Learning Variational Word Masks to Improve the Interpretability of Neural Text Classifiers

Learning Variational Word Masks to Improve the Interpretability of Neural Text Classifiers Hanjie Chen, Yangfeng Ji Department of Computer Science University of Virginia Charlottesville, VA, USA {hc9mx, yangfeng}@virginia.edu Abstract To build an interpretable neural text classifier, most of the prior work has focused on design- ing inherently interpretable models or finding faithful explanations. A new line of work on improving model interpretability has just started, and many existing methods require ei- ther prior information or human annotations as additional inputs in training. To address this limitation, we propose the variational word mask (VMASK) method to automatically learn task-specific important words and reduce irrel- evant information on classification, which ulti- mately improves the interpretability of model predictions. The proposed method is evaluated with three neural text classifiers (CNN, LSTM, and BERT) on seven benchmark text classifi- cation datasets. Experiments show the effec- tiveness of VMASK in improving both model prediction accuracy and interpretability. 1 Introduction Neural network models have achieved remarkable performance on text classification due to their ca- pacity of representation learning on natural lan- guage texts (Zhang et al., 2015; Yang et al., 2016; Joulin et al., 2017; Devlin et al., 2018). However, the lack of understanding of their prediction behav- iors has become a critical issue for reliability and trustworthiness and hindered their applications in the real world (Lipton, 2016; Ribeiro et al., 2016; Jacovi and Goldberg, 2020). Many explanation methods have been proposed to provide post-hoc explanations for neural networks (Ribeiro et al., 2016; Lundberg and Lee, 2017; Sundararajan et al., 2017), but they are only able to explain model pre- dictions and cannot help improve their interpretabil- ity. In this work, we consider interpretability as an intrinsic property of neural network models. Fur- thermore, we hypothesize that neural network mod- els with similar network architectures could have Ex. Model Text & Explanation 1 A An exceedingly clever piece of cinema B An exceedingly clever piece of cinema 2 A It becomes gimmicky instead of compelling B It becomes gimmicky instead of compelling Table 1: Model A and B are two neural text classifiers with similar network architectures. They all make cor- rect sentiment predictions on both texts (ex. 1: posi- tive; ex. 2: negative). Two post-hoc explanation meth- ods, LIME (Ribeiro et al., 2016) and SampleShapley (Kononenko et al., 2010), are used to explain the model predictions on example 1 and 2 respectively. Top three important words are shown in pink or blue for model A and B. Whichever post-hoc method is used, explana- tions from model B are easier to understand because the sentiment keywords “clever” and “gimmicky” are highlighted. different levels of interpretability, even though they may have similar prediction performance. Table 1 shows explanations extracted from two neural text classifiers with similar network architectures.1 Al- though both models make correct predictions of the sentiment polarities of two input texts (positive for example 1 and negative for example 2), they have different explanations for their predictions. In both examples, no matter which explanation gen- …

Content truncated to 500 words