Hierarchical Hybrid Translation between English and German

Hierarchical Hybrid Translation between English and German Yu Chen Andreas Eisele German Research Center for Artificial Intelligence (DFKI) GmbH Saarbr¨ucken, Germany {Yu.Chen, Andreas.Eisele}@dfki.de Abstract We present new results from a hybrid com- bination of rule-based machine translation (RBMT) with a variant of statistical ma- chine translation (SMT) that supports hier- archical structures and is therefore able to preserve more of the linguistic structures obtained from the RBMT system than ver- sions of SMT that operate on flat phrases alone. Having shown in (Chen and Eisele, 2010) for the first time that a tighter in- tegration of hierachical MT systems from different paradigms leads to consistent im- provements for translation from German to English in various experimental settings, the current paper generalizes the approach to translation from English to German, where we observe similar improvements. These findings indicate that hybrid com- binations of MT paradigms can benefit from structural similarities in the underly- ing models, which makes us expect even stronger benefits from a tight integration of different approaches. . 1 Introduction Research in machine translation has made sig- nificant progress in recent years. Statistical ma- chine translation (SMT) systems became espe- cially dominant in this area, motivated by the abil- ity to create new systems from existing training data without much effort, but also encouraged by the fact that many evaluation measures that solely rely on string matching allow to implement in- cremental improvements without having to solve c⃝ 2010 European Association for Machine Translation. the really hard issues. However, human assess- ments (Callison-Burch et al., 2009) show that rule- based systems can still translate better than SMT systems in many cases. The errors produced by different types of systems are somehow comple- mentary (Thurmair, 2004). In addition to improv- ing MT techniques for certain kinds of systems, another stream of research in MT aims at combin- ing existing methods, that is, to build hybrid ap- proaches. One way to integrate SMT and RBMT is to apply a variant of standard statistical methods to induce information from translations made by RBMT systems and to incorporate the information into the core of a SMT system (Eisele et al., 2008). Most RBMT systems benefit from the large ac- curate lexicons and complex grammars that took enormous human efforts through decades. By the hybrid combination, the linguistic information supplied in RBMT systems are expected to fill gaps in lexical knowledge of the SMT system, which is particularly lacking when translating texts in domains different from the training data. The hybrid framework in (Eisele et al., 2008) outperforms the original SMT system that acts as the core, however the improvements over the RBMT systems were not consistent when the RBMT systems actually created better translations compared to the SMT system. It is mostly because this hybrid combination method is unable to make use of well-formedness in RBMT, which, in fact, is one of the most significant advantages of sys- tems based on linguistic knowledge. On the con- trary, the correct syntactic structures are decom- …

Content truncated to 500 words