Deps-SAN: Neural Machine Translation with Dependency-Scaled Self-Attention Network

Abstract

The neural machine translation model assumes that syntax knowledge can be learned from the bilingual corpus via an attention network automatically. However, the attention network trained in weak supervision actually cannot capture the deep structure of the sentence. Naturally, we expect to introduce external syntax knowledge to guide the learning of the attention network. Thus, we propose a novel, parameterfree, dependency-scaled self-attention network, which integrates explicit syntactic dependencies into the attention network to dispel the dispersion of attention distribution. Finally, two knowledge sparse techniques are proposed to prevent the model from overfitting noisy syntactic dependencies. Experiments and extensive analyses on the IWSLT14 German-to-English and WMT16 German-to-English translation tasks validate the effectiveness of our approach. 1

HybridVocab: Towards Multi-Modal Machine Translation via Multi-Aspect Alignment

“…Previous research of Transformer [29,32,46] found that lower networks are inclined to focus on structural information among words, while higher layers (usually above the 4th layer) preferred to capture sentence-level semantics. Based on this cognition, we fuse visual features that are semantically different from the text at higher layers, which moderate the corruption in the original semantic modelling of sentences.…”