Generating Token-Level Explanations for Natural Language Inference

Proceedings of NAACL-HLT 2019, pages 963–969 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 963 Generating Token-Level Explanations for Natural Language Inference James Thorne University of Cambridge jt719@cam.ac.uk Andreas Vlachos University of Cambridge av308@cam.ac.uk Christos Christodoulopoulos Amazon chrchrs@amazon.co.uk Arpit Mittal Amazon mitarpit@amazon.co.uk Abstract The task of Natural Language Inference (NLI) is widely modeled as supervised sentence pair classification. While there has been a lot of work recently on generating explanations of the predictions of classifiers on a single piece of text, there have been no attempts to gen- erate explanations of classifiers operating on pairs of sentences. In this paper, we show that it is possible to generate token-level expla- nations for NLI without the need for training data explicitly annotated for this purpose. We use a simple LSTM architecture and evaluate both LIME and Anchor explanations for this task. We compare these to a Multiple Instance Learning (MIL) method that uses thresholded attention make token-level predictions. The approach we present in this paper is a novel extension of zero-shot single-sentence tagging to sentence pairs for NLI. We conduct our ex- periments on the well-studied SNLI dataset that was recently augmented with manually annotation of the tokens that explain the en- tailment relation. We find that our white-box MIL-based method, while orders of magnitude faster, does not reach the same accuracy as the black-box methods. 1 Introduction Large-scale datasets for Natural Language Infer- ence (NLI) (Bowman et al., 2015; Williams et al., 2018) have enabled the development of many deep-learning models (Rockt¨aschel et al., 2016; Peters et al., 2018; Radford et al., 2018). The task is modeled as 3-way classification of the entail- ment relation between a pair of sentences. Model performance is assessed through accuracy on a held-out test set. While state-of-the-art models achieve high accuracy, their complexity makes it difficult to interpret their behavior. Explaining the predictions made by classifiers has been of increasing concern (Doshi-Velez and Premise: Children smiling and waving at a camera Hypothesis: The kids are frowning Label: Contradiction Figure 1: Example of token-level highlights from the e- SNLI dataset (Camburu et al., 2018). Annotators were provided a premise and hypothesis and asked to high- light words considered essential to explain the label. Kim, 2017). It has been studied in natural lan- guage processing through both black-box analy- sis, and through modifications to the models un- der investigation; we refer to the latter approaches as white-box. Common black-box techniques gen- erate explanations of predictions through training meta-models by perturbing input tokens (Ribeiro et al., 2016; Nguyen, 2018; Ribeiro et al., 2018) or through interpretation of model sensitivity to input tokens (Li et al., 2016; Feng et al., 2018). White- box methods induce new features (Aubakirova and Bansal, 2016), augment models to generate expla- nations accompanying their predictions (Lei et al., 2016; Camburu et al., 2018), or expose model in- ternals such as magnitude of hidden states (Linzen et al., 2016), gradients (as a proxy for model sensi- tivity to input …

Content truncated to 500 words