Towards the automatic acquisition of lexical selection rules

MT Summit VII Sept.1999 Towards the Automatic Acquisition of Lexical Selection Rules Jin Yang SYSTRAN Software. Inc. 7855 Fay Avenue, Suite 300 La Jolla, CA 92037, USA jyang@systransoft.com Abstract This paper is a study of a certain type of collocations and implication and applica- tion to acquisition of lexical selection rules in transfer-approach MT systems. Collo- cations reveal the co-occurrence possibil- ities of linguistic units in one language, which often require lexical selection rules to enhance the natural flow and clarity of MT output. The study presents an auto- matic acquisition and human verification process to acquire collocations and sug- gest possible candidates for lexical selec- tion rules. The mechanism has been used in the development and enhancement of the Chinese-English and Japanese-English MT systems, and can be easily adapted to other language pairs. Future work in- cludes expanding its usage to more lan- guage pairs and furthering its application to MT customers. 1 Introduction In a transfer-based machine translation (MT) system, lexical choice is realized in the transfer stage by struc- tural transfer rules, lexical transfer rules and bilin- gual 1exicons. While the structural transfer accounts for the transformation of syntactic structures between two languages, the lexical transfer is responsible for selecting a proper translation based on various inter- woven constraints (e.g., syntax, semantics, domain, and pragmatics). Among various types of lexical se- lection rules, representation and treatment of colloca- tions is one of them. Since collocation-an arbitrary and recurrent word combination [l]-is one of the im- portant linguistic phenomena, the representation and treatment of collocations is indispensable for natural MT output [15]. Many theoretical and applied studies relating to collocations have been conducted from different points of view. One representative work from the computa- tional linguistics point of view is Smadja's study on collocations [9, 10, 11, 12, 13], which not only reviewed the theoretical and applied studies related to collo- cations, but also explored the relevance to computa- tional linguistics. Of the three types of collocations1 categorized by Smadja [11]. the type of “predicative relations” is the point of interest of this paper. “A predicative relation consists of two words repeatedly used together in a similar syntactic relation. These lexical relations are the most flexible types of collo- cations. They are hard to identify since they often correspond to interrupted word sequences in the cor- pus” ([11], pp.148). One example occurs when a noun and a verb are used together in a verb-object pattern to form a predicative relation (e.g.. "raise-question'', "reach-agreement"). Similarly, there are other pat- terns such as adjective-noun "top-priority'', subject- verb "temperature-rise", and verb-complement "blow- out" etc. For MT, collocations in one language usu- ally require special translation treatment via lexical selection rules, instead of using the default transla- tion coded in the bilingual lexicon. Additionally, the collocations of predicative relations are usually com- positionally represented and treated. In a transfer- based MT system, they are represented by context- dependent lexical selection rules. Table 1 illustrates some examples for the translation of verb "raise" …

Content truncated to 500 words