Translation of Unseen Bigrams by Analogy Using an SVM Classifier

Translation of Unseen Bigrams by Analogy Using an SVM Classifier Hao Wang Lu Lyu Yves Lepage Graduate School of Information, Production and Systems Waseda University 2-7 Hibikino, Wakamatsu-ku, Kitakyushu, Fukuoka 808-0135, Japan {oko ips@ruri., lulv90@ruri., yves.lepage@}waseda.jp Abstract Detecting language divergences and predict- ing possible sub-translations is one of the most essential issues in machine translation. Since the existence of translation divergences, it is impractical to straightforward translate from source sentence into target sentence while keeping the high degree of accuracy and with- out additional information. In this paper, we investigate the problem from an emerging and special point of view: bigrams and the cor- responding translations. We first profile cor- pora and explore the constituents of bigrams in the source language. Then we translate un- seen bigrams based on proportional analogy and filter the outputs using an Support Vector Machine (SVM) classifier. The experiment re- sults also show that even a small set of features from analogous can provide meaningful infor- mation in translating by analogy. 1 Introduction Over the last decade, phrase-based statistical ma- chine translation (Koehn et al., 2003) systems have demonstrated that they can produce reasonable qual- ity when ample training data is available, especially for language pairs with similar word order. How- ever, the PB-SMT model has not yet been capable of satisfying the various translation tasks for very dif- ferent languages (Isozaki et al., 2010). The existence of translation divergences makes the straightforward transfer from source sentences into target sentences hard. Though many previous pieces of work (Dorr, 1994; Habash et al., 2002; Dorr et al., 2004) have at- tempted to take account for divergences and to deal with this linguistic problem using various translation approaches. This paper further inquires the topic. Since sentence consists of bigrams, instead of analysing the syntactic structures of the whole sen- tence or part of the sentence as in (Ding and Palmer, 2005), we explore the possibilities of translating un- seen bigrams based on an analogy learning method. We investigate the coverage of translated bigrams in the test set and inspect the probability of translat- ing a bigram using analogy. Analogical learning has been investigated by several authors. To cite a few, Lepage et al. (2005) showed that proportional anal- ogy can capture some syntactic and lexical struc- tures across languages. Langlais et al. (2007) in- vestigated the more specific task of translating un- seen words. Bayoudh et al. (2007) explored generat- ing new learning examples from very scarce original learning data using analogy to train an SVM classi- fier. Dandapat et al. (2010) performed transliteration by analogical learning for English-to-Hindi. In the issue of translation using analogy, one of the main drawbacks should be addressed is the prob- lem of ”over-generative”. Analogy is able to cap- ture the most divergences of translation in the most cases, yet it generates a great number of solutions that are ungrammatical and incorrect. In this pa- per, we propose to translate useen bigrams as re- constructing with the principle of analogy learning. …

Content truncated to 500 words