediction error), may easily lead to such collapsed solution.In contrast, our proposed framework updates each of the encoders in different ways.From Equation (4), the online encoder is optimized to minimize the loss, while the target encoder is updated to slowly approximate the online encoder.That is, the direction of updating the target encoder (\theta-\xi) totally differs from that of updating the online encoder (-\nabla_{\theta}\mathcal{L}_{\theta,\xi}),and this effectively keeps both the encoders from converging to the collapsed solution.Several recent work on bootstrapping-based representation learning (Grill et al., 2020; Chen and He, 2021) empirically demonstrated that this kind of dynamics (i.e., updating two networks differently) allows to avoid the collapsed solution without any explicit term to prevent it.Similarly to Section 3.2, the online encoder is trained by minimizing \mathcal{L}_{\theta,\xi}(\psi(u,\mathcal{V}_{u}),\psi(v,\mathcal{U}_{v})), and the target encoder is slowly updated by the momentum mechanism.After the optimization is finished, the interaction score is inferred by f_{\theta}(u,\mathcal{V}_{u}) and f_{\theta}(v,\mathcal{U}_{v}) (Equation (5)).Figure 2 shows an example of our data augmentation which injects a certain level of perturbations to the neighbors. | Correct | null |
The online encoder is updated to minimize the error between the output and the target and updated by user-indicated gradients, but target network is updated based on the momentum update and updated as the moving average of the online encoder . | BUIR makes use of two distinct encoder networks that have the same structure: online encoder f_{\theta} and target encoder f_{\xi}.They are parameterized by \theta and \xi, respectively.The key idea of BUIR is to train the online encoder by using outputs of the target encoder as its target, while gradually improving the target encoder as well.The main difference of BUIR from existing end-to-end learning frameworks is that f_{\theta} and f_{\xi} are updated in different ways.The online encoder is trained to minimize the error between its output and the target, whereas the target network is slowly updated based on the momentum update (Heet al., 2020b) so as to keep its output consistent.To sum up, the parameters of the online encoder and target encoder are optimized by(4)\begin{split}\theta&\leftarrow\theta-\eta\cdot\nabla_{\theta}\mathcal{L}_{\theta,\xi}\\\xi&\leftarrow\tau\cdot\xi+(1-\tau)\cdot\theta.\end{split}\eta is the learning rate for stochastic optimization, and \tau\in[0,1] is a momentum coefficient (also called as target decay) for momentum-based moving average.The online encoder f_{\theta} (and the predictor q_{\theta}) is effectively optimized by the gradients back-propagated from the loss (Equation (3)), while the target encoder f_{\xi} is updated as the moving average of the online encoder.By taking a large value of \tau, the target encoder slowly approximates the online encoder.This momentum-based update makes \xi evolve more slowly than \theta, which enables to bootstrap the representations by providing enhanced but consistent targets to the online encoders (Heet al., 2020b; Grill et al., 2020).Figure 1 illustrates the overall framework of BUIR with the simple one-hot encoders.Bypassing the collapsed solution. It is obvious that the loss in Equation (3) admits the collapsed solution with respect to \theta and \xi, which means both the encoders generate the same representations for all users and items.For this reason, the conventional end-to-end learning strategy, which optimizes both f_{\theta} and f_{\xi} to minimize the loss (i.e., cross-prediction error), may easily lead to such collapsed solution.In contrast, our proposed framework updates each of the encoders in different ways.From Equation (4), the online encoder is optimized to minimize the loss, while the target encoder is updated to slowly approximate the online encoder.That is, the direction of updating the target encoder (\theta-\xi) totally differs from that of updating the online encoder (-\nabla_{\theta}\mathcal{L}_{\theta,\xi}),and this effectively keeps both the encoders from converging to the collapsed solution.Several recent work on bootstrapping-based representation learning (Grill et al., 2020; Chen and He, 2021) empirically demonstrated that this kind of dynamics (i.e., updating two networks differently) allows to avoid the collapsed solution without any explicit term to prevent it.Similarly to Section 3.2, the online encoder is trained by minimizing \mathcal{L}_{\theta,\xi}(\psi(u,\mathcal{V}_{u}),\psi(v,\mathcal{U}_{v})), and the target encoder is slowly updated by the momentum mechanism.After the optimization is finished, the interaction score is inferred by f_{\theta}(u,\mathcal{V}_{u}) and f_{\theta}(v,\mathcal{U}_{v}) (Equation (5)).Figure 2 shows an example of our data augmentation which injects a certain level of perturbations to the neighbors. | Incorrect | Change Concept |
According to the authors, assuming unobserved user-item pairs negative will lead to limited performance since there are some cases of positive but unobserved, and the number of this case is increased. | Nevertheless, the negative sampling approach has critical limitations in the following aspects.First, the underlying assumption about negative interactions becomes less valid as user-item interactions get sparser.This is because as fewer positive interactions are observed, the number of ”positive but unobserved” interactions increases, which consequently makes it even harder to sample correct negative ones.Such uncertainty of supervision eventually degrades the performance for top-K recommendation.Second, the convergence speed and the final performance depend on the specific choice of distributions for negative sampling.For example, sampling negative pairs from a non-uniform distribution (Rendle andFreudenthaler, 2014; Dinget al., 2019) (e.g., the multinomial distribution which models the probability of each interaction being actually negative) can improve the final performance, but inevitably incurs high computational costs, especially when a lot of users and items should be considered. | Correct | null |
According to the authors, assuming unobserved user-item pairs negative will not lead to limited performance since there are some cases of positive but unobserved, and the number of this case is increased. | Nevertheless, the negative sampling approach has critical limitations in the following aspects.First, the underlying assumption about negative interactions becomes less valid as user-item interactions get sparser.This is because as fewer positive interactions are observed, the number of ”positive but unobserved” interactions increases, which consequently makes it even harder to sample correct negative ones.Such uncertainty of supervision eventually degrades the performance for top-K recommendation.Second, the convergence speed and the final performance depend on the specific choice of distributions for negative sampling.For example, sampling negative pairs from a non-uniform distribution (Rendle andFreudenthaler, 2014; Dinget al., 2019) (e.g., the multinomial distribution which models the probability of each interaction being actually negative) can improve the final performance, but inevitably incurs high computational costs, especially when a lot of users and items should be considered. | Incorrect | Opposite |
To prevent the problem of collapsed sollution, the authors update target encoder and online encoder differently. | Pointing out that the contrastive methods need to carefully treat the negative instances during the training for effectiveness and efficiency, the most recent work proposed a bootstrapping-based self-supervised learning framework (Grill et al., 2020; Chen and He, 2021), which is capable of avoiding the collapsed solution without the help of negative instances.Inspired by bootstrapping methods in deep reinforcement learning (Mnihet al., 2015; Mnih et al., 2016), it directly bootstraps the representation of images by using two neural networks that iteratively learn from each other.This approach achieves the state-of-the-art performance for various downstream tasks in computer vision, and also shows better robustness to the choice of data augmentations used for self-supervision.Bypassing the collapsed solution. It is obvious that the loss in Equation (3) admits the collapsed solution with respect to \theta and \xi, which means both the encoders generate the same representations for all users and items.For this reason, the conventional end-to-end learning strategy, which optimizes both f_{\theta} and f_{\xi} to minimize the loss (i.e., cross-prediction error), may easily lead to such collapsed solution.In contrast, our proposed framework updates each of the encoders in different ways.From Equation (4), the online encoder is optimized to minimize the loss, while the target encoder is updated to slowly approximate the online encoder.That is, the direction of updating the target encoder (\theta-\xi) totally differs from that of updating the online encoder (-\nabla_{\theta}\mathcal{L}_{\theta,\xi}),and this effectively keeps both the encoders from converging to the collapsed solution.Several recent work on bootstrapping-based representation learning (Grill et al., 2020; Chen and He, 2021) empirically demonstrated that this kind of dynamics (i.e., updating two networks differently) allows to avoid the collapsed solution without any explicit term to prevent it. | Correct | null |
To prevent the problem of collapsed sollution, the authors avoid updating target encoder and online encoder differently. | Pointing out that the contrastive methods need to carefully treat the negative instances during the training for effectiveness and efficiency, the most recent work proposed a bootstrapping-based self-supervised learning framework (Grill et al., 2020; Chen and He, 2021), which is capable of avoiding the collapsed solution without the help of negative instances.Inspired by bootstrapping methods in deep reinforcement learning (Mnihet al., 2015; Mnih et al., 2016), it directly bootstraps the representation of images by using two neural networks that iteratively learn from each other.This approach achieves the state-of-the-art performance for various downstream tasks in computer vision, and also shows better robustness to the choice of data augmentations used for self-supervision.Bypassing the collapsed solution. It is obvious that the loss in Equation (3) admits the collapsed solution with respect to \theta and \xi, which means both the encoders generate the same representations for all users and items.For this reason, the conventional end-to-end learning strategy, which optimizes both f_{\theta} and f_{\xi} to minimize the loss (i.e., cross-prediction error), may easily lead to such collapsed solution.In contrast, our proposed framework updates each of the encoders in different ways.From Equation (4), the online encoder is optimized to minimize the loss, while the target encoder is updated to slowly approximate the online encoder.That is, the direction of updating the target encoder (\theta-\xi) totally differs from that of updating the online encoder (-\nabla_{\theta}\mathcal{L}_{\theta,\xi}),and this effectively keeps both the encoders from converging to the collapsed solution.Several recent work on bootstrapping-based representation learning (Grill et al., 2020; Chen and He, 2021) empirically demonstrated that this kind of dynamics (i.e., updating two networks differently) allows to avoid the collapsed solution without any explicit term to prevent it. | Incorrect | Opposite |
The authors show augmented views of positive interactions can lead the performance improvement, especially in sparser datasets by showing the experimental result of stochastic data augmentation achieved a big improvement compared to the case of using the fixed neighborhood information as encoder inputs. | In Figure 4, our stochastic data augmentation (i.e., P>0) brings a significant improvement compared to the case of using the fixed neighborhood information (i.e., P=0) as encoder inputs.This result shows that the augmented views of positive interactions encourage BUIR to effectively learn users’ preference on items even in much sparse dataset.Interestingly, in case of the Ciao dataset which is less sparse than CiteULike, the benefit of our augmentation linearly increases with the maximum drop probability.This is because there is room for producing more various views (i.e., larger perturbation) based on a relatively more number of neighbors, and it eventually helps to boost the recommendation performance.To sum up, our framework that adopts the neighbor augmentation function successfully relieves the data sparsity issue of the OCCF problem, by leveraging the augmented views of few positive interactions.Our extensive evaluation on real-world implicit feedback datasets shows that BUIR consistently performs the best for top-K recommendation among a wide range of OCCF methods.In particular, the performance improvement becomes more significant in sparser datasets, with the help of utilizing augmented views of positive interactions as well as eliminating the effect of uncertain negative interactions.In addition, comparison results on a downstream task, which classifies the items into their category, support that BUIR learns more effective representations than other OCCF baselines. | Correct | null |
The authors show augmented views of positive interactions can lead the performance improvement, especially in less sparser datasets by showing the experimental result of stochastic data augmentation achieved a big improvement compared to the case of using the fixed neighborhood information as encoder inputs. | In Figure 4, our stochastic data augmentation (i.e., P>0) brings a significant improvement compared to the case of using the fixed neighborhood information (i.e., P=0) as encoder inputs.This result shows that the augmented views of positive interactions encourage BUIR to effectively learn users’ preference on items even in much sparse dataset.Interestingly, in case of the Ciao dataset which is less sparse than CiteULike, the benefit of our augmentation linearly increases with the maximum drop probability.This is because there is room for producing more various views (i.e., larger perturbation) based on a relatively more number of neighbors, and it eventually helps to boost the recommendation performance.To sum up, our framework that adopts the neighbor augmentation function successfully relieves the data sparsity issue of the OCCF problem, by leveraging the augmented views of few positive interactions.Our extensive evaluation on real-world implicit feedback datasets shows that BUIR consistently performs the best for top-K recommendation among a wide range of OCCF methods.In particular, the performance improvement becomes more significant in sparser datasets, with the help of utilizing augmented views of positive interactions as well as eliminating the effect of uncertain negative interactions.In addition, comparison results on a downstream task, which classifies the items into their category, support that BUIR learns more effective representations than other OCCF baselines. | Incorrect | Change Concept |
Through experiments, the authors demonstrated that the performance of the model (i. , MRR@10) decreased without knowledge propagation and that it was comparable to vanilla ERNIE, which demonstrated that multi-hop neighbors were essential for ranking performance. This result can be attributed to how using multi-hope neighbors allows for knowledge to propagate between query and passage. | Knowledge propagation via meta-graph.It is worth noting that, the above-defined knowledge injection process only leverages knowledge embeddings learned by TransE on the global graph \overline{\mathcal{G}}. Particularly, it lacks considering the knowledge that bridges the semantics between query and passage. To this end, we introduce a Graph Meta Network (GMN) module that refines knowledge with the constructed meta-graph \mathbf{G}_{\mathbf{q},\mathbf{p}}, The multi-hop paths of \mathbf{G}_{\mathbf{q},\mathbf{p}} allow the knowledge to be propagated between query and passage, which can enhance the relevance signal to be captured by the model, and thus alleviate the semantic gap.By applying a K-layer GMN in each layer of the knowledge injector, the output entity representation \hat{\mathbf{E}}_{e_{h}}^{(K)} can ensemble knowledge from all the K-hop neighbors. As described in Section 4.1.2 that all the paths of \mathbf{G}_{\mathbf{q},\mathbf{p}} between \mathbf{q} and \mathbf{p} is within K hops, the GMN module can attentively propagate knowledge along the paths from entities in \mathbf{p} to those in \mathbf{q}, and vice versa, which can enrich the semantics of the entities that benefit the relevance modeling.Table 3 shows the performance comparisons between different settings of knowledge injector, which is statistically significant. From this table, we can observe the following phenomena. (1) MRR@10 of KERM without interaction and propagation process decreases at least 1\% respectively. This indicates both knowledge interaction and propagation processes play an indispensable role in ranking performance. (2) The performance of KERM without propagation is comparable to vanilla ERNIE. Not only query and passage entities, but also their multi-hop neighbors are essential for the ranking performance. (3) MRR@10 of KERM without knowledge interaction drops the most. It suggests the simple and straightforward way to aggregate knowledge graph with text does not work in the passage re-ranking scenario. The text and knowledge graph need to be refined with each other mutually in the interaction, which will be further analyzed in detail as follows. | Correct | null |
Through experiments, the authors demonstrated that the performance of the model (i. , MRR@10) decreased without knowledge propagation and that it was comparable to fine-tuned ERNIE, which demonstrated that multi-hop neighbors were essential for ranking performance. This result can be attributed to how using multi-hope neighbors allows for knowledge to propagate between query and passage. | Knowledge propagation via meta-graph.It is worth noting that, the above-defined knowledge injection process only leverages knowledge embeddings learned by TransE on the global graph \overline{\mathcal{G}}. Particularly, it lacks considering the knowledge that bridges the semantics between query and passage. To this end, we introduce a Graph Meta Network (GMN) module that refines knowledge with the constructed meta-graph \mathbf{G}_{\mathbf{q},\mathbf{p}}, The multi-hop paths of \mathbf{G}_{\mathbf{q},\mathbf{p}} allow the knowledge to be propagated between query and passage, which can enhance the relevance signal to be captured by the model, and thus alleviate the semantic gap.By applying a K-layer GMN in each layer of the knowledge injector, the output entity representation \hat{\mathbf{E}}_{e_{h}}^{(K)} can ensemble knowledge from all the K-hop neighbors. As described in Section 4.1.2 that all the paths of \mathbf{G}_{\mathbf{q},\mathbf{p}} between \mathbf{q} and \mathbf{p} is within K hops, the GMN module can attentively propagate knowledge along the paths from entities in \mathbf{p} to those in \mathbf{q}, and vice versa, which can enrich the semantics of the entities that benefit the relevance modeling.Table 3 shows the performance comparisons between different settings of knowledge injector, which is statistically significant. From this table, we can observe the following phenomena. (1) MRR@10 of KERM without interaction and propagation process decreases at least 1\% respectively. This indicates both knowledge interaction and propagation processes play an indispensable role in ranking performance. (2) The performance of KERM without propagation is comparable to vanilla ERNIE. Not only query and passage entities, but also their multi-hop neighbors are essential for the ranking performance. (3) MRR@10 of KERM without knowledge interaction drops the most. It suggests the simple and straightforward way to aggregate knowledge graph with text does not work in the passage re-ranking scenario. The text and knowledge graph need to be refined with each other mutually in the interaction, which will be further analyzed in detail as follows. | Incorrect | Change Concept |
This work proposes an aggregation module that employs a PLM and a Graph Neural Network (GMN) to model the interaction between explicit and implicit knowledge. The PLM encodes text to obtain word representations (i. , implicit knowledge), and the Graph Neural Network (GMN) encodes knowledge meta-graphs to obtain entity representations (i. , explicit knowledge). This module aggregates the word and entity representations to aggregate the implicit and explicit knowledge. | Knowledge propagation via meta-graph.It is worth noting that, the above-defined knowledge injection process only leverages knowledge embeddings learned by TransE on the global graph \overline{\mathcal{G}}. Particularly, it lacks considering the knowledge that bridges the semantics between query and passage. To this end, we introduce a Graph Meta Network (GMN) module that refines knowledge with the constructed meta-graph \mathbf{G}_{\mathbf{q},\mathbf{p}}, The multi-hop paths of \mathbf{G}_{\mathbf{q},\mathbf{p}} allow the knowledge to be propagated between query and passage, which can enhance the relevance signal to be captured by the model, and thus alleviate the semantic gap.For knowledge aggregation, we design a novel interaction module between text and knowledge graph to combine the implicit and explicit knowledge. To derive implicit knowledge from text, we employ PLM as text encoder. To be aligned with implicit knowledge, knowledge meta graph is encoded with a multi-layer graph neural network (i.e. k-hop), namely Graph Meta Network (GMN). Each transformer layer outputs word representations. Each graph meta network layer outputs entity representations. Both word and entity representations are aggregated as the input of the following transformer and GMN layer, respectively in a novelly designed module, namely knowledge injector. Therefore through knowledge aggregation, implicit knowledge from text corpus and explicit knowledge from existing knowledge graph can mutually boost each other to achieve a better re-ranking performance, in which the issues in Challenge 2. could be mitigated.Overall, our contributions can be summarized as follows:•It is the first attempt to solve the knowledge enhanced PLMs problem for passage re-ranking. The key motivation lies in that bridging the semantic gap between the query and passage with the help of both kinds of knowledge.•We design a novel knowledge graph distillation method. It refines a reliable knowledge graph from the existing one globally and constructs a knowledge meta graph based on the refined graph locally.•We propose a novel aggregation of PLM and graph neural network framework to model the interaction between explicit knowledge and implicit knowledge.•Experimental results show the effectiveness of KERM on both general and domain specific data, achieving state-of-the-art performance for passage re-ranking. We also conduct a comprehensive study for the effects of each module in our method. The code is available at https://github.com/DQ0408 /KERM. | Correct | null |
This work proposes an aggregation module that employs a LLM and a Graph Neural Network (GMN) to model the interaction between explicit and implicit knowledge. The LLM encodes text to obtain word representations (i. , implicit knowledge), and the Graph Neural Network (GMN) encodes knowledge meta-graphs to obtain entity representations (i. , explicit knowledge). This module aggregates the word and entity representations to aggregate the implicit and explicit knowledge. | Knowledge propagation via meta-graph.It is worth noting that, the above-defined knowledge injection process only leverages knowledge embeddings learned by TransE on the global graph \overline{\mathcal{G}}. Particularly, it lacks considering the knowledge that bridges the semantics between query and passage. To this end, we introduce a Graph Meta Network (GMN) module that refines knowledge with the constructed meta-graph \mathbf{G}_{\mathbf{q},\mathbf{p}}, The multi-hop paths of \mathbf{G}_{\mathbf{q},\mathbf{p}} allow the knowledge to be propagated between query and passage, which can enhance the relevance signal to be captured by the model, and thus alleviate the semantic gap.For knowledge aggregation, we design a novel interaction module between text and knowledge graph to combine the implicit and explicit knowledge. To derive implicit knowledge from text, we employ PLM as text encoder. To be aligned with implicit knowledge, knowledge meta graph is encoded with a multi-layer graph neural network (i.e. k-hop), namely Graph Meta Network (GMN). Each transformer layer outputs word representations. Each graph meta network layer outputs entity representations. Both word and entity representations are aggregated as the input of the following transformer and GMN layer, respectively in a novelly designed module, namely knowledge injector. Therefore through knowledge aggregation, implicit knowledge from text corpus and explicit knowledge from existing knowledge graph can mutually boost each other to achieve a better re-ranking performance, in which the issues in Challenge 2. could be mitigated.Overall, our contributions can be summarized as follows:•It is the first attempt to solve the knowledge enhanced PLMs problem for passage re-ranking. The key motivation lies in that bridging the semantic gap between the query and passage with the help of both kinds of knowledge.•We design a novel knowledge graph distillation method. It refines a reliable knowledge graph from the existing one globally and constructs a knowledge meta graph based on the refined graph locally.•We propose a novel aggregation of PLM and graph neural network framework to model the interaction between explicit knowledge and implicit knowledge.•Experimental results show the effectiveness of KERM on both general and domain specific data, achieving state-of-the-art performance for passage re-ranking. We also conduct a comprehensive study for the effects of each module in our method. The code is available at https://github.com/DQ0408 /KERM. | Incorrect | Change Concept |
This work conducted ablation studies to investigate the contribution of each component in the performance of KERM. By testing different settings for the knowledge injector, this work found that performance decreases without knowledge interaction and also without knowledge propagation. By testing the model without global or local distillation, they also demonstrated that performance decreases without global distillation and efficiency decreases without either global or local distillation. These experiments demonstrate that each component of KERM contributes to passage re-ranking performance quantitatively. | Here we compare ranking performances of KERM and other PLMs based re-rankers on the first two widely used query sets. Moreover, ablation studies for each component of KERM are also explored. All experimental results were reported under the same BM25 setting.Table 3 shows the performance comparisons between different settings of knowledge injector, which is statistically significant. From this table, we can observe the following phenomena. (1) MRR@10 of KERM without interaction and propagation process decreases at least 1\% respectively. This indicates both knowledge interaction and propagation processes play an indispensable role in ranking performance. (2) The performance of KERM without propagation is comparable to vanilla ERNIE. Not only query and passage entities, but also their multi-hop neighbors are essential for the ranking performance. (3) MRR@10 of KERM without knowledge interaction drops the most. It suggests the simple and straightforward way to aggregate knowledge graph with text does not work in the passage re-ranking scenario. The text and knowledge graph need to be refined with each other mutually in the interaction, which will be further analyzed in detail as follows.Knowledge graph distillation is performed in both global and local perspectives. To explore their roles in the ranking performance, we remove the graph pruning globally and sentence selection locally respectively, keep other settings unchanged, and derive KERM without graph pruning and sentence selection respectively. From results on TREC 2019 DL in Table 5, observations are listed as below. (1) Without global graph pruning, MRR@10 and the average edge score, calculated through Eq.(3), decrease the most, and the time efficiency drops slightly. This indicates the original knowledge graph exists noise data that affect performance. (2) Without sentence selection, the time efficiency drops the most and the average edge score decreases slightly, which proves that not every sentence in a passage has a positive effect on semantic matching. Overall, knowledge graph distillation is significant to KERM. | Correct | null |
This work conducted ablation studies to investigate the contribution of each component in the performance of KERM. By testing different settings for the knowledge injector, this work found that performance doesn't change without knowledge interaction and also without knowledge propagation. However, by testing the model without global or local distillation, they demonstrated that performance decreases without global distillation and efficiency decreases without either global or local distillation. These experiments demonstrate that each component of KERM contributes to passage re-ranking performance quantitatively. | Here we compare ranking performances of KERM and other PLMs based re-rankers on the first two widely used query sets. Moreover, ablation studies for each component of KERM are also explored. All experimental results were reported under the same BM25 setting.Table 3 shows the performance comparisons between different settings of knowledge injector, which is statistically significant. From this table, we can observe the following phenomena. (1) MRR@10 of KERM without interaction and propagation process decreases at least 1\% respectively. This indicates both knowledge interaction and propagation processes play an indispensable role in ranking performance. (2) The performance of KERM without propagation is comparable to vanilla ERNIE. Not only query and passage entities, but also their multi-hop neighbors are essential for the ranking performance. (3) MRR@10 of KERM without knowledge interaction drops the most. It suggests the simple and straightforward way to aggregate knowledge graph with text does not work in the passage re-ranking scenario. The text and knowledge graph need to be refined with each other mutually in the interaction, which will be further analyzed in detail as follows.Knowledge graph distillation is performed in both global and local perspectives. To explore their roles in the ranking performance, we remove the graph pruning globally and sentence selection locally respectively, keep other settings unchanged, and derive KERM without graph pruning and sentence selection respectively. From results on TREC 2019 DL in Table 5, observations are listed as below. (1) Without global graph pruning, MRR@10 and the average edge score, calculated through Eq.(3), decrease the most, and the time efficiency drops slightly. This indicates the original knowledge graph exists noise data that affect performance. (2) Without sentence selection, the time efficiency drops the most and the average edge score decreases slightly, which proves that not every sentence in a passage has a positive effect on semantic matching. Overall, knowledge graph distillation is significant to KERM. | Incorrect | Change Concept |
This work proposes using knowledge graph distillation as it can help retain only informative knowledge needed for passage re-ranking. By investigating the effect of global and local distillation separately, this work found that the MRR@10 score and efficiency decreased slightly without global distillation, and that time efficiency decreased the most without local distillation. Therefore, this work demonstrates that both global and local distillation of knowledge graphs is useful for re-ranking tasks in terms of performance and efficiency. | Existing knowledge graphs are usually incomplete and noisy. It is unsuitable for direct introduction of them to the current model. Specially, there is no knowledge base particularly for passage re-ranking task. For example, ConceptNet (Speeret al., 2017) is a general knowledge graph that contains common sense knowledge, where the information might not be useful for our passage re-ranking task. Therefore, it is critical for us to propose a knowledge graph distillation process from both global and local perspectives.Knowledge graph distillation is performed in both global and local perspectives. To explore their roles in the ranking performance, we remove the graph pruning globally and sentence selection locally respectively, keep other settings unchanged, and derive KERM without graph pruning and sentence selection respectively. From results on TREC 2019 DL in Table 5, observations are listed as below. (1) Without global graph pruning, MRR@10 and the average edge score, calculated through Eq.(3), decrease the most, and the time efficiency drops slightly. This indicates the original knowledge graph exists noise data that affect performance. (2) Without sentence selection, the time efficiency drops the most and the average edge score decreases slightly, which proves that not every sentence in a passage has a positive effect on semantic matching. Overall, knowledge graph distillation is significant to KERM.The main goal of this paper is to reasonably introduce external knowledge graph to PLMs for passage re-ranking. We first design a novel knowledge meta graph construction method to distill reliable and query related knowledge from a general and noisy knowledge graph. The knowledge meta graph bridges the semantic gap between each query and passage. Then we propose a knowledge injector layer for mutually updating text and knowledge representations, which transformers word to entity representations for graph meta network, vice versa. Knowledge Enhanced Ranking Model is pretrained with Masked Language Model (MLM) Sentence Relation Prediction (SRP) [38] tasks, and fine-tuned with cross entropy loss function for passage re-ranking task. Experimental results on public benchmark datasets show the effectiveness of the proposed method compared with state-of-the-art baselines without external knowledge due to its first attempt. The role of each module in KERM is also comprehensively analyzed. Since this work was limited to the one-to-one meta-graph of a query-passage pair built online, continued efforts are needed to make knowledge enhancement more efficient for both retrieval and re-ranking stage.Despite that the knowledge graph distillation in our method is empirically shown to be effective for the final performance, the implementation of graph pruning and meta-graph construction is still based on simple heuristics. A more promising way of formulating a useful meta-graph is to jointly learn a graph generator with the reranker in an end-to-end fashion, which enables more flexibility.Besides, it is currently infeasible to exploit the external knowledge in the retrieval stage, which needs to exhaustively build massive meta-graphs for a large scale of candidates. A further study could focus on how to use external knowledge in PLM based retriever.For knowledge graph distillation, we propose a novel pipeline to establish knowledge meta graphs, which only retain informative knowledge for passage re-ranking. Specifically, we first distill a graph globally for passage re-ranking scenario from an existing knowledge graph by pruning some unreliable or noisy relations based on TransE embedding. Then for a specific query-passage pair, we extract entities from both the query and passage, and construct a query-document bipartite entity graph based on query and passage entities and their k-hop neighbors, namely knowledge meta graph. Challenge 1. could be addressed in this distillation process. | Correct | null |
End of preview. Expand
in Data Studio
SciCiteVal
Introduction
Our dataset is specifically designed for citation verification, it contains manual annotations labeling citations as Correct, Incorrect or Unrelated. For citations labeled Incorrect, we further define five detailed sub-categories to characterize the nature of the inaccuracy. Each data sample is structured as a pair, consisting of the citation context from the citing paper and the corresponding evidence passage from the cited paper that justifies the label.
Our dataset contains 1,034 citations distributed across three categories: 302 Correct citations, 302 Incorrect citations, and 430 Unrelated citations. The citations are from scientific papers in machine learning and biology domains. Both Correct and Incorrect citations are adapted from the QASA dataset, whereas unrelated citations are extracted from real citations that cite unrelated articles.
Language
English
Supported Tasks
- Text Classification
Files Stucture
- Few_shot_matrix contains confusion-matrix figures of each benchmarking model with few-shot prompt.
- Results_cleaned contains benchmarking results of all the models with zero-shot and few-shot prompts in tsv format.
- Zero_shot_matrix contains confusion-matrix figures of each benchmarking model with zero-shot prompt.
- experiment_dataset.tsv is the file that we use as input for benchmarking models
- Qasa_annotated.tsv is the file that contains details of annotations and modifications we made with Qasa dataset. This file preserves the original QASA dataset IDs.
Dataset Information
The dataset contains four columns: "Citation Context", "Cited Content", "Label", and "Distortion Category"
- Citation_context (String) contains the citation context transformed from the "Composition" of the Qasa dataset by our annotators
- Cited_content (String) contains the content that justifies the corresponding citation context, originally from the "Evidence" of the Qasa dataset
- Label (String) contains the label for the corresponding citation.
- Twist_category (String) contains the label of distortion for the manually-distorted citation context.
Annotation Process
Correct Citations: We manually verified and transformed QASA "composition" to citation. Compositions consisting only of "Yes"/"No" answers were excluded.
Incorrect Citations: Two annotators with NLP backgrounds read the original Qasa "composition" and systematically distorted it to produce a citation context that inaccurately references the "Cited Content". Minimal changes were made to improve fluency.
Unrelated Citations: These were manually collected from scientific papers across diverse fields (e.g., social science, biology). Only complete sentences from the main body of the text were included, content from tables or figures was excluded.
Referenced Dataset
Citation Information
Will be updated soon
- Downloads last month
- 47