Pervasive Attention: 2D Convolutional Neural Networks for Sequence-to-Sequence Prediction

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 97–107 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 97 Pervasive Attention: 2D Convolutional Neural Networks for Sequence-to-Sequence Prediction Maha Elbayad1,2 Laurent Besacier1 Jakob Verbeek2 Univ. Grenoble Alpes, CNRS, Grenoble INP, Inria, LIG, LJK, F-38000 Grenoble France 1 firstname.lastname@univ-grenoble-alpes.fr 2 firstname.lastname@inria.fr Abstract Current state-of-the-art machine translation systems are based on encoder-decoder archi- tectures, that first encode the input sequence, and then generate an output sequence based on the input encoding. Both are interfaced with an attention mechanism that recombines a fixed encoding of the source tokens based on the decoder state. We propose an alterna- tive approach which instead relies on a sin- gle 2D convolutional neural network across both sequences. Each layer of our network re- codes source tokens on the basis of the out- put sequence produced so far. Attention-like properties are therefore pervasive throughout the network. Our model yields excellent re- sults, outperforming state-of-the-art encoder- decoder systems, while being conceptually simpler and having fewer parameters. 1 Introduction Deep neural networks have made a profound im- pact on natural language processing technology in general, and machine translation in particular (Blunsom, 2013; Sutskever et al., 2014; Cho et al., 2014; Jean et al., 2015; LeCun et al., 2015). Ma- chine translation (MT) can be seen as a sequence- to-sequence prediction problem, where the source and target sequences are of different and vari- able length. Current state-of-the-art approaches are based on encoder-decoder architectures (Blun- som, 2013; Sutskever et al., 2014; Cho et al., 2014; Bahdanau et al., 2015). The encoder “reads” the variable-length source sequence and maps it into a vector representation. The decoder takes this vector as input and “writes” the target sequence, updating its state each step with the most recent word that it generated. The basic encoder-decoder model is generally equipped with an attention model (Bahdanau et al., 2015), which repetitively re-accesses the source sequence during the decod- ing process. Given the current state of the decoder, a probability distribution over the elements in the source sequence is computed, which is then used to select or aggregate features of these elements into a single “context” vector that is used by the decoder. Rather than relying on the global rep- resentation of the source sequence, the attention mechanism allows the decoder to “look back” into the source sequence and focus on salient positions. Besides this inductive bias, the attention mecha- nism bypasses the problem of vanishing gradients that most recurrent architectures encounter. However, the current attention mechanisms have limited modeling abilities and are generally a simple weighted sum of the source representations (Bahdanau et al., 2015; Luong et al., 2015), where the weights are the result of a shallow matching between source and target elements. The atten- tion module re-combines the same source token codes and is unable to re-encode or re-interpret the source sequence while decoding. To address these limitations, we propose an al- ternative neural MT architecture, based …

Content truncated to 500 words