Ambiguity Reduction for Machine Translation: Human-Computer Collaboration

Ambiguity Reduction for Machine Translation: Human-Computer Collaboration Marcus Sammer† Kobi Reiter† Stephen Soderland† Katrin Kirchhoff‡ Oren Etzioni† Turing Center† University of Washington Seattle, WA 98195 Electrical Engineering‡ University of Washington Seattle, WA 98195 {sammer,dbgalur,soderlan,katrin,etzioni}@{cs†/ee‡}.washington.edu Abstract Statistical Machine Translation (SMT) accuracy degrades when there is only a limited amount of training, or when the training is not from the same domain or genre of text as the target application. However, cross-domain applications are typical of many real world tasks. We demonstrate that SMT accuracy can be improved in a cross-domain application by using a controlled language (CL) in- terface to help reduce lexical ambiguity in the input text. Our system, CL-MT, presents a monolingual user with a choice of word senses for each content word in the input text. CL-MT temporarily ad- justs the underlying SMT system's phrase table, boosting the scores of translations that include the word senses preferred by the user and lowering scores for disfa- vored translations. We demonstrate that this improves translation adequacy in 33.8% of the sentences in Spanish to En- glish translation of news stories, where the SMT system was trained on proceed- ings of the European Parliament. 1 Introduction Statistical Machine Translation (SMT) is sensitive to the genre and domain of the training data that is used to train the translation model. The best perfor- mance is typically achieved when the texts to be translated are drawn from the same population of texts as the training data. Unfortunately, many real world applications are for target domains or genres for which readily available parallel training corpo- ra do not exist. Mismatches between training and test data result in deteriorated performance. One source of translation errors is lexical ambi- guity in the input text, which may result in lexical errors in the translation. State-of-the-art SMT sys- tems use a phrase-based approach to translation (Och et al., 1999; Koehn et al., 2003; Tillmann, 2003; Zens and Ney, 2004), where translations are obtained by concatenating translations of chunks of words (as opposed to single words) in the input sentence. When training and test data are matched, a phrase-based SMT system can implicitly perform word-sense disambiguation (WSD) and choose the correct translation because the local context of the input word is taken into consideration. In fact, ad- ditional explicit WSD has not been shown to be helpful (e.g. Carpuat and Wu, 2005). Under mis- matched conditions, however, lexical ambiguity may become a much more significant source of er- rors because word senses occurring in the test data may never have occurred in similar context in the training data. In this paper we present CL-MT, a hybrid con- trolled language-statistical MT system where cross-domain SMT is improved by human guided WSD. This study is part of a larger research effort on utilizing machine translation technology to en- hance human-human communication. A typical application scenario is on-the-fly automatic email translation, where two users that do not share the same language are engaged in an email exchange, 193 Proceedings of the 7th …

Content truncated to 500 words