Evaluating the Impact of Using a Domain-specific Bilingual Lexicon on the Performance of a Hybrid Machine Translation Approach

Proceedings of Recent Advances in Natural Language Processing, pages 579–587, Hissar, Bulgaria, Sep 7–9 2015. Evaluating the Impact of Using a Domain-specific Bilingual Lexicon on the Performance of a Hybrid Machine Translation Approach Nasredine Semmar, Othman Zennaki, Meriama Laib CEA, LIST, Vision and Content Engineering Laboratory F-91191, Gif-sur-Yvette, France {nasredine.semmar,othman.zennaki,meriama.laib}@cea.fr Abstract This paper describes an Example-Based Ma- chine Translation prototype and presents an evaluation of the impact of using a domain- specific vocabulary on its performance. This prototype is based on a hybrid approach which needs only monolingual texts in the target lan- guage and consists to combine translation can- didates returned by a cross-language search engine with translation hypotheses provided by a finite-state transducer. The results of this combination are evaluated against a statistical language model of the target language in order to obtain the n-best translations. To measure the performance of this hybrid approach, we achieved several experiments using corpora on two domains from the European Parliament proceedings (Europarl) and the European Medicines Agency documents (Emea). The obtained results show that the proposed ap- proach outperforms the state-of-the-art Statis- tical Machine Translation system Moses when texts to translate are related to the specialized domain. 1 Introduction Current Machine Translation (MT) technology has serious limitations: there are, on the one hand, the rule-based systems which require hand- crafted linguistic rules and their manual con- struction is time consuming and expensive, and, on the other hand, the statistical systems which try to learn how to translate by analyzing the translation patterns found in large collections of human translations and these systems are effec- tive only when large amounts of parallel corpora are available. However, parallel corpora are only available for a limited number of language pairs and domains. In several fields, available corpora are not sufficient to make Statistical Machine Translation (SMT) approaches operational. We present, in this paper, an Example-Based Machine Translation (EBMT) prototype and we study the impact of using a domain-specific lexi- con on its performance. The EBMT prototype is based on a hybrid approach which uses only a monolingual corpus in the target language. This corpus is considered as a textual database of a cross-language search engine. For each sentence to translate (query in natural language), the cross-language search engine is used to provide a set of sentences in the target language. These sentences are combined with translation hypoth- eses provided by a finite-state transducer. The result of this combination is evaluated against a statistical language model learned from the target language corpus in order to produce the n-best translations. We believe that this is the first ap- plication of cross-language information retrieval in machine translation (Semmar and Bouamor 2011; Semmar et al., 2011; Semmar et al., 2014). The remainder of this paper is organized as follows: Section 2 describes the main approaches used in machine translation and presents previ- ous works addressing the task of domain adapta- tion in statistical machine translation. Section 3 introduces the hybrid approach used to imple- ment the EBMT prototype …

Content truncated to 500 words