Hybridity in MT. Experiments on the Europarl Corpus

Hybridity in MT: Experiments on the Europarl Corpus Declan Groves∗ and Andy Way National Centre for Language Technology School of Computing Dublin City University Dublin 9, Ireland {dgroves,away}@computing.dcu.ie Abstract (Way & Gough, 2005) demonstrate that their Marker-based EBMT system is ca- pable of outperforming a word-based SMT system trained on reasonably large data sets. (Groves & Way, 2005) take this a stage further and demonstrate that while the EBMT system also outperforms a phrase-based SMT (PBSMT) system, a hybrid ‘example-based SMT’ system incorporating marker chunks and SMT sub-sentential alignments is capable of outperforming both baseline translation models for French–English translation. In this paper, we show that similar gains are to be had from constructing a hy- brid ‘statistical EBMT’ system capable of outperforming the baseline system of (Way & Gough, 2005). Using the Eu- roparl (Koehn, 2005) training and test sets we show that this time around, al- though all ‘hybrid’ variants of the EBMT system fall short of the quality achieved by the baseline PBSMT system, merging elements of the marker-based and SMT data, as in (Groves & Way, 2005), to cre- ate a hybrid ‘example-based SMT’ sys- tem, outperforms the baseline SMT and EBMT systems from which it is derived. Furthermore, we provide further evidence in favour of hybrid systems by adding an SMT target language model to all EBMT system variants and demonstrate that this too has a positive effect on transla- tion quality. 1 Introduction Almost all research in MT being carried out today is corpus-based. Within this field, by far the most dominant paradigm is PB- ∗This work is supported by an IRCSET Ph.D. Fellowship Award. SMT, but much important work continues to be carried out in EBMT. Until the re- cent work of (Way & Gough, 2005), no com- parative studies of any flavour of SMT and EBMT had appeared in print. While this work demonstrated that the Marker-based EBMT system of (Way & Gough, 2005) was capable of outperforming a word-based SMT system, (Groves & Way, 2005) showed that the EBMT system was also capable of higher translation quality than a PB- SMT system constructed from freely avail- able resources. However, perhaps more im- portantly for the MT research community as a whole, this paper also demonstrated that a novel hybrid ‘example-based SMT’ system incorporating marker chunks and SMT sub- sentential alignments was capable of outper- forming both baseline translation models, for French–English. In this paper, we continue that line of research by developing a new hybrid ‘sta- tistical EBMT’ system which outperforms the equivalent baseline system and a hybrid ‘example-based’ SMT system which outper- forms the baseline EBMT and SMT sys- tems from which it is derived. Crucially, therefore, the most important message aris- ing from our research is that MT devel- opers need to combine aspects from both SMT and EBMT if further gains are to be made; that is, despite the obvious conver- gence of the two paradigms, the remaining differences between SMT and EBMT (Way & Gough, 2005) are …

Content truncated to 500 words