Rapid assembly of a large-scale French-English MT system

Rapid assembly of a large-scale French-English MT system Jessie Pinkham§, Monica Corston-Oliver°, Martine Smets§, and Martine Pettenaro§ §Microsoft Research One Microsoft Way Redmond, WA 98052 {jessiep, martines, martinep}@microsoft.com °Butler Hill Group http://www.butlerhill.com moco@butlerhill.com Abstract Past research has shown that the ideal MT system should be modular and devoid of language pair specific information in its design. We describe here the assembly of TAMTAM (Traduction Automatique Microsoft), the French-English research MT system under development at Microsoft, which was constructed from a combination of pre-existing rule-based components and automatically created components. At this stage, the system has not been adapted either computationally or linguistically to the French-English context and yet it performs only slightly below the French-English Systran system in independent blind human evaluations Keywords Hybrid MT, EBMT, automatic transfer learning, French-English Introduction TAMTAM (Traduction Automatique Microsoft) is a French-English translation system which uses a French broad coverage analyzer, a large multi-purpose French dictionary, a French-English bilingual lexicon, an application-independent English natural language generation component and a transfer component. The transfer component consists of high-quality transfer patterns automatically acquired from sentence-aligned bilingual corpora using an alignment grammar and algorithm described in detail in Menezes (2001) (see Figure 1). The system is best characterized as a data-driven hybrid system, with rule-based analysis and generation, example- based transfer, and some statistically derived lexical input. The automatic alignment procedure used to create the example base relies on the same parser employed during analysis and also makes use of its own small set of rules for determining permissible alignments. A moderately sized French-English dictionary, containing only word pairs and their parts of speech, provides translation candidates for the alignment procedure and is also used as a backup source of translations during transfer. Statistical techniques supply additional translation pair candidates for alignment and identify certain multi-word terms for parsing and transfer. A complete description of the architecture and technical details appears in Richardson et al. (2001). We will focus here on describing the assembly and early stages of development of the French-English MT system, from its inception in October 2000 to now (April 2001). We report on evaluations conducted so far (in November, February and March), which show significant progress without any hand-coded linguistic changes specific to this language pair. Review of previous work Frank (1999) describes an effort to leverage work on parallel grammar development toward Machine Translation. The ParGram project involves parallel development in four languages (English, French, German, Norwegian). The effort stresses the importance of LFG representations, particularly f-structures, which encode functional properties shared across multiple languages, in the construction of an MT prototype covering fewer than 100 sentences. Though we do not rely on the LFG theory as a formalism, we have in our Logical Form component the same advantage of functional/semantic properties shared across several languages. The transfer rules written for the prototype described by Frank are extremely complex, demonstrating that hand- coded transfer rules would be unsuitable for large-scale MT system development. Frank states that automatic efforts for acquisition of transfer …

Content truncated to 500 words