Enjoy the Salience: Towards Better Transformer-based Faithful Explanations with Word Salience

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8189–8200 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8189 Enjoy the Salience: Towards Better Transformer-based Faithful Explanations with Word Salience George Chrysostomou Nikolaos Aletras Department of Computer Science, University of Sheffield United Kingdom {gchrysostomou1, n.aletras}@sheffield.ac.uk Abstract Pretrained transformer-based models such as BERT have demonstrated state-of-the-art pre- dictive performance when adapted into a range of natural language processing tasks. An open problem is how to improve the faithfulness of explanations (rationales) for the predictions of these models. In this paper, we hypoth- esize that salient information extracted a pri- ori from the training data can complement the task-specific information learned by the model during fine-tuning on a downstream task. In this way, we aim to help BERT not to forget assigning importance to informative input to- kens when making predictions by proposing SALOSS; an auxiliary loss function for guid- ing the multi-head attention mechanism dur- ing training to be close to salient information extracted a priori using TextRank. Experi- ments for explanation faithfulness across five datasets, show that models trained with SA- LOSS consistently provide more faithful ex- planations across four different feature attribu- tion methods compared to vanilla BERT. Us- ing the rationales extracted from vanilla BERT and SALOSS models to train inherently faith- ful classifiers, we further show that the lat- ter result in higher predictive performance in downstream tasks.1 1 Introduction Pretrained transformer-based (Vaswani et al., 2017) language models (LMs) such as BERT (Devlin et al., 2019), have achieved state-of-the-art results in various language understanding tasks (Wang et al., 2019b,a). Despite their success, their highly complex nature consisting of millions of parame- ters, makes them difficult to interpret (Jain et al., 2020). This has motivated new research on under- standing and explaining their predictions. Previous work has explored whether LMs en- code syntactic knowledge by studying their multi- 1Code: https://github.com/GChrysostomou/ saloss. head attention distributions (Clark et al., 2019; Htut et al., 2019; Voita et al., 2019). Recent studies have evaluated the faithfulness of explanations2 for predictions made by these models (Vashishth et al., 2019; Atanasova et al., 2020; Jain et al., 2020). In general, LMs can provide faithful ex- planations, particularly using attention (Jain et al., 2020), but still fall behind other simpler architec- tures (Atanasova et al., 2020) possibly due to in- creased information mixing and higher contextual- ization in the model (Brunner et al., 2020; Pascual et al., 2021; Tutek and Snajder, 2020). Recent stud- ies have attempted to improve the explainability of non transformer-based models, by guiding them through an auxiliary objective towards informative input importance distributions (e.g. human or ad- versarial priors) (Ross et al., 2017a; Liu and Avci, 2019; Moradi et al., 2021). In a similar direction, we propose Salient Loss (SALOSS), an auxiliary objective that allows the multi-head attention of the model to learn from salient information (i.e. token importance) during training to reduce the effects of information mix- ing (Pascual et al., 2021). We compute a priori …

Content truncated to 500 words