Why do you think that? Exploring Faithful Sentence-Level Rationales Without Supervision

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1080–1095 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1080 Why do you think that? Exploring Faithful Sentence–Level Rationales Without Supervision Max Glockner and Ivan Habernal and Iryna Gurevych Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt https://www.ukp.tu-darmstadt.de Abstract Evaluating the trustworthiness of a model’s prediction is essential for differentiating be- tween ‘right for the right reasons’ and ‘right for the wrong reasons’. Identifying textual spans that determine the target label, known as faithful rationales, usually relies on pipeline approaches or reinforcement learning. How- ever, such methods either require supervision and thus costly annotation of the rationales or employ non-differentiable models. We pro- pose a differentiable training–framework to create models which output faithful rationales on a sentence level, by solely applying super- vision on the target task. To achieve this, our model solves the task based on each rationale individually and learns to assign high scores to those which solved the task best. Our eval- uation on three different datasets shows com- petitive results compared to a standard BERT blackbox while exceeding a pipeline counter- part’s performance in two cases. We further exploit the transparent decision–making pro- cess of these models to prefer selecting the correct rationales by applying direct supervi- sion, thereby boosting the performance on the rationale–level.1 1 Introduction Large pre-trained language models, such as BERT (Devlin et al., 2018) or RoBERTa (Liu et al., 2019b) gain impressive results on a large variety of NLP tasks, including reasoning and inference (Rogers et al., 2020). Despite this success, research shows that their strong performance can rely, to some extent, on dataset–specific artifacts and not necessarily on the ability to solve the underlying task (Gururangan et al., 2018; Schuster et al., 2019; Gardner et al., 2020). Thus, these observations un- dermine the models’ trustworthiness and impede 1Code available at https://github.com/UKPLab/ emnlp2020-faithful-rationales Lorem ipsum dolor sit amet, consectetur adipisici elit, sed eiusmod tempor incidunt ut labore et dolore magna aliqua. Robbie Collins is a British film critic. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquid ex ea ommodi consequat. Robbie Collins is a British film critic. SUPPORT Robbie Collin is British. Select Predict Document Faithful explanation Target label Query Figure 1: Example of the proposed rationale selecting process on one of the datasets (FEVER): Given a query and a document, our model selects the best rationale and predicts the label solely based on this selection. their deployment in situations where ‘blindly trust- ing’ the model is deemed irresponsible (Sokol and Flach, 2020). Explainability has thus emerged as an increasingly popular field (Gilpin et al., 2018; Guidotti et al., 2018). We aim at faithful explanations – the identifi- cation of the actual reason for the model’s predic- tion, which is essential for accountability, fairness, and credibility (Chakraborty et al., 2017; Wu and Mooney, 2019) to evaluate whether a model’s pre- diction is based on the correct evidence. The re- cently published …

Content truncated to 500 words