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

By Maha Elbayad, Laurent Besacier, Jakob Verbeek. 2018

84 citations recorded by the source corpus.

Current state-of-the-art machine translation systems are based on encoder-decoder architectures, 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 alternative approach which instead relies on a single 2D convolutional neural network across both sequences. Each layer of our network re-codes source tokens on the basis of the output sequence produced so far. Attention-like properties are therefore pervasive throughout the network. Our model yields excellent results, outperforming state-of-the-art encoder-decoder systems, while being conceptually simpler and having fewer parameters.

Subfields: Artificial Intelligence, Electrical and Electronic Engineering, Molecular Biology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Jean-Marc Andreoli
2. Masato Neishi
3. Naoki Yoshinaga
4. Bryan Eikema
5. Marc Dymetman

In Molecular Biology

1. Jean-Marc Andreoli

Citations in the corpus

• On the Relation between Position Information and Sentence Length in Neural Machine Translation (2019, background): “Only a few studies do the analysis on the translation quality in terms of sentence length (Elbayad et al., 2018; Zhang et al., 2019).”
• Convolution, attention and structure embedding. (2019, background): “We propose a generic model to achieve this, and claim that it captures the essence of many attention mechanisms: [19, 7, 20, 2, 3, 10].”
• Revisiting Low-Resource Neural Machine Translation: A Case Study (2019, unknown): “…et al., 2016) 5 21.8 BSO (Wiseman and Rush, 2016) 25.5 NPMT+LM (Huang et al., 2018) 30.1 MRT (Edunov et al., 2018) 32.84 ± 0.08 Pervasive Attention (Elbayad et al., 2018) 33.8 Transformer Baseline (Wu et al., 2019) 34.4 Dynamic Convolution (Wu et al., 2019) 35.2 our PBSMT (1) 28.19 ± 0.01 our NMT…”