Improving Phrase-Based SMT Using Cross-Granularity Embedding Similarity

• Published in: Baltic Journal of Modern Computing Vol. 4; no. 2; p. 129
• Main Authors: Passban, Peyman, Hokamp, Chris, Way, Andy, Liu, Qun
• Format: Journal Article
• Language: English
• Published: Riga University of Latvia 01-01-2016
• Subjects: Machine translation; Neural networks; Segments; Semantics; Similarity; Tables (data); Training; Translations
• ISSN: 2255-8942; 2255-8950

The phrase-based statistical machine translation (PBSMT) model can be viewed as a log-linear combination of translation and language model features. Such a model typically relies on the phrase table as the main resource for bilingual knowledge, which in its most basic form consists of aligned phrases, along with four probability scores. These scores only indicate the co-occurrence of phrase pairs in the training corpus, and not necessarily their semantic relatedness. The basic phrase table is also unable to incorporate contextual information about the segments where a particular phrase tends to occur. In this paper, we define six new features which express the semantic relatedness of bilingual phrases. Our method utilizes both source and target side information to enrich the phrase table. The new features are inferred from a bilingual corpus by a neural network (NN). We evaluate our model on the English-Farsi (En-Fa) and English-Czech (En-Cz) pairs and observe considerable improvements in the all En[Lef-right arrow]Fa and En[Lef-right arrow]Cz directions.