Advances in Machine Translation Research in IBM

[MT Summit III, Washington, DC, USA] Advances in Machine Translation Research in IBM Mori Rimon* Michael C. McCord Ulrike Schwall Pilar Martínez IBM Scientific Center IBM T.J. Watson IBM Scientific Center IBM Scientific Center Haifa, Israel Research Center Heidelberg, Germany Madrid, Spain Hawthorne, NY, USA Abstract IBM is engaged in advanced research and devel- opment projects on various aspects of machine translation, between several language pairs. The activities reported on hero are all parts of a rather large-scale, international effort, fol- lowing Michael McCord’s LMT approach. The paper focuses on seven selected topics: recent enhancements made in the Slot Grammar for- malism and the specific analysis components; specification of a semantic type hierarchy and its use for verb sense disambiguation; incorpo- ration of statistical techniques in the transla- tion process; anaphora resolution; linkage of target morphology modules; methods for the construction of large MT lexicons; and interac- tive disambiguation. 1 Introduction This paper reports on a sel of advanced research and de- velopment projects, coordinated under one international framework, being carried out in various IBM centers in Europe and the USA. The projects tackle various as- pects of machine translation, and the achievements are demonstrated by working prototypes. The presentation of this work in MT Summit III will include examples from our three major prototypes at this time: English- German, German-English, and English-Spanish, (Work is also underway on five other language pairs.) The underlying technical approach follows that of the LMT system by Michael McCord. In the basic LMT design, translation is accomplished through four ma- jor steps (all implemented in Prolog): lexical analysis, producing clauses describing inflected and derived input *The authors represent larger groups in the IBM centers in the United States, Germany, Spain and Israel, collaborating on this effort. Credit for contributions is due to G. Arrarte, E. Bentur, A. Bernth, B. Bläser, B. Carranza, I. Dagan, S. Lappin, H. Leass, M. Neff, U. Rackow, T. Redondo, D. Segev (Ben-Ari), A. Storrer and I. Zapata. Specific references are made in the paper to major individual contributions. The mail address of the technical leader of the coordinated projects is: Mori Rimon, IBM Scientific Center, the Technion City, Haifa 32000, Israel. E-mail: rimon@haifasc3.vnet.ibm.com words and their transfers; source syntactic analysis, pro- ducing an internal representation showing both surface structure and deep (logical) relations; transfer, divid- ed into a compositional phase which produces a data structure isomorphic to the output of source analysis but having target (base) words and target features, and a re- structuring phase which produces a target surface tree; and finally target morphological generation, producing the target sentence (or phrase). The system includes a large shell with language-general functions. The overall LMT architecture is described in detail elsewhere [McCord 85, 89a,b]. Here we concentrate on a few selected topics, representing interesting recent en- hancements to the original LMT design. The one central guideline we follow is to maintain a reasonable balance between advanced research goals and practical possibilities. Thus, while we seek solid method- ological grounds and …

Content truncated to 500 words