Bilingual phrase-to-phrase alignment for arbitrarily-small datasets

Bilingual phrase-to-phrase alignment for arbitrarily-small datasets Kevin Flanagan 667644@swansea.ac.uk Department of Languages, Translation and Communication, Swansea University, Swansea SA2 8PP, U.K. Abstract This paper presents a novel system for sub-sentential alignment of bilingual sentence pairs, however few, using readily-available machine-readable bilingual dictionaries. Performance is evaluated against an existing gold-standard parallel corpus where word alignments are annotated, showing results that are a considerable improvement on a comparable system and on GIZA++ performance for the same corpus. Since naïve application of the system for N languages would require N(N - 1) dictionaries, it is also evaluated using a pivot language, where only 2(N - 1) dictionaries would be required, with surprisingly similar performance. The system is proposed as an alternative to statistical methods, for use with very small cor- pora or for ‘on-the-fly’ alignment. 1. Introduction The process of extracting phrase pairs from parallel corpora is relevant in several contexts, in particular when inducing a translation model with a Statistical Machine Translation (SMT) system, such as that described by Koehn, Och et al. (2003). Phrase pairs are derived from alignments between parts of a bilingual sentence pair, based on word-alignment probabilities. A related procedure is alignment of parts of parse trees for a bilingual sentence pair, to create models for syntax-based machine translation (MT) (Tinsley, Zhechev et al. 2007, Lavie, Parlikar et al. 2008). What these tasks have in common is a requirement for initial word-level alignment information to be established, typically using a tool such as GIZA++ (Och and Ney 2003). In turn, this requires the corpus to be large to achieve useful results, and above a cer- tain minimum size to produce any results at all. A different context where alignment below sentence level can be of relevance is when re- calling content from Translation Memory (TM) systems, whether in order to propose transla- tions for sub-sentential matches (Simard 2003, Planas 2005) or to identify where to edit inex- act sentence-level matches (Kranias and Samiotou 2004, Esplà Gomis, Sánchez Martínez et al. 2011). While alignments of that kind can be achieved using statistical methods like those of GIZA++, the size requirements of those methods restrict their usefulness to cases where TMs are sufficiently large. Pre-training GIZA++ on a separate, large corpus does not give very good results, as shown by Esplà Gomis, Sánchez Martínez et al. (2012). The system presented here aligns parts of bilingual sentence pairs in isolation, without any training step, so has no minimum size requirement. It is intended to be of particular rele- vance to TM applications, but may also be used to generate sub-sentential alignments in other contexts where they are required and where the data available is limited. It produces hierar- chically-arranged pairs of word spans representing the alignments between sentence parts, including alignment of individual words where possible. Alignments between longer spans can be rendered as word alignments for evaluation or use in other contexts. The system is in Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC …

Content truncated to 500 words