Refining Language Models with Compositional Explanations

Abstract

Pre-trained language models have been successful on text classification tasks, but are prone to learning spurious correlations from biased datasets, and are thus vulnerable when making inferences in a new domain. Prior work reveals such spurious patterns via post-hoc explanation algorithms which compute the importance of input features. Further, the model is regularized to align the importance scores with human knowledge, so that the unintended model behaviors are eliminated. However, such a regularization technique lacks flexibility and coverage, since only importance scores towards a pre-defined list of features are adjusted, while more complex human knowledge such as feature interaction and pattern generalization can hardly be incorporated. In this work, we propose to refine a learned language model for a target domain by collecting human-provided compositional explanations regarding observed biases. By parsing these explanations into executable logic rules, the human-specified refinement advice from a small set of explanations can be generalized to more training examples. We additionally introduce a regularization term allowing adjustments for both importance and interaction of features to better rectify model behavior. We demonstrate the effectiveness of the proposed approach on two text classification tasks by showing improved performance in target domain as well as improved model fairness after refinement.

Accepted to NeurIPS 2021. Camera-ready version. Code: https://github.com/INK-USC/expl-refinement

References in corpus (8)

• Distilling the Knowledge in a Neural Network
• Learning to Map Sentences to Logical Form: Structured Classification with Probabilistic Categorial Grammars
• Model Adaptation: Unsupervised Domain Adaptation without Source Data
• Visualizing Deep Neural Network Decisions: Prediction Difference Analysis
• Contextualizing Hate Speech Classifiers with Post-hoc Explanation
• Learning from Explanations with Neural Execution Tree
• Regularizing Black-box Models for Improved Interpretability (HILL 2019 Version)
• FIND: Human-in-the-Loop Debugging Deep Text Classifiers

Cited by in corpus (3)

• Informing clinical assessment by contextualizing post-hoc explanations of risk prediction models in type-2 diabetes
• Toward Operationalizing Pipeline-aware ML Fairness: A Research Agenda for Developing Practical Guidelines and Tools
• Clarify: Improving Model Robustness With Natural Language Corrections