RANCC: Rationalizing Neural Networks via Concept Clustering

Proceedings of the 28th International Conference on Computational Linguistics, pages 3214–3224 Barcelona, Spain (Online), December 8-13, 2020 3214 RANCC: Rationalizing Neural Networks via Concept Clustering Housam Khalifa Bashier Babiker 1, Mi-Young Kim2, Randy Goebel1 Alberta Machine Intelligence Institute 1 Department of Computing Science, University of Alberta 2 Department of Science, Augustana Faculty, University of Alberta {khalifab, miyoung2, rgoebel}@ulaberta.ca Abstract We propose a new self-explainable model for Natural Language Processing (NLP) text classifi- cation tasks. Our approach constructs explanations concurrently with the formulation of classifi- cation predictions. To do so, we extract a rationale from the text, then use it to predict a concept of interest as the final prediction. We provide three types of explanations: 1) rationale extraction, 2) a measure of feature importance, and 3) clustering of concepts. In addition, we show how our model can be compressed without applying complicated compression techniques. We exper- imentally demonstrate our explainability approach on a number of well-known text classification datasets. 1 Introduction Deep neural network (DNN) models provide powerful and sophisticated approaches for addressing Nat- ural Language Processing (NLP) text classification tasks. Yet their underlying behaviour is often opaque, especially in sensitive domains which can critically influence a user’s decision, such as in legal and med- ical domains. As a result, we need to create DNNs that are explainable, in other words, that can provide explanations for their predictions. Our work here focuses on understanding predictions made by a DNN model, to provide explanations at inference time. There has, alternatively, been attention on models at- tempt to make a neural network explainable, for instance (Sundararajan et al., 2017) and (Ribeiro et al., 2016), which create post-hoc explanations to support explainability. There are also other methods which focused on learning explanations concurrently with the prediction. For example, (Lei et al., 2016) and (Bastings et al., 2019) proposed a neural network architecture for text classification which “justifies” its predictions by selecting relevant tokens in the input text. But this interpretable representation is then ad- justed by a complex neural network, so the method is transparent as to what aspect of the input it uses for prediction, but not how it captures the salient features. Our work also focuses on extracting a rationale (which can also be defined as an excerpt or justification) simultaneously while computing a prediction. Figure 1: Our alternative explanations for NLP text classification. In addition, we insist that explanations are provided interactively, so that a user can switch from one explanation to another to gain a better understanding about the class prediction (sometimes referred to as al- ternative explanations) as in the work of (Atakishiyev et al., 2020)). Overall, we in- vestigate how to improve the level of ab- straction for DNNs and go beyond mea- suring feature importance. To address this problem, we present a new alternative ex- planation mechanism, i.e., clustering simi- lar rationales into distinct clusters concur- rently with rationale extraction. Our explanation mechanism for DNNs can be summarized as follows: This work is licensed under a …

Content truncated to 500 words