Further Experiments with Shallow Hybrid MT Systems

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 77–81, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Further Experiments with Shallow Hybrid MT Systems Christian Federmann1, Andreas Eisele1, Hans Uszkoreit1,2, Yu Chen1, Sabine Hunsicker1, Jia Xu1 1: Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz GmbH, Saarbr¨ucken, Germany 2: Universit¨at des Saarlandes, Saarbr¨ucken, Germany {cfedermann,eisele,uszkoreit,yuchen,sabine.hunsicker,jia.xu}@dfki.de Abstract We describe our hybrid machine trans- lation system which has been developed for and used in the WMT10 shared task. We compute translations from a rule- based MT system and combine the re- sulting translation “templates” with par- tial phrases from a state-of-the-art phrase- based, statistical MT engine. Phrase sub- stitution is guided by several decision factors, a continuation of previous work within our group. For the shared task, we have computed translations for six lan- guage pairs including English, German, French and Spanish. Our experiments have shown that our shallow substitu- tion approach can effectively improve the translation result from the RBMT system; however it has also become clear that a deeper integration is needed to further im- prove translation quality. 1 Introduction In recent years the quality of machine translation (MT) output has improved greatly, although each paradigm suffers from its own particular kind of errors: statistical machine translation (SMT) of- ten shows poor syntax, while rule-based engines (RBMT) experience a lack in vocabulary. Hybrid systems try to avoid these typical errors by com- bining techniques from both paradigms in a most useful manner. In this paper we present the improved version of the hybrid system we developed last year’s shared task (Federmann et al., 2009). We take the out- put from an RBMT engine as basis for our hybrid translations and substitute noun phrases by trans- lations from an SMT engine. Even though a gen- eral increase in quality could be observed, our sys- tem introduced errors of its own during the substi- tution process. In an internal error analysis, these degradations were classified as follows: - the translation by the SMT engine is incorrect - the structure degrades through substitution (because of e.g. capitalization errors, double prepositions, etc.) - the phrase substitution goes astray (caused by alignment problems, etc.) Errors of the first class cannot be corrected, as we have no way of knowing when the translation by the SMT engine is incorrect. The other two classes could be eliminated, however, by introduc- ing additional steps for pre- and post-processing as well as improving the hybrid algorithm itself. Our current error analysis based on the results of this year’s shared task does not show these types of errors anymore. Additionally, we extended our coverage to also include the language pairs English↔French and English↔Spanish in both directions as well as English→German, compared to last year’s initial experiments for German→English only. We were able to achieve an increase in translation quality for this language set, which shows that the substi- tution method works for different language config- urations. 2 Architecture Our hybrid translation system takes translation output from …

Content truncated to 500 words