Interactive Rationale Extraction for Text Classification

Interactive Rationale Extraction for Text Classification Jiayi Dai1 and Mi-Young Kim2 and Randy Goebel1,3 1Department of Computing Science University of Alberta 2Department of Science, Augustana Faculty University of Alberta 3Alberta Machine Intelligence Institute University of Alberta {dai1, miyoung2, rgoebel}@ualberta.ca Abstract Deep neural networks show superior perfor- mance in text classification tasks, but their poor interpretability and explainability can cause trust issues. For text classification problems, the identification of textual sub-phrases or “ra- tionales” is one strategy for attempting to find the most influential portions of text, which can be conveyed as critical in making classification decisions. Selective models for rationale extrac- tion faithfully explain a neural classifier’s pre- dictions by training a rationale generator and a text classifier jointly: the generator identifies rationales and the classifier predicts a category solely based on the rationales. The selected rationales are then viewed as the explanations for the classifier’s predictions. Through ex- change of such explanations, humans interact to achieve higher performance in problem solving. To imitate the interactive process of humans, we propose a simple interactive rationale extrac- tion architecture that selects a pair of rationales and then makes predictions from two indepen- dently trained selective models. We show how this architecture outperforms both base models for text classification tasks on datasets IMDB movie reviews and 20 Newsgroups in terms of predictive performance. 1 Introduction Selective (or select-predict) models for rationale ex- traction in text classification (Lei et al., 2016; Bast- ings et al., 2019), with the general structure shown in Figure 1, are designed to extract a set of words, namely a rationale (Zaidan et al., 2007), from an original text where, for prediction purposes, the rationale is expected to be sufficient as the input for the classification model to obtain the same predic- tion based on the whole text. For the purpose of interpretability, the rationale should be concise and 0The implementation is provided on https://github. com/JiayiDai/RationaleExtraction. contiguous. A rationale extraction model is faith- ful (Lipton, 2018) if the extracted rationales are truly the information used for classification (Jain et al., 2020). The problem of extracting rationales that satisfy the criteria above is complex from a machine learning perspective and becomes more difficult with only instance-level supervision (i.e., without token-level annotations) (Jain et al., 2020). One model’s identification of rationales can suffer from high variance because of the complex train- ing process. An ensemble of more than one model helps to reduce variance, which leads to the explo- ration of how to take use of two rationale extraction models and how to make a choice when the two models make different predictions. When two humans have different answers to a problem, they tend to exchange their reasons or ex- planations, after which there might be a change of mind. To show why this interaction of humans is ef- fective, we use the problem of proving a mathemat- ical conjuncture as an instance: because searching for a correct mathematical proof, which then leads to a correct claim about the conjuncture, is …

Content truncated to 500 words