TectoMT: Highly Modular MT System with Tectogrammatics Used as Transfer Layer

Proceedings of the Third Workshop on Statistical Machine Translation, pages 167–170, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics TectoMT: Highly Modular MT System with Tectogrammatics Used as Transfer Layer∗ Zdenˇek ˇZabokrtsk´y, Jan Pt´aˇcek, Petr Pajas Institute of Formal and Applied Linguistics Charles University, Prague, Czech Republic {zabokrtsky,ptacek,pajas}@ufal.mff.cuni.cz Abstract We present a new English→Czech machine translation system combining linguistically motivated layers of language description (as defined in the Prague Dependency Treebank annotation scenario) with statistical NLP ap- proaches. 1 Introduction We describe a new MT system (called Tec- toMT) based on the conventional analysis-transfer- synthesis architecture. We use the layers of language description defined in the Prague Dependency Tree- bank 2.0 (PDT for short, (Hajiˇc and others, 2006)), namely (1) word layer – raw text, no linguistic annotation, (2) morphological layer – sequence of tagged and lemmatized tokens, (3) analytical layer – each sentence represented as a surface-syntactic dependency tree, and (4) tectogrammatical layer – each sentence represented as a deep-syntactic de- pendency tree in which only autosemantic words do have nodes of their own; prefixes w-, m-, a-, or t- will be used for denoting these layers.1 We use ‘Praguian’ tectogrammatics (introduced in (Sgall, 1967)) as the transfer layer because we believe that, first, it largely abstracts from language-specific (inflection, agglutination, func- tional words. . . ) means of expressing non-lexical ∗The research reported in this paper is financially supported by grants GAAV ˇCR 1ET101120503 and MSM0021620838. 1In addition, we use also p-layer (phrase structures) as an a-layer alternative, the only reason for which is that we do not have a working a-layer parser for English at this moment. meanings, second, it allows for a natural transfer factorization, and third, local tree contexts in t-trees carry more information (esp. for lexical choice) than local linear contexts in the original sentences. In order to facilitate separating the transfer of lex- icalization from the transfer of syntactization, we in- troduce the concept of formeme. Each t-node’s has a formeme attribute capturing which morphosyntac- tic form has been (in the case of analysis) or will be (synthesis) used for the t-node in the surface sen- tence shape. Here are some examples of formemes we use for English: n:subj (semantic noun (sn) in subject position), n:for+X (sn with preposition for), n:X+ago (sn with postposition ago), n:poss (posses- sive form of sn), v:because+fin (semantic verb (sv) as a subordinating finite clause introduced by be- cause), v:without+ger (sv as a gerund after without), adj:attr (semantic adjective (sa) in attributive posi- tion), adj:compl (sa in complement position). The presented system intensively uses the PDT technology (data formats, software tools). Special attention is paid to modularity: the translation is im- plemented (in Perl) as a long sequence of processing modules (called blocks) with relatively tiny, well- defined tasks, so that each module is independently testable, improvable, or substitutable. TectoMT al- lows to easily combine blocks based on different approaches, from blocks using complex probabilis- tic solutions (e.g., B2, B6, B35, see the next section), through …

Content truncated to 500 words