What to Learn, and How: Toward Effective Learning from Rationales

Findings of the Association for Computational Linguistics: ACL 2022, pages 1075 - 1088 May 22-27, 2022 c⃝2022 Association for Computational Linguistics What to Learn, and How: Toward Effective Learning from Rationales Samuel Carton University of Chicago carton@uchicago.edu Surya Kanoria University of Colorado Boulder surya.kanoria@colorado.edu Chenhao Tan University of Chicago chenhao@uchicago.edu Abstract Learning from rationales seeks to augment model prediction accuracy using human- annotated rationales (i.e. subsets of input to- kens) that justify their chosen labels, often in the form of intermediate or multitask supervi- sion. While intuitive, this idea has proven elu- sive in practice. We make two observations about human rationales via empirical analy- ses: 1) maximizing rationale supervision ac- curacy is not necessarily the optimal objective for improving model accuracy; 2) human ratio- nales vary in whether they provide sufficient information for the model to exploit for pre- diction. Building on these insights, we pro- pose several novel loss functions and learn- ing strategies, and evaluate their effectiveness on three datasets with human rationales. Our results demonstrate consistent improvements over baselines in both label and rationale ac- curacy, including a 3% accuracy improvement on MultiRC. Our work highlights the impor- tance of understanding properties of human ex- planations and exploiting them accordingly in model training. 1 Introduction In the past several years, explainability has become a prominent issue in machine learning, addressing concerns about the safety and ethics of using large, opaque models for decision-making. As interest has grown in explanations for understanding model behavior, so has interest grown in soliciting gold- standard explanations from human annotators and using them to inject useful inductive biases into models (Hase and Bansal, 2021). Many such ex- planation datasets have become available recently (Wiegreffe and Marasovi´c, 2021). A common format for explanations in NLP is the rationale, a subset of input tokens that are relevant to the decision. A popular architecture for gen- erating such explanations is the rationale model, (A) Unsupervised rationale [CLS] susan wanted to have a birthday party . she called all of her friends . she has five friends . her mom said that susan can invite them all to the party . her first friend could not go to the party because she was sick . her second friend was going out of town . her third friend was not so sure if her parents would let her . the fourth friend said maybe . the fifth friend could go to the party for sure . susan was a little sad . on the day of the party , all five friends showed up . each friend had a present for susan . susan was happy and sent each friend a thank you card the next week . [SEP] how many people did susan call ? | | 5 [SEP] Prediction: False (B) Human rationale [CLS] susan wanted to have a birthday party . she called all of her friends . she has five friends . her mom said that susan can invite them all to …

Content truncated to 500 words