Partial-input baselines show that NLI models can ignore context, but they don’t. (2022.naacl-main)
Copied to clipboard
| Challenge: | Researchers have shown that many datasets contain statistical biases, or "annotation artifacts" that systems leverage to correctly predict entailment. |
| Approach: | They propose to use edited contexts to examine RoBERTa models' sensitivity to edited context to examine their model's sensitivity. |
| Outcome: | The proposed model can learn to condition on context, despite being trained on artifact-ridden datasets. |
Similar Papers
Misleading Failures of Partial-input Baselines (P19-1)
Copied to clipboard
| Challenge: | Recent work establishes dataset difficulty and removes annotation artifacts via partial-input baselines. |
| Approach: | They propose to use partial-input baselines to establish dataset difficulty . they show how trivial patterns only visible in the full input can evade partial-output baseline . |
| Outcome: | The proposed model can solve 15% of previously-thought "hard" examples. |
“You are grounded!”: Latent Name Artifacts in Pre-trained Language Models (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models perpetuate biases originating in their training corpus to downstream models. |
| Approach: | They focus on the representations of given names in pre-trained language models and show that name perturbation can have an effect on downstream tasks. |
| Outcome: | The proposed model can be used to model the representation of given names in pre-trained language models on reading comprehension probes where name perturbation changes the model answers. |
Don’t Take the Premise for Granted: Mitigating Artifacts in Natural Language Inference (P19-1)
Copied to clipboard
| Challenge: | Natural Language Inference (NLI) datasets often contain hypothesis-only biases . authors propose probabilistic methods to build models that are more robust to such artifacts - a new study shows . |
| Approach: | They propose probabilistic methods to build models that are more robust to biases . authors train on datasets containing biase . |
| Outcome: | The proposed methods can make NLI models more robust to dataset-specific artifacts . the methods transfer better than a baseline architecture in 9 out of 12 NLI datasets compared with baseline architectures based on the proposed methods . |
Which Spurious Correlations Impact Reasoning in NLI Models? A Visual Interactive Diagnosis through Data-Constrained Counterfactuals (2023.acl-demo)
Copied to clipboard
| Challenge: | a spurious correlation exists when a feature correlates with the target label while there is no causal relationship between the feature and the label. |
| Approach: | They propose a dashboard that allows users to generate diverse and challenging examples by drawing inspiration from GPT-3 suggestions. |
| Outcome: | The proposed dashboard enables users to generate diverse and challenging examples by drawing inspiration from GPT-3 suggestions and make refinements based on the feedback. |
Revisiting Pathologies of Neural Models under Input Reduction (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent studies have shown that modern neural models tend to be miscalibrated. |
| Approach: | They examine why models produce high-confidence predictions on inputs that appear nonsensical to humans . previous work suggested that models fail to assign low probabilities due to model overconfidence . |
| Outcome: | The proposed methods can be extended to reduce the number of examples but with the cost of miscalibration. |
Is My Model Using the Right Evidence? Systematic Probes for Examining Evidence-Based Tabular Reasoning (2022.tacl-1)
Copied to clipboard
| Challenge: | Existing models that claim to reason about evidence should avoid spurious patterns . tabular inputs are well-suited for the study—they admit systematic probes . |
| Approach: | They propose to use tabular data to test whether models can reason about evidence . they show that a RoBERTa-based model fails to reason on the following counts . |
| Outcome: | The proposed model fails to reason on tabular data on the following counts . the model is over-sensitive to annotation artifacts and ignores relevant parts of the evidence . |
Do Neural Language Models Overcome Reporting Bias? (2020.coling-main)
Copied to clipboard
| Challenge: | Recent studies show that pre-trained language models can overcome reporting bias by estimating the plausibility of rare but unspoken facts. |
| Approach: | They revisit the experiments conducted by Gordon and Van Durme (2013) . they find that pre-trained language models overestimate the very rare . |
| Outcome: | The proposed approach overestimates the rare at the expense of the rare, while minimizing reporting bias. |
Probing Neural Network Comprehension of Natural Language Arguments (P19-1)
Copied to clipboard
| Challenge: | Argument Reasoning Comprehension Task (ARCT) focuses on inferences, not just discovering warrants. |
| Approach: | They propose to build an adversarial dataset on which all models achieve random accuracy. |
| Outcome: | The proposed dataset provides a more robust assessment of argument comprehension and should be adopted as the standard in future work. |
No Context Needed: Contextual Quandary In Idiomatic Reasoning With Pre-Trained Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | idiomatic expressions (IEs) are a non-compositional aspect of a text that makes it difficult for a model to comprehend . general purpose PTLMs are negatively affected by the context, as performance increases with its removal. |
| Approach: | They propose to use idiomatic expressions to infer additional meaning from IEs . they argue that only IE-aware models are suitable for idiom- matic reasoning tasks . |
| Outcome: | The proposed models can reason in the presence of idiomatic expressions, the authors show . they show that general purpose PTLMs are negatively affected by the context . |
Can NLI Models Verify QA Systems’ Predictions? (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Recent question answering systems perform well on benchmark datasets, but are not always well-calibrated to spot spurious answers under distribution shifts. |
| Approach: | They propose to use natural language inference to verify whether answers are correct . they leverage large pre-trained models and recent prior datasets to construct powerful question conversion and decontextualization modules. |
| Outcome: | The proposed approach improves the confidence estimation of a QA model across different domains, evaluated in a selective QA setting. |