Czech-English Dependency Tree-based Machine Translation

Czech-English Dependency-based Machine Translation Martin 'emejrek, Jan Cufin, and MI Havelka Institute of Formal and Applied Linguistics, and Center for Computational Linguistics Charles University in Prague fcmejrek,curin,havelkal@ufal.mff.cuni.cz Abstract We present some preliminary results of a Czech-English translation system based on dependency trees. The fully auto- mated process includes: morphological tagging, analytical and tectogrammat- ical parsing of Czech, tectogrammati- cal transfer based on lexical substitu- tion using word-to-word translation dic- tionaries enhanced by the information from the English-Czech parallel corpus of WSJ, and a simple rule-based system for generation from English tectogram- matical representation. In the evalua- tion part, we compare results of the fully automated and the manually annotated processes of building the tectogrammat- ical representation.' 1 Introduction The experiment described in this paper is an at- tempt to develop a full MT system based on de- pendency trees (DBMT). Dependency trees repre- sent the sentence structure as concentrated around the verb and its valency. We use tectogrammatical dependency trees capturing the linguistic meaning of the sentence. In a tectogrammatical dependency tree, only autosemantic (lexical) words are repre- sented as nodes, dependencies (edges) are labeled 1This research was supported by the following grants: MSMT 'd2 Grant No. LNO0A063 and NSF Grant No. IIS- 0121285. by tectogrammatical functors denoting the seman- tic roles, the information conveyed by auxiliary words is stored in attributes of the nodes. For de- tails about the tectogrammatical representation see Haji6ova et al. (2000), an example of a tectogram- matical tree can be found in Figure 3. MAGENTA (Haji6 et al., 2002) is an exper- imental framework for machine translation im- plemented during 2002 NLP Workshop at CLSP, Johns Hopkins University in Baltimore. Modules for parsing of Czech, lexical transfer, a prototype of a statistical tree-to-tree transducer for structural transformations used during transfer and genera- tion, and a language model for English based on dependency syntax are integrated in one pipeline. For processing the Czech part of the data, we reuse some modules of the MAGENTA sys- tem, but instead of MAGENTA's statistical tree- to-tree transducing module and subsequent lan- guage model, we implement a rule-based method for generating English output directly from the tectogrammatical representation. First, we summarize resources available for the experiments (Section 2). Section 3 describes the automatic procedures used for the preparation of both training and testing data, including morpho- logical tagging, and analytical and tectogrammat- ical parsing of Czech input. In Section 4 we de- scribe the process of the filtering of dictionaries used in the transfer procedure (for its character- ization, see Section 5). The generation process consisting mainly of word reordering and lexical insertions is explained in Section 6, an example il- lustrating the generation steps is presented in Sec- 83 tion 7. For the evaluation of the results we use the BLEU score (Papineni et al., 2001). Section 8 compares translations generated from automati- cally built and manually annotated tectogrammat- ical representations. We also compare the results with the output generated by the statistical trans- lation system GIZA++/ISI ReWrite …

Content truncated to 500 words