Explorations in using grammatical dependencies for contextual phrase translation disambiguation

Explorations in Using Grammatical Dependencies for Contextual Phrase Translation Disambiguation Aur´elien Max1, Rafik Makhloufi1, and Philippe Langlais2 1 LIMSI-CNRS & Universit´e Paris-Sud 11, Orsay, France 2 DIRO, Universit´e de Montr´eal, Canada Abstract. Recent research has shown the importance of using source context information to disambiguate source phrases in phrase-based Sta- tistical Machine Translation. Although encouraging results have been obtained, those studies mostly focus on translating into a less inflected target language. In this article, we present an attempt at using source context information to translate from English into French. In addition to information extracted from the immediate context of a source phrase, we also exploit grammatical dependencies information. While automatic evaluation does not exhibit a significant difference, a manual evaluation we conducted provides evidence that our context-aware translation en- gine outperforms its context-insensitive counterpart. 1 Introduction One notable shortcoming of the now standard phrase-based approach to Statis- tical Machine Translation (PBSMT) [1] is that a unique conditional probability distribution p(e|f) is considered for translating all the occurrences of a source phrase f,3 while obviously, differences in translation may happen due to the context in which the phrase occurs. A systematic inspection of the output of a state-of-the-art PBSMT engine has been conducted by [2]. Among other interesting things, it shows that sense errors (including wrong lexical choice and source ambiguity) account for 21.9% of the observed errors when translating from English into Spanish, and for 28.2% when translating from Spanish into English. Incorrect forms (including incorrect verb tense or person, and incorrect gender or number agreement) account for 33.9% of errors when translating from English into Spanish, and for 9.9% from Spanish into English. Although the latter type of errors may be recovered to some extent by a target language model, those translation errors essentially arise because the engine fails to use grammatical dependencies present in the source sentences, such as subject-verb or verb-object relations. The integration of source contextual information in phrase-based SMT al- ready gave rise to a number of interesting works [3–7]. [3] and [4] embedded 3 We use the standard notation where f is a source phrase, and e is a target phrase. 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 114 existing word-sense disambiguation systems into a phrase-based SMT system. Both systems use local collocations and word and POS from the immediate con- text. In [3], bag-of-word context and basic dependency features are used. [5] trained Support Vector Machine classifiers for every possible candidate trans- lation of a source phrase. They considered words of the immediate context (5 tokens to the left and to the right), n-grams (up to size 3) of words, POS and lemmas, as well as chunking labels. [6] trained a global memory-based classi- fier that performs implicit smoothing of the probability estimates. The classifier makes use of words and POS information of the immediate context of the phrase (2 tokens to the left and to the right). The experimental conditions and the gains reported in the aforementioned studies differ significantly. However, …

Content truncated to 500 words