Unsupervised Token-wise Alignment to Improve Interpretation of Encoder-Decoder Models

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 74–81 Brussels, Belgium, November 1, 2018. c⃝2018 Association for Computational Linguistics 74 Unsupervised Token-wise Alignment to Improve Interpretation of Encoder-Decoder Models Shun Kiyono 1∗ Sho Takase 2 Jun Suzuki 1,2,4 Naoaki Okazaki 3 Kentaro Inui 1,4 Masaaki Nagata 2 1 Tohoku University 3 Tokyo Institute of Technology2 NTT Communication Science Laboratories 4 RIKEN Center for Advanced Intelligence Project {kiyono,jun.suzuki,inui}@ecei.tohoku.ac.jp, {takase.sho, nagata.masaaki}@lab.ntt.co.jp, okazaki@c.titech.ac.jp Abstract Developing a method for understanding the in- ner workings of black-box neural methods is an important research endeavor. Convention- ally, many studies have used an attention ma- trix to interpret how Encoder-Decoder-based models translate a given source sentence to the corresponding target sentence. However, re- cent studies have empirically revealed that an attention matrix is not optimal for token-wise translation analyses. We propose a method that explicitly models the token-wise align- ment between the source and target sequences to provide a better analysis. Experiments show that our method can acquire token-wise align- ments that are superior to those of an attention mechanism1. 1 Introduction The Encoder-Decoder model with an attention mechanism (EncDec) (Sutskever et al., 2014; Cho et al., 2014; Bahdanau et al., 2015; Luong et al., 2015) has been an epoch-making development that has led to great progress in many natural language generation tasks, such as machine translation (Bah- danau et al., 2015), dialog generation (Shang et al., 2015), and headline generation (Rush et al., 2015). An enormous number of studies have attempted to enhance the ability of EncDec. Furthermore, several intensive studies have also attempted to analyze and interpret the inside of black-box EncDec models, especially how they translate a given source sentence to the correspond- ing target sentence (Ding et al., 2017). One typical approach to this is to visualize an attention matrix, which is a collection of attention vectors (Bahdanau et al., 2015; Luong et al., 2015; Tu et al., 2016). ∗ This work is a product of collaborative research pro- gram of Tohoku University and NTT Communication Science Laboratories. 1Our code for reproducing the experiments is available at https://github.com/butsugiri/UAM The assumption behind this interpretation is that the attention matrix has a “soft” token-wise align- ment between the source and target sequences, and thus we can use EncDec models to skim which to- kens in the source are converted to which tokens in the target. However, recent studies have empirically re- vealed that an attention model can operate not only for token-wise alignment but also for other func- tionalities, such as reordering (Ghader and Monz, 2017; Liu et al., 2016). In addition, Luong et al. (2015) reported that the quality of attention matrix- based alignment is quantitatively inferior to that of the Berkeley aligner (Liang et al., 2006). Koehn and Knowles (2017) also reported that attention matrix-based alignment is significantly different from that acquired from an off-the-shelf aligner for English-German language pairs. From these re- cent findings, the goal of this paper is to provide a method that …

Content truncated to 500 words