Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ERASER: A Benchmark to Evaluate Rationalized NLP Models

Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, Byron C. Wallace

TL;DR

ERASER introduces a unified benchmark for evaluating NLP rationales by collecting seven datasets with human-annotated rationales and providing standardized metrics for agreement and faithfulness. It benchmarks hard- and soft-selection approaches to rationale extraction, showing diverse performance across tasks and highlighting the need for adaptable, long-input-aware models. The resource and baseline methods aim to catalyze progress in interpretable NLP and enable fairer cross-task comparisons, with open avenues for improved evaluation metrics and multilingual rationales. Overall, ERASER sets the foundation for systematic study of how explanations relate to model predictions in real-world NLP tasks.

Abstract

State-of-the-art models in NLP are now predominantly based on deep neural networks that are opaque in terms of how they come to make predictions. This limitation has increased interest in designing more interpretable deep models for NLP that reveal the `reasoning' behind model outputs. But work in this direction has been conducted on different datasets and tasks with correspondingly unique aims and metrics; this makes it difficult to track progress. We propose the Evaluating Rationales And Simple English Reasoning (ERASER) benchmark to advance research on interpretable models in NLP. This benchmark comprises multiple datasets and tasks for which human annotations of "rationales" (supporting evidence) have been collected. We propose several metrics that aim to capture how well the rationales provided by models align with human rationales, and also how faithful these rationales are (i.e., the degree to which provided rationales influenced the corresponding predictions). Our hope is that releasing this benchmark facilitates progress on designing more interpretable NLP systems. The benchmark, code, and documentation are available at https://www.eraserbenchmark.com/

ERASER: A Benchmark to Evaluate Rationalized NLP Models

TL;DR

ERASER introduces a unified benchmark for evaluating NLP rationales by collecting seven datasets with human-annotated rationales and providing standardized metrics for agreement and faithfulness. It benchmarks hard- and soft-selection approaches to rationale extraction, showing diverse performance across tasks and highlighting the need for adaptable, long-input-aware models. The resource and baseline methods aim to catalyze progress in interpretable NLP and enable fairer cross-task comparisons, with open avenues for improved evaluation metrics and multilingual rationales. Overall, ERASER sets the foundation for systematic study of how explanations relate to model predictions in real-world NLP tasks.

Abstract

State-of-the-art models in NLP are now predominantly based on deep neural networks that are opaque in terms of how they come to make predictions. This limitation has increased interest in designing more interpretable deep models for NLP that reveal the `reasoning' behind model outputs. But work in this direction has been conducted on different datasets and tasks with correspondingly unique aims and metrics; this makes it difficult to track progress. We propose the Evaluating Rationales And Simple English Reasoning (ERASER) benchmark to advance research on interpretable models in NLP. This benchmark comprises multiple datasets and tasks for which human annotations of "rationales" (supporting evidence) have been collected. We propose several metrics that aim to capture how well the rationales provided by models align with human rationales, and also how faithful these rationales are (i.e., the degree to which provided rationales influenced the corresponding predictions). Our hope is that releasing this benchmark facilitates progress on designing more interpretable NLP systems. The benchmark, code, and documentation are available at https://www.eraserbenchmark.com/

Paper Structure

This paper contains 19 sections, 3 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Datasets in ERASER
  4. Metrics
  5. Agreement with human rationales
  6. Measuring faithfulness
  7. Baseline Models
  8. Hard selection
  9. Soft selection
  10. Evaluation
  11. Conclusions and Future Directions
  12. Acknowledgements
  13. Dataset Preprocessing
  14. Annotation details
  15. Hyperparameter and training details
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Examples of instances, labels, and rationales illustrative of four (out of seven) datasets included in ERASER. The 'erased' snippets are rationales.
  • Figure 2: Illustration of faithfulness scoring metrics, comprehensiveness and sufficiency, on the Commonsense Explanations (CoS-E) dataset. For the former, erasing the tokens comprising the provided rationale ($\tilde{x}_i$) ought to decrease model confidence in the output 'Forest'. For the latter, the model should be able to come to a similar disposition regarding 'Forest' using only the rationales $r_i$.