The LDV-COMBO system for SMT

Proceedings of the Workshop on Statistical Machine Translation, pages 166–169, New York City, June 2006. c⃝2006 Association for Computational Linguistics The LDV-COMBO system for SMT Jes´us Gim´enez and Llu´ıs M`arquez TALP Research Center, LSI Department Universitat Polit`ecnica de Catalunya Jordi Girona Salgado 1–3, E-08034, Barcelona {jgimenez,lluism}@lsi.upc.edu Abstract We describe the LDV-COMBO system pre- sented at the Shared Task. Our approach explores the possibility of working with alignments at different levels of abstrac- tion using different degrees of linguis- tic analysis from the lexical to the shal- low syntactic level. Translation mod- els are built on top of combinations of these alignments. We present results for the Spanish-to-English and English-to- Spanish tasks. We show that liniguistic in- formation may be helpful, specially when the target language has a rich morphology. 1 Introduction The main motivation behind our work is to introduce linguistic information, other than lexical units, to the process of building word and phrase alignments. In the last years, many efforts have been devoted to this matter (Yamada and Knight, 2001; Gildea, 2003). Following our previous work (Gim´enez and M`arquez, 2005), we use shallow syntactic informa- tion to generate more precise alignments. Far from full syntactic complexity, we suggest going back to the simpler alignment methods first described by IBM (1993). Our approach exploits the possibil- ity of working with alignments at two different lev- els of granularity, lexical (words) and shallow pars- ing (chunks). Apart from redefining the scope of the alignment unit, we may use different linguistic data views. We enrich tokens with features further than lexical such as part-of-speech (PoS), lemma, and chunk IOB label. For instance, suppose the case illustrated in Fig- ure 1 where the lexical item ‘plays’ is seen acting as a verb and as a noun. Considering these two words, with the same lexical realization, as a single token adds noise to the word alignment process. Repre- senting this information, by means of linguistic data views, as ‘playsV BZ’ and ‘playsNNS’ would allow us to distinguish between the two cases. Ideally, one would wish to have still deeper information, moving through syntax onto semantics, such as word senses. Therefore, it would be possible to distinguish for instance between two realizations of ‘plays’ with different meanings: ‘heP RP playsV BG guitarNN’ and ‘heP RP playsV BG footballNN’. Of course, there is a natural trade-off between the use of linguistic data views and data sparsity. Fortunately, we hava data enough so that statistical parameter estimation re- mains reliable. The approach which is closest to ours is that by Schafer and Yarowsky (2003) who suggested a com- bination of models based on shallow syntactic anal- ysis (part-of-speech tagging and phrase chunking). They followed a backoff strategy in the application of their models. Decoding was based on Finite State Automata. Although no significant improvement in MT quality was reported, results were promising taking into account the short time spent in the de- velopment of the linguistic tools utilized. Our system is further described in Section 2. Re- …

Content truncated to 500 words