Rethinking Self-Attention: Towards Interpretability in Neural Parsing

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 731–742 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 731 Rethinking Self-Attention: Towards Interpretability in Neural Parsing Khalil Mrini1, Franck Dernoncourt2, Quan Tran2, Trung Bui2, Walter Chang2, and Ndapa Nakashole1 1 University of California, San Diego, La Jolla, CA 92093 khalil@ucsd.edu, nnakashole@eng.ucsd.edu 2Adobe Research, San Jose, CA 95110 {franck.dernoncourt, qtran, bui, wachang}@adobe.com Abstract Attention mechanisms have improved the per- formance of NLP tasks while allowing mod- els to remain explainable. Self-attention is cur- rently widely used, however interpretability is difficult due to the numerous attention distri- butions. Recent work has shown that model representations can benefit from label-specific information, while facilitating interpretation of predictions. We introduce the Label At- tention Layer: a new form of self-attention where attention heads represent labels. We test our novel layer by running constituency and dependency parsing experiments and show our new model obtains new state-of-the-art re- sults for both tasks on both the Penn Treebank (PTB) and Chinese Treebank. Additionally, our model requires fewer self-attention layers compared to existing work. Finally, we find that the Label Attention heads learn relations between syntactic categories and show path- ways to analyze errors. 1 Introduction Attention mechanisms (Bahdanau et al., 2014; Lu- ong et al., 2015) provide arguably explainable atten- tion distributions that can help to interpret predic- tions. For example, for their machine translation predictions, Bahdanau et al. (2014) show a heat map of attention weights from source language words to target language words. Similarly, in trans- former architectures (Vaswani et al., 2017), a self- attention head produces attention distributions from the input words to the same input words, as shown in the second row on the right side of Figure 1. However, self-attention mechanisms have multiple heads, making the combined outputs difficult to interpret. Recent work in multi-label text classification (Xiao et al., 2019) and sequence labeling (Cui and Zhang, 2019) shows the efficiency and interpretabil- ity of label-specific representations. We introduce Label Attention Head Self-Attention Head qi WKX i WVX i ai Repeated X Select the person Select the person Matrix Projection Select the person From other heads WKX i WVX i Ai Select the person WQX i Aggregating with output from other heads Select the person Matrix Projection Select the person Select the person X Select the person Figure 1: Comparison of the attention head architec- tures of our proposed Label Attention Layer and a Self- Attention Layer (Vaswani et al., 2017). The matrix X represents the input sentence “Select the person”. the Label Attention Layer: a modified version of self-attention, where each classification label corre- sponds to one or more attention heads. We project the output at the attention head level, rather than after aggregating all outputs, to preserve the source of head-specific information, thus allowing us to match labels to heads. To test our proposed Label Attention Layer, we build upon the parser of Zhou and Zhao (2019) and establish a new state of the art for both constituency …

Content truncated to 500 words