Generating Hierarchical Explanations on Text Classification via Feature Interaction Detection

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5578–5593 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5578 Generating Hierarchical Explanations on Text Classification via Feature Interaction Detection Hanjie Chen, Guangtao Zheng, Yangfeng Ji Department of Computer Science University of Virginia Charlottesville, VA, USA {hc9mx, gz5hp, yangfeng}@virginia.edu Abstract Generating explanations for neural networks has become crucial for their applications in real-world with respect to reliability and trust- worthiness. In natural language processing, ex- isting methods usually provide important fea- tures which are words or phrases selected from an input text as an explanation, but ignore the interactions between them. It poses challenges for humans to interpret an explanation and con- nect it to model prediction. In this work, we build hierarchical explanations by detecting feature interactions. Such explanations visual- ize how words and phrases are combined at dif- ferent levels of the hierarchy, which can help users understand the decision-making of black- box models. The proposed method is evalu- ated with three neural text classifiers (LSTM, CNN, and BERT) on two benchmark datasets, via both automatic and human evaluations. Ex- periments show the effectiveness of the pro- posed method in providing explanations that are both faithful to models and interpretable to humans. 1 Introduction Deep neural networks have achieved remark- able performance in natural language processing (NLP) (Devlin et al., 2018; Howard and Ruder, 2018; Peters et al., 2018), but the lack of under- standing on their decision making leads them to be characterized as blackbox models and increases the risk of applying them in real-world applica- tions (Lipton, 2016; Burns et al., 2018; Jumelet and Hupkes, 2018; Jacovi et al., 2018). Understanding model prediction behaviors has been a critical factor in whether people will trust and use these blackbox models (Ribeiro et al., 2016). A typical work on understanding decision- making is to generate prediction explanations for each input example, called local explanation gen- eration. In NLP, most of existing work on local explanation generation focuses on producing word- level or phrase-level explanations by quantifying contributions of individual words or phrases to a model prediction (Ribeiro et al., 2016; Lundberg and Lee, 2017; Lei et al., 2016; Plumb et al., 2018). Figure 1: Different explanations for a NEGATIVE movie review a waste of good performance, where the color of each block represents the contribu- tion of the corresponding word/phrase/clause (feature) to the model prediction. From the hierarchical expla- nation, we obtain a set of features in each timestep (t), where the most important one is waste of good. Figure 1 (a) and (b) present a word-level and a phrase-level explanation generated by the LIME (Ribeiro et al., 2016) and the Contextual Decom- position (CD) (Murdoch et al., 2018) respectively for explaining sentiment classification. Both ex- planations provide scores to quantify how a word or a phrase contributes to the final prediction. For example, the explanation generated by LIME cap- tures a keyword waste and the explanation from CD identifies an important phrase …

Content truncated to 500 words