Arabic-English Semantic Word Class Alignment to Improve Statistical Machine Translation

Proceedings of Recent Advances in Natural Language Processing, pages 663–671, Hissar, Bulgaria, Sep 7–9 2015. Arabic-English Semantic Class Alignment to Improve Statistical Machine Translation Ines Turki Khemakhem MIRACL Laboratory, University of Sfax-TUNISIA ines_turki@yahoo.fr Salma Jamoussi MIRACL Laboratory, University of Sfax- TUNISIA salma.jamoussi @isimsf.rnu.tn Abdelmajid Ben Hamadou MIRACL Laboratory, University of Sfax-TUNISIA abdelmajid.benhamadou @isimsf.rnu.tn Abstract Clustering words is a widely used technique in statistical natural language processing. It re- quires syntactic, semantic, and contextual fea- tures. Especially, semantic clustering is gain- ing a lot of interest. It consists in grouping a set of words expressing the same idea or shar- ing the same semantic properties. In this paper, we present a new method to in- tegrate semantic classes in a Statistical Ma- chine Translation (SMT) context to improve the Arabic-English translation quality. In our method, we first apply a semantic word clustering algorithm for English. We then pro- ject the obtained semantic word classes from the English side to the Arabic side. This pro- jection is based on available word alignments provided by the alignment step using GIZA++ tool. Finally, we apply a new process to incor- porate semantic classes in order to improve the SMT quality. The experimental results show that introducing semantic word classes achieves 4 % of relative improvement on the BLEU score for the Arabic → English transla- tion task. 1 Introduction In the past decade, statistical machine translation (SMT) has been advanced from word based SMT to phrase and syntax based SMT. Although this advancement produces major improvements in BLEU scores, important meaning errors still harm the quality of SMT translations. More recently, research in statistical machine translation has witnessed many attempts to inte- grate semantic feature into SMT models, to gen- erate not only grammatical but also meaning pre- served translations. Integrating semantic features into SMT tasks aims at improving translation adequacy. In a bi- lingual corpus, different senses of words in the source language can have different translations in the target language, as the context in which they appear. This motivates the introduction of semantic word classes in statistical machine translation. A semantic word class is represented by a set of words expressing the same idea and sharing the same semantic properties. For example, the words plane, train, boat, bus can all correspond to the semantic class “transport”. Semantic word clustering is a technique for parti- tioning sets of words into subsets of semantically similar words. It is increasingly becoming a ma- jor technique used in SMT task. Furthermore, most of the SMT system well suit- ed for processing English and other languages with a relatively rigid word order, while lan- guages with complicated morphological para- digms still pose difficulties as Arabic. In this paper, we present a new method to inte- grate the underlined semantic classes in a SMT context to improve the Arabic-English transla- tion quality. We first describe the semantic word clustering algorithm for English and we proceed to directly project the obtained semantic word classes from English side into Arabic …

Content truncated to 500 words