Zero-Shot Rationalization by Multi-Task Transfer Learning from Question Answering

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2187–2197 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2187 Zero-Shot Rationalization by Multi-Task Transfer Learning from Question Answering Po-Nien Kung Tse-Hsuan Yang Yi-Cheng Chen Sheng-Siang Yin Department of Computer Science and Information EngineeringYun-Nung Chen National Taiwan University, Taipei, Taiwan {b06902012, b06902032, b06902011, b06902103}@ntu.edu.tw y.v.chen@ieee.org Abstract Extracting rationales can help human under- stand which information the model utilizes and how it makes the prediction towards better in- terpretability. However, annotating rationales requires much effort and only few datasets contain such labeled rationales, making super- vised learning for rationalization difficult. In this paper, we propose a novel approach that leverages the benefits of both multi-task learn- ing and transfer learning for generating ratio- nales through question answering in a zero- shot fashion. For two benchmark rationaliza- tion datasets, the proposed method achieves comparable or even better performance of ra- tionalization without any supervised signal, demonstrating the great potential of zero-shot rationalization for better interpretability.1 1 Introduction Resolving NLP tasks by deep neural networks has been proven to be effective, and it is also important to investigate how the models make such a decision. For example, only providing the prediction to med- ical tasks may not be enough, and providing the associated reasons is more crucial for the practical applications. Therefore, there has been increasing attempts that focus on interpretability or explain- ability of the machine-learned models. There are different ways of explaining how machines make the decision (Ribeiro et al., 2016; Alvarez-Melis and Jaakkola, 2017; Lee et al., 2019; Liu et al., 2019a), and one of these methods is to extract ra- tionales (Lei et al., 2016; DeYoung et al., 2019). However, most prior work focused on extracting rationales in a supervised manner (DeYoung et al., 2019), but not all datasets contain such annotated rationales for model learning, making the rational- ization task difficult and impractical. 1The source code and the processed data is available at: https://github.com/MiuLab/ ZeroShotRationale. Rationalization is defined as a task that focuses on extracting the rationales from the input texts for better justification and interpretation. Lei et al. (2016) is the first work that attempted to extract rationales in order to justify the model’s answers, where a rationale generator extracts the context and a predictor generates the answer based on the ex- tracted rationales. This method shows great preci- sion in extracting rationales. Recently, Yu et al. (2019) proposed an introspective model, an ex- tension of the prior work that further improved the comprehensiveness of the extracted rationales. Moreover, DeYoung et al. (2019) proposed to learn rationale extraction in a supervised manner and prepared the benchmark experiments in diverse rationalization tasks. From the experimental re- sults, it can be found that supervised learning for rationalization may not be always better than the unsupervised method due to the complex reasoning process. Considering that in the practical application, the target domain may not contain the annotated ra- tionales for supervised training, transferring the knowledge about rationalization to …

Content truncated to 500 words