Use of Domain-Specific Language Resources in Machine Translation

Use of Domain-Specific Language Resources in Machine Translation Sanja ˇStajner, Andreia Querido, Nuno Rendeiro, Jo˜ao Rodrigues, and Ant´onio Branco Department of Informatics, Faculty of Sciences University of Lisbon, Portugal {sanja.stajner, joao.rodrigues, andreia.querido, nuno.rendeiro, antonio.branco}@di.fc.ul.pt Abstract In this paper, we address the problem of Machine Translation (MT) for a specialised domain in a language pair for which only a very small domain-specific parallel corpus is available. We conduct a series of experiments using a purely phrase-based SMT (PBSMT) system and a hybrid MT system (TectoMT), testing three different strategies to overcome the problem of the small amount of in-domain training data. Our results show that adding a small size in-domain bilingual terminology to the small in-domain training corpus leads to the best improvements of a hybrid MT system, while the PBSMT system achieves the best results by adding a combination of in-domain bilingual terminology and a larger out-of-domain corpus. We focus on qualitative human evaluation of the output of two best systems (one for each approach) and perform a systematic in-depth error analysis which revealed advantages of the hybrid MT system over the pure PBSMT system for this specific task. Keywords: Qualitative human evaluation, domain-specific machine translation, hybrid machine translation, TectoMT 1. Introduction Phrase-based statistical machine translation (PBSMT) sys- tems are considered the state of the art for language pairs for which large amounts of parallel data for training are available. For vast majority of language pairs (English- Portuguese among them), however, the available corpora are usually limited on one or two particular domains, e.g. legal documents (JRC-Acquis corpus), or parliamentary discussions (Europarl corpus). In those cases, for domain- specific MT, it is believed that rule-based or hybrid MT systems have higher potential to overcome the problems of data sparsity. In this paper, we focus on English to Portuguese MT for Information Technology (IT) domain, motivated by the fol- lowing real world usage scenario: an user asks a question in Portuguese, the question is machine translated into En- glish, the answer is searched for in an English database, automatically translated back to Portuguese and presented to the user (Figure 1). Figure 1: The task. All experiments have been performed within the QTLeap project,1 which aims to investigate an articulated method- ology for machine translation based on deep language engi- neering approaches. As no freely available corpora for the EN-PT language pair exists for the IT domain, a small EN- PT corpus was compiled under the QTLeap project in order to enable in-domain examples and guide the hand-crafted rules for the synthesis phase of the hybrid MT system being 1www.qtleap.eu developed. This QTLeap corpus consists of recorded inter- actions of real users with experts to obtain technical support via chat, which were translated by professional translators into the eight languages of the project.2 We perform a series of MT experiments using two sys- tems: (1) the TectoMT (ˇZabokrtsk´y et al., 2008) adapted to English-Portuguese translation (Silva et al., 2015), as a hybrid system with hybrid analysis, rule-based synthesis and statistically based transfer, …

Content truncated to 500 words