Meta-Learning with Variational Semantic Memory for Word Sense Disambiguation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 477–488 August 1–6, 2021. ©2021 Association for Computational Linguistics 477 Improving the Faithfulness of Attention-based Explanations with Task-specific Information for Text Classification George Chrysostomou Nikolaos Aletras Department of Computer Science, University of Sheffield United Kingdom {gchrysostomou1, n.aletras}@sheffield.ac.uk Abstract Neural network architectures in natural lan- guage processing often use attention mech- anisms to produce probability distributions over input token representations. Attention has empirically been demonstrated to im- prove performance in various tasks, while its weights have been extensively used as explanations for model predictions. Re- cent studies (Jain and Wallace, 2019; Ser- rano and Smith, 2019; Wiegreffe and Pin- ter, 2019) have showed that it cannot gener- ally be considered as a faithful explanation (Jacovi and Goldberg, 2020) across encoders and tasks. In this paper, we seek to im- prove the faithfulness of attention-based expla- nations for text classification. We achieve this by proposing a new family of Task-Scaling (TaSc) mechanisms that learn task-specific non-contextualised information to scale the original attention weights. Evaluation tests for explanation faithfulness, show that the three proposed variants of TaSc improve attention- based explanations across two attention mech- anisms, five encoders and five text classifica- tion datasets without sacrificing predictive per- formance. Finally, we demonstrate that TaSc consistently provides more faithful attention- based explanations compared to three widely- used interpretability techniques.1 1 Introduction Natural Language Processing (NLP) approaches for text classification are often underpinned by large neural network models (Cho et al., 2014; De- vlin et al., 2019). Despite the high accuracy and efficiency of these models in dealing with large amounts of data, an important problem is their in- creased complexity that makes them opaque and hard to interpret by humans which usually treat 1Code is available at: https://github.com/ GChrysostomou/tasc.git them as black boxes (Zhang et al., 2018; Linzen et al., 2019). Attention mechanisms (Bahdanau et al., 2015) produce a probability distribution over the input to compute a vector representation of the entire token sequence as the weighted sum of its con- stituent vectors. A common practice is to provide explanations for a given prediction and qualitative model analysis by assigning importance to input tokens using scores provided by attention mecha- nisms (Chen et al., 2017; Wang et al., 2016; Jain et al., 2020; Sun and Lu, 2020) as a mean towards model interpretability (Lipton, 2016; Miller, 2019). A faithful explanation is one that accurately rep- resents the true reasoning behind a model’s pre- diction (Jacovi and Goldberg, 2020). A series of recent studies illustrate that explanations obtained by attention weights do not always provide faith- ful explanations (Serrano and Smith, 2019) while different text encoders can affect attention inter- pretability, e.g. results can differ when using a recurrent or non-recurrent encoder (Wiegreffe and Pinter, 2019). A limitation of attention as an indicator of input importance is that it refers to the word in context due to information mixing in the model …

Content truncated to 500 words