Translation Combination using Factored Word Substitution

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 70–74, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Translation Combination using Factored Word Substitution Christian Federmann1, Silke Theison2, Andreas Eisele1,2, Hans Uszkoreit1,2, Yu Chen2, Michael Jellinghaus2, Sabine Hunsicker2 1: Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz GmbH, Saarbr¨ucken, Germany 2: Universit¨at des Saarlandes, Saarbr¨ucken, Germany {cfedermann,eisele,uszkoreit}@dfki.de, {sith,yuchen,micha,sabineh}@coli.uni-sb.de Abstract We present a word substitution approach to combine the output of different machine translation systems. Using part of speech information, candidate words are deter- mined among possible translation options, which in turn are estimated through a pre- computed word alignment. Automatic substitution is guided by several decision factors, including part of speech, local context, and language model probabili- ties. The combination of these factors is defined after careful manual analysis of their respective impact. The approach is tested for the language pair German- English, however the general technique it- self is language independent. 1 Introduction Despite remarkable progress in machine transla- tion (MT) in the last decade, automatic translation is still far away from satisfactory quality. Even the most advanced MT technology as summarized by (Lopez, 2008), including the best statistical, rule- based and example-based systems, produces out- put rife with errors. Those systems may employ different algorithms or vary in the linguistic re- sources they use which in turn leads to different characteristic errors. Besides continued research on improving MT techniques, one line of research is dedicated to bet- ter exploitation of existing methods for the com- bination of their respective advantages (Macherey and Och, 2007; Rosti et al., 2007a). Current approaches for system combination in- volve post-editing methods (Dugast et al., 2007; Theison, 2007), re-ranking strategies, or shal- low phrase substitution. The combination pro- cedure applied for this pape tries to optimize word-level translations within a ”trusted” sentence frame selected due to the high quality of its syntac- tic structure. The underlying idea of the approach is the improvement of a given (original) translation through the exploitation of additional translations of the same text. This can be seen as a simplified version of (Rosti et al., 2007b). Considering our submission from the shared translation task as the ”trusted” frame, we add translations from four additional MT systems that have been chosen based on their performance in terms of automatic evaluation metrics. In total, the combination system performs 1,691 substitutions, i.e., an average of 0.67 substitutions per sentence. 2 Architecture Our system combination approach computes a combined translation from a given set of machine translations. Below, we present a short overview by describing the different steps in the derivation of a combined translation. Compute POS tags for translations. We apply part-of-speech (POS) tagging to prepare the selection of possible substitution candidates. For the determination of POS tags we use the Stuttgart TreeTagger (Schmid, 1994). Create word alignment. The alignment between source text and translations is needed to identify translation options within the differ- ent systems’ translations. Word alignment is computed using the GIZA++ toolkit (Och and …

Content truncated to 500 words