High-Accuracy Phrase Translation Acquisition Through Battle-Royale Selection

Proceedings of Recent Advances in Natural Language Processing, pages 516–524, Hissar, Bulgaria, 7-13 September 2013. High-Accuracy Phrase Translation Acquisition Through Battle-Royale Selection Lionel Nicolas Egon W. Stemle Klara Kranebitter Verena Lyding Institute for Specialised Communication and Multilingualism, European Academy of Bozen/Bolzano {lionel.nicolas,egon.stemle,klara.kranebitter,verena.lyding}@eurac.edu Abstract In this paper, we report on an unsupervised greedy-style process for acquiring phrase translations from sentence-aligned paral- lel corpora. Thanks to innovative selection strategies, this process can acquire multi- ple translations without size criteria, i.e. phrases can have several translations, can be of any size, and their size is not con- sidered when selecting their translations. Even though the process is in an early development stage and has much room for improvements, evaluation shows that it yields phrase translations of high pre- cision that are relevant to machine trans- lation but also to a wider set of applica- tions including memory-based translation or multi-word acquisition. 1 Introduction This paper reports on work in progress to acquire contiguous phrase translations from sentence- aligned parallel corpora in an unsupervised way. The described process has three key features: it allows to acquire multiple translations for each phrase, the acquired translations can comprise phrases of any length,1 and it does not rely on any relation between the sizes of the phrases (no fertility criteria). In addition, its performance, es- pecially its precision, allows for competition with the state-of-the-art. Furthermore, the acquired phrase translations can be used for performing ma- chine translation, and memory-based translation; phrase/word alignment; multi-word, paraphrase, and synonymy acquisition; and error correction. The process starts by generating an exhaustive set of candidate translations and coarsely filters them. It then provides the remaining set to a 1We only use a loose maximum length restriction in order to limit exponential computation greedy fine-grained selection that processes one candidate translation at each iteration. The itera- tion stops when no candidate translations remain. The main contributions of this paper are (1) to introduce a set of filters for the coarse filter- ing of candidate translations, and (2) to describe a greedy-style process for performing a fine-grained selection of translations. In section 2 and 6, we describe the state-of- the-art and, in section 3 and 4, the process itself. We then present its results in section 5, compare it with related work in section 6, highlight future works in section 7 and conclude in section 8. 2 Related works The process described here can be considered in between two lines of approaches: bilingual lexi- con acquisition and phrase translations extraction from word alignments or translations. Methods performing bilingual lexicon acqui- sition focus on short phrases, mostly with one or two tokens. They generally use association measures to rank candidate translations and ap- ply several thresholds to decide which ones to keep (Gale and Church, 1991; Melamed, 1995; Wu and Xia, 1994). Most association measures used focus on recurrent occurrences, except meth- ods like Widdows et al. (2002) which apply mea- sures from semantic similarity approaches. Some approaches rely on either or both part-of-speech knowledge …

Content truncated to 500 words