Translation similarity model based on bilingual compositional semantics

The authors propose a translation similarity model based on bilingual compositional semantics to integrate the bilingual semantic similarity feature into decoding process to improve translation quality. In the proposed model, monolingual compositional vectors for phrases are obtained at the source and target side respectively using distributional approach. These monolingual vectors are then projected onto the same semantic space and therefore transformed into bilingual compositional vectors. Base on this semantic space, translation similarity between source phrases and their corresponding target phrases is calculated. The similarities are integrated into the decoder as a new feature. Experiments on Chinese-to-English NIST06 and NIST08 test sets show that the proposed model significantly outperforms the baseline by 0.56 and 0.42 BLEU points respectively.