Automatic domain recognition for machine translation

MT Summit VII Sept. 1999 Automatic Domain Recognition for Machine Translation Elke D. Lange & Jin Yang SYSTRAN Software. Inc. 7855 Fay Avenue. Suite 300 La Jolla, CA 92037, USA {elange, jyang}@systransoft.com Abstract This paper describes an ongoing project which has the goal of improving machine translation quality by increasing knowledge about the text to be translated. A ba- sic piece of such knowledge is the domain or subject field of the text. When this is known, it is possible to improve mean- ing selection appropriate to that domain. Our current effort consists in automating both recognition of the text’s domain and the assignment of domain-specific transla- tions. Results of our implementation show that the approach of using terminology cat- egorization already existing in the machine translation system is very promising. 1 Introduction Along with the increasing growth of the World Wide Web, on-line translation has become an area of success for Machine Translation (MT) and also presents new challenges. It is agreed that, for a translation service catering to a wide audience, enhancement of robust- ness including automatic identification of language, domain and style level is required [2, 5, 9, 10]. The au- tomatic setting of translation parameters has become one of our development priorities, thus increasing ease- of-use and translation quality at the same time. One area of great importance is the automatic selection of target language meanings depending on the domain of the text. The automatic recognition of a text’s domain is similar to text categorization, which can be defined as the context-based assignment of one or more pre- defined categories to texts [6]. Interest in automatic categorization of texts usually is related to the fields of information retrieval and machine learning. Papers on the subject of automatic domain recognition for MT were seldom found until recently when the problem became evident for the on-line translation environ- ment. Another closely related research area is word sense disambiguation, which must be an MT devel- oper’s concern in providing high-quality translation. “This becomes very important when one is trying to further improve the quality of commercial MT systems which already have a store of knowledge required to produce fairly satisfactory translation'” [10]. We found two related works, which shared our motivation and goals of improving translation quality by recognizing domains. The “domain recognizer” in a Web-based English-Korean system [1] uses a text categorization technique, identifying 25 domains, and demonstrated a 45% accuracy for top domain and a 75% accuracy when the second top domain is applied. Another study [10] showed 12.0% improvement on the translation quality of Japanese-English MT. The present paper describes an ongoing project that has the dual goal of automatically identifying the domain of a text and of improving MT quality by automatically linking domain identification to mean- ing selection. The goal of the project-Automatic Do- main Recognition-is to increase translation accuracy by providing an automatic means of assigning domain- specific meanings to the translation of individual words. In order to do this, the system is …

Content truncated to 500 words