The MetaMorpho Translation System

Proceedings of the Third Workshop on Statistical Machine Translation, pages 111–114, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics The MetaMorpho translation system Attila Novák, László Tihanyi and Gábor Prószéky MorphoLogic Orbánhegyi út 5, Budapest 1126, Hungary {novak,tihanyi,proszeky}@morphologic.hu Abstract In this article, we present MetaMorpho, a rule based machine translation system that was used to create MorphoLogic’s submission to the WMT08 shared Hungarian to English translation task. The architecture of Meta- Morpho does not fit easily into traditional categories of rule based systems: the building blocks of its grammar are pairs of rules that describe source and target language structures in a parallel fashion and translated structures are created while parsing the input. 1 Introduction Three rule-based approaches to MT are tradition- ally distinguished: direct, interlingua and transfer. The direct method uses a primitive one-stage proc- ess in which words in the source language are re- placed with words in the target language and then some rearrangement is done. The main idea behind the interlingua method is that the analysis of any source language should result in a language- independent representation. The target language is then generated from that language-neutral repre- sentation. The transfer method first parses the sen- tence of the source language. It then applies rules that map the lexical and grammatical segments of the source sentence to a representation in the target language. The MetaMorpho machine translation system de- veloped at MorphoLogic (Prószéky and Tihanyi, 2002), cannot be directly classified in either of the above categories, although it has the most in com- mon with the transfer type architecture. 2 Translation via immediate transfer In the MetaMorpho system, both productive rules of grammar and lexical entries are stored in the form of patterns, which are like context-free rules enriched with features. Patterns may contain more-or-less underspecified slots, ranging from general productive rules of grammar through more- or-less idiomatic phrases to fully lexicalized items. The majority of the patterns (a couple of hundreds of thousands in the case of our English grammar) represent partially lexicalized items. The grammar operates with pairs of patterns that consist of one source pattern used during bot- tom-up parsing and one or more target patterns that are applied during top-down generation of the translation. While traditional transfer and interlin- gua based systems consist of separate parsing and generating rules, in a MetaMorpho grammar, each parsing rule has its associated generating counter- part. The translation of the parsed structures is al- ready determined during parsing the source language input. The actual generation of the target language representations does not involve any ad- ditional transfer operations: target language struc- tures corresponding to substructures of the source language parse tree are combined and the leaves of the resulting tree are interpreted by a morphologi- cal generator. We call this solution “immediate transfer” as it uses no separate transfer steps or target transformations. The idea behind this architecture has much in common with the way semantic compositionality was formalized by Bach (1976) in the from …

Content truncated to 500 words