Source-linked AI summary
Explain Yourself! Leveraging Language Models for Commonsense Reasoning
Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, Richard Socher
TL;DR
Commonsense reasoning remains difficult for deep learning models because it can require world knowledge beyond the input. This paper collects human explanations in CoS-E and uses them to train CAGE, which generates explanations for answer prediction. CAGE improves reported state-of-the-art performance on CommonsenseQA, while the paper also examines explanation transfer and limitations related to bias and scope.
Problem
It remains unclear how well commonsense models reason and how much their reasoning relies on world knowledge.
Method
The paper collects human explanations in CoS-E and trains a language model in CAGE to generate explanations used by a commonsense reasoning classifier.
Results
10%: CAGE outperforms the best reported baseline on CommonsenseQA, while explanation transfer is also studied on out-of-domain tasks.
Takeaways & Limitations
Human and automatically generated explanations provide a basis for studying and leveraging neural commonsense reasoning, including transfer beyond CommonsenseQA.
Takeaways & Limitations
CoS-E and models trained from it inherit observed gender bias, including a higher proportion of female pronouns in negative contexts.
Abstract
from arXiv · showhide
Deep learning models perform poorly on tasks that require commonsense reasoning, which often necessitates some form of world-knowledge or reasoning over information not immediately present in the input. We collect human explanations for commonsense reasoning in the form of natural language sequences and highlighted annotations in a new dataset called Common Sense Explanations (CoS-E). We use CoS-E to train language models to automatically generate explanations that can be used during training and inference in a novel Commonsense Auto-Generated Explanation (CAGE) framework. CAGE improves the state-of-the-art by 10% on the challenging CommonsenseQA task. We further study commonsense reasoning in DNNs using both human and auto-generated explanations including transfer to out-of-domain tasks. Empirical results indicate that we can effectively leverage language models for commonsense reasoning.
1 Introduction
The paper introduces CoS-E, a dataset of human commonsense explanations, and CAGE, which generates explanations for CommonsenseQA. CAGE uses these explanations to improve commonsense reasoning and achieves approximately 65% accuracy on CQA v1.0.
- Dataset and motivation: CoS-E adds human open-ended explanations and highlighted words to CommonsenseQA examples.The highlights identify words humans considered important for selecting the correct answer.
- Method: CAGE trains a language model to generate explanations from questions and answer choices, then supplies them to a second reasoning model.The generated explanation is concatenated with the original question and answer choices.
- Results: 10%: CAGE outperforms the best reported baseline on CommonsenseQA.The framework both improves performance and produces explanations intended to justify predictions.
- Results: Approximately 65% accuracy: CAGE achieves state-of-the-art performance on CQA v1.0.The paper also reports experiments on the later, more challenging CQA v1.11 release.
2 Background and Related Work
CommonsenseQA and related benchmarks test reasoning about situations, events, and world knowledge, but strong language models can still perform poorly. CoS-E augments such benchmarks with explanations that support analysis of how models reason.
- Commonsense reasoning benchmarks: Story Cloze, SWAG, Winograd Schemas, and CommonsenseQA evaluate different forms of commonsense reasoning.These tasks include predicting story endings, next scenes, or answers requiring relations between situations and world knowledge.
- CommonsenseQA: CommonsenseQA contains 9500 questions with one correct answer and two distractors drawn from the same source concept.Its construction is intended to require inference from the question rather than exploitation of distributional biases.
- CommonsenseQA: State-of-the-art language models perform very poorly compared with human participants on CommonsenseQA.The benchmark therefore exposes a gap between current model performance and human commonsense reasoning.
- Explanations: CoS-E provides explanations that can be used to study, analyze, and evaluate models’ reasoning capabilities.This extends CommonsenseQA beyond answer accuracy alone.
- Natural-language explanations: Prior explanation methods have produced mixed outcomes, including reported performance loss on e-SNLI, whereas explanations for CQA improved performance.The paper positions CQA as a setting where explanations may aid both interpretability and accuracy.
3 Common Sense Explanations (CoS-E)
CoS-E is built from human explanations collected for CommonsenseQA using crowdsourcing, with both highlighted justifications and open-ended text. Although the annotations are noisy and often do not contain the answer, they are sufficient to train explanation-generating models.
- Data collection: Amazon Mechanical Turk was used to collect CoS-E explanations on the more difficult random split of CommonsenseQA.The dataset and experiments use this random split as the main evaluation setting.
- Annotation procedure: Annotators received each question, answer choices, and ground-truth answer, then highlighted relevant words and wrote a brief explanation.The collected data covered 7610 training and 950 development examples for the random split.
- Quality control: In-browser checks rejected explanations without highlights, explanations shorter than four words, and direct substrings of questions or answer choices.Only one annotator was used per example, followed by additional quality checks.
- Dataset analysis: 58%: CoS-E explanations contain the ground-truth answer.The dataset also includes explanations containing distractors or using elimination of incorrect choices.
- Dataset utility: CoS-E can improve performance even when explanations have no word overlap with answer choices.The authors report state-of-the-art results using CoS-E only during training, despite remaining annotation noise.
4 Algorithm
CAGE separates explanation generation from answer prediction: a language model learns from CoS-E, and a classifier uses generated explanations with the original question and choices. The paper also distinguishes reasoning explanations from label-conditioned rationalizations and studies transfer to other tasks.
- CAGE framework: CAGE trains a language model to generate an explanation e from a question q, answer choices, and human explanation eh.The generated explanation is learned from CoS-E and is used as supplementary input to classification.
- Reasoning: In reasoning mode, the language model generates explanations before answer prediction without conditioning on the answer label.This is intended to provide additional context during inference for commonsense question answering.
- Evaluation: CAGE reasoning outperforms the reported state of the art on CommonsenseQA by 10%.The paper also generates explanations for SWAG and Story Cloze and trains classifiers using those explanations.
- Rationalization: In rationalization mode, the language model conditions on predicted labels and generates post-hoc explanations.The paper treats these outputs as rationalizations rather than commonsense reasoning because the label is provided.
- Classification: BERT predicts among answer choices using inputs that concatenate the question, explanation when available, and each candidate answer.When explanations are used only during training, they are omitted during evaluation.
5 Experimental Results
Experiments evaluate human and automatically generated explanations on CQA, oracle settings, a harder CQA version, and transfer to out-of-domain tasks. Explanation-based models improve CQA performance, while transferred explanations produce little change on SWAG and Story Cloze.
- CQA results: 72% accuracy results from using CAGE-reasoning during both training and validation, compared with 64% for the BERT baseline.Adding open-ended human explanations during training yields a 2% boost over the baseline.
- CQA results: 10% absolute gain over the previous state-of-the-art results from replacing CoS-E training explanations with CAGE reasoning during both training and inference on CQA v1.0.The test comparison is summarized in Table 3.
- Oracle analysis: 90% accuracy is reached in the oracle setting when open-ended human explanations are provided during both training and inference.This setting is explicitly described as unfair because the explanation provider had access to the ground-truth answer.
- Oracle analysis: 53% accuracy with selected-word explanations contrasts with 85% using open-ended human explanations when the original question is excluded.These variants use explanations during both training and validation.
- CQA v1.11: Limited open-ended explanations without words overlapping answer choices still improve over the BERT baseline on CQA v1.11.The result is presented as evidence that explanations provide information beyond simply mentioning correct or incorrect answers.
- Out-of-domain transfer: Explanation transfer to SWAG and Story Cloze causes a decrease of less than 0.6% compared with the baseline on both tasks.The transfer evaluation uses explanations from CQA on out-of-domain tasks.
6 Analysis and Discussion
The analysis examines when explanations improve commonsense question answering, how well generated explanations support prediction, and where CAGE remains limited. Results show strong performance on CQA, but explanation quality, transfer, and classifier use of generated information remain uneven.
- Explanation quality: 4.1 BLEU and 32 perplexity were achieved by CAGE, compared with 0.8 BLEU for language models that were not fine-tuned.BLEU measures syntactic n-gram precision, while perplexity measures token-level next-word prediction.
- Error analysis: Explanations are particularly relevant to longer, more compositional questions, where the baseline BERT model performs poorly.The reported average question length was 14 words for such cases versus 13 words for questions CAGE predicted incorrectly.
- Human evaluation: 42% of Turkers matched the model’s answer from CAGE-reasoning alone, compared with 52% when given CoS-E-open-ended explanations.CAGE-rationalization produced the same 42% human agreement and often differed from CAGE-reasoning only in word ordering or answer-choice substitutions.
- Robustness: 30% accuracy followed training on adversarial misleading explanations, down from 60% and below the 50% validation baseline.Among errors from misleading explanations, 57% were correctly answered by the model trained with true CoS-E explanations.
- Error analysis: CAGE explanations can contain the correct answer while the BERT classifier still selects the wrong choice, especially when answer options are semantically close.On CQA v1.11, examples included correct terms such as “dresser drawer” and “cleanness” in generated explanations, yet the classifier failed to exploit them.
- Out-of-domain transfer: Transfer to SWAG and Story Cloze caused a slight performance decrease despite explanations that often contained relevant information.The classifier was unable to use information that appeared grounded in SWAG images or pointed toward the correct Story Cloze ending.
7 Conclusion and Future Work
The paper introduces CoS-E and CAGE to generate and use explanations for commonsense reasoning. It reports state-of-the-art CQA performance and studies explanation transfer, while identifying broader task coverage and bias as future concerns.
- Contributions: CoS-E is a dataset of human commonsense explanations built on CommonsenseQA, and CAGE trains language models to generate explanations for classifier-based prediction.The framework connects human explanations during language-model fine-tuning with generated explanations supplied to a downstream commonsense reasoning model.
- Findings: The approach achieves state-of-the-art performance on a difficult commonsense reasoning task and supports explanation transfer to out-of-domain datasets.The paper also reports comprehensive error analyses of language-model explanations.
- Future work: CAGE currently generates explanations before answer prediction, leaving joint answer-and-explanation training and broader task coverage as future directions.The paper suggests that explanation datasets across many tasks could support a more general explanatory language model for unseen tasks.
- Ethical considerations: Biases accumulated during pretraining or fine-tuning may propagate into explanations and downstream models, requiring careful monitoring.The authors observed gender disparity in CoS-E, including a higher proportion of female pronouns in negative contexts.