Using Noisy Biligual Data for Statistical Machine Translation

Using Noisy Bilingual Data for Statistical Machine Translation Stephan Vogel Interactive Systems Lab Language Technologies Institute Carnegie Mellon University vogel+@cs.cmu.edu Abstract SMT systems rely on sufficient amount of parallel corpora to train the trans- lation model. This paper investigates possibilities to use word-to-word and phrase-to-phrase translations extracted not only from clean parallel corpora but also from noisy comparable corpora. Translation results for a Chinese to En- glish translation task are given. 1 Introduction Statistical machine translation systems typically use a translation model trained on bilingual data and a language model for the target language, trained on perhaps some larger monolingual data. Often the amount of clean parallel data is limited. This leads to the question of whether translation quality can be improved by using additional nois- ier bilingual data. Some approaches, like (Fung and MxKeown, 1997), have been developed to extract word trans- lations from non-parallel corpora. In (Munteanu and Marcu, 2002) bilingual suffix trees are used to extract parallel sequences of words from a com- parable corpus. 95% of those phrase translation pairs were judged to be correct. However, no re- sults where reported if these additional translation correspondences resulted in improved translation quality. 2 The SMT System Statistical translation as introduced in (Brown et al., 1993) is based on word-to-word translations. The SMT system used in this study relies on multi- word to multi-word translations. The term phrase translations will be used throughout this paper without implying that these multi-word translation pairs are phrases in some linguistic sense. Phrase translations can be extracted from the Viterbi alignment of the alignment model. Phrase translation pairs are seen only a few times. Actually, most of the longer phrases are seen only once in even the larger corpora. Using relative frequency to estimate the translation prob- ability would make most of the phrase translation probabilities 1.0. This would lead to two conse- quences: First, phrase translation would always be preferred over a translation generated using word translations from the statistical and manual lexicons, even if the phrase translation is wrong, due to misalignment. Secondly, two translations would often have the same probability. As the language model probability is larger for shorter phrases this will usually result in overall shorter sentences, which sometimes are too short. To make phrase translations comparable to the word translations the translation probability is cal- culated on the basis of the word translation proba- bilities resulting from IBM 1-type alignment. n 1 P(fZIelk) = Ep(filei) i=m i=k This now gives the desired property that longer (1) 175 translations get higher probabilities. If the addi- tional word should not be part of the phrase trans- lation then these additional probabilities kb ei) which go into the sum will be small, i.e. the phrase translation probabilities will be very similar and the language model gives a bias toward the shorter translation. If, however, this additional word is ac- tually the translation of one of the words in the source phrase then the additional probabilities go- ing into the summation …

Content truncated to 500 words