Self-training with Few-shot Rationalization

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10702–10712 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10702 Self-training with Few-shot Rationalization Meghana Moorthy Bhat† Alessandro Sordoni‡ Subhabrata Mukherjee‡ †The Ohio State University ‡Microsoft Research bhat.89@osu.edu, {alsordon, submukhe}@microsoft.com Abstract While pre-trained language models have ob- tained state-of-the-art performance for several natural language understanding tasks, they are quite opaque in terms of their decision-making process. While some recent works focus on ra- tionalizing neural predictions by highlighting salient concepts in text as justifications or ratio- nales, they rely on thousands of labeled train- ing examples for both task labels as well as an- notated rationales for every instance. Such ex- tensive large-scale annotations are infeasible to obtain for many tasks. To this end, we de- velop a multi-task teacher-student framework based on self-training language models with limited task-specific labels and rationales, and judicious sample selection to learn from infor- mative pseudo-labeled examples1. We study several characteristics of what constitutes a good rationale and demonstrate that the neu- ral model performance can be significantly im- proved by making it aware of its rationalized predictions particularly in low-resource set- tings. Extensive experiments in several bench- mark datasets demonstrate the effectiveness of our approach. 1 Introduction Recent success in several natural language under- standing tasks can be attributed to training large- scale and complex neural network models. While these models work very well for specific tasks, they offer limited insights into their inner working and are often used as black-box predictors. To address these shortcomings, recent works (DeYoung et al., 2020; Paranjape et al., 2020; Yu et al., 2019) have focused on designing interpretable NLP systems that can explain the model’s predictions. A typical approach to study this decision-making process has been via annotating rationales as a short and suf- ficient part of the input text leading to the specific 1Code available at https://aka.ms/RationaleST prediction that can be also used as auxiliary supervi- sion for training. Appropriate use of such rationales can improve the downstream task performance as the model learns to focus on the task-relevant parts of the input (Pruthi et al., 2020b). However, a significant resource challenge is to obtain large-scale annotated rationales to train these models as explored in fully supervised setting in recent work (DeYoung et al., 2020). This re- quires models to have access to both instance-level task labels as well as token-level binary rationale labels depicting whether a token should be included in the rationale or not. Such extensive annotations are infeasible to obtain for many tasks, hence de- vising models that can effectively exploit a limited number of annotated rationales is utterly important. Therefore, our objective is two-fold: (1) improve downstream task performance and (2) improve ra- tionale extraction – with few labeled examples for the downstream task and corresponding rationales. Recent works (Mukherjee and Awadallah, 2020; Wang et al., 2020; Xie et al., 2019) on few-shot learning have explored self-training as a mecha- nism to train neural network models with limited …

Content truncated to 500 words