Hierarchical Sub-sentential Alignment with Anymalign

Hierarchical Sub-sentential Alignment with Anymalign Adrien Lardilleux LIMSI-CNRS Orsay, France adrien.lardilleux@limsi.fr Franc¸ois Yvon LIMSI-CNRS/University Paris Sud Orsay, France francois.yvon@limsi.fr Yves Lepage Waseda University, IPS Waseda, Japan yves.lepage@waseda.jp Abstract We present a sub-sentential alignment al- gorithm that relies on association scores between words or phrases. This algorithm is inspired by previous work on alignment by recursive binary segmentation and on document clustering. We evaluate the re- sulting alignments on machine translation tasks and show that we can obtain state-of- the-art results, with gains up to more than 4 BLEU points compared to previous work, with a method that is simple, independent of the size of the corpus to be aligned, and directly computes symmetric alignments. This work also provides new insights re- garding the use of “heuristic” alignment scores in statistical machine translation. 1 Introduction Sub-sentential alignment consists in identifying translation units in sentence-aligned parallel cor- pora, i.e. in texts in which each sentence has been matched with its translation. This task constitutes the first step in the process of training most data- driven machine translation (MT) systems (statistical or example-based). The most prominent approach nowadays is phrase-based statistical machine trans- lation (SMT), where the core model is a translation table derived from sub-sentential mappings. This ta- ble consists in a pre-computed list of phrase1 pairs, where each (source, target) pair is associated with a certain number of scores loosely reflecting the likelihood that source translates to target. The problem of identifying sub-sentential map- pings from parallel texts, e.g. between isolated words or n-grams of words, is well-known, and nu- merous proposals have been put forward to perform this task. Those methods roughly fall into two main c⃝ 2012 European Association for Machine Translation. 1In this context, a phrase is a sequence of words and does not necessarily correspond to a syntactic phrase. categories. On the one hand, the probabilistic ap- proach, introduced by Brown et al. (1988), consid- ers the problem of identifying links between words or groups of words in parallel sentences. This ap- proach consists in defining a probabilistic model of the parallel corpus, the parameters of which are es- timated by a global maximization process which si- multaneously considers all possible associations in the corpus. The goal is to determine the best set of alignment links between all source and target words of every parallel sentence pair. The most famous representatives in this category are the IBM models (Brown et al., 1993) for aligning isolated words, which have given rise to an impressive series of variants and amendments (see e.g. (Vogel et al., 1996; Wu, 1997; Deng and Byrne, 2005; Liang et al., 2006; Fraser and Marcu, 2007; Ganchev et al., 2008), to cite a few). Generalizing word align- ment models to phrase alignment proves to be a much more difficult problem, and in the view of work of Marcu and Wong (2002) and Vogel (2005), such alignments are generally produced by heuristi- cally combining asymmetric 1–n word alignments (“oriented”) in both directions (Koehn et al., 2003; …

Content truncated to 500 words