Maximizing Component Quality in Bilingual Word-Aligned Segmentations

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 30–38, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Maximizing Component Quality in Bilingual Word-Aligned Segmentations Spyros Martzoukos Christophe Costa Florˆencio Christof Monz Intelligent Systems Lab Amsterdam, University of Amsterdam Science Park 904, 1098 XH Amsterdam, The Netherlands {S.Martzoukos, C.CostaFlorencio, C.Monz}@uva.nl Abstract Given a pair of source and target language sentences which are translations of each other with known word alignments between them, we extract bilingual phrase-level segmenta- tions of such a pair. This is done by identi- fying two appropriate measures that assess the quality of phrase segments, one on the mono- lingual level for both language sides, and one on the bilingual level. The monolingual mea- sure is based on the notion of partition refine- ments and the bilingual measure is based on structural properties of the graph that repre- sents phrase segments and word alignments. These two measures are incorporated in a ba- sic adaptation of the Cross-Entropy method for the purpose of extracting an N-best list of bilingual phrase-level segmentations. straight-forward application of such lists inA Statistical Machine Translation (SMT) yields a conservative phrase pair extraction method that reduces phrase-table sizes by 90% with insignificant loss in translation quality. 1 Introduction Given a pair of source and target language sen- tences which are translations of each other with known word alignments between them, the problem of extracting high quality bilingual phrase segmen- tations is defined as follows: Maximize the quality of phrase segments, i.e., groupings of consecutive words, in both language sides, subject to constraints imposed by the underlying word alignments. The purpose of this work is to provide a solution to this maximization problem and investigate the effect of the resulting high quality bilingual phrase segments on SMT. For brevity, ‘phrase-level sentence segmen- tation’ and ‘phrase segment’ will henceforth be sim- ply referred to as ‘segmentation’ and ‘segment’ re- spectively. The exact definition of segments’ quality depends on the application. Our notion of a segmentation of maximum quality is defined as the set of consecutive words of the sentence that captures maximum col- locational and/or grammatical characteristics. This implies that a sequence of tokens is identified as a segment if its fully compositional expressive power is higher than the expressive power of any combina- tion of partial compositions. Since this definition is fairly general it is thus suitable for most NLP tasks. In particular, it is tailored to the type of segments that are suitable for the purposes of SMT and is in line with previous work (Blackwood et al., 2008; Paul et al., 2010). With this definition in mind, we introduce a monolingual segment quality measure that is based on assessing the cost of converting one segmentation into another by means of an elementary operation. This operation, namely the ‘splitting’ of a segment into two segments, together with all possible seg- mentations of a sentence are known to form a par- tially ordered set (Guo, 1997). …

Content truncated to 500 words