Capturing Language Speci c Constraints on Lexical Selection with Feature-Based Lexicalized Tree-Adjoining Grammars

The success of a Machine Translation (MT) application depends on its ability to perform lexical selection , that is, to choose lexical items in the target language that most closely match the lexical items in the input source. This task is particularly dif-cult in cases, such as those which arise in translating from English to Chinese and Korean, where the target language imposes lexical constraints which are non-existent or completely diierent in the source. We present an implementation of an English-Korean MT system using Feature-Based, Lexicalized Tree-Adjoining Grammar (FB-LTAG), and demonstrate its ability to handle diiculties involving lexical selection between those two languages. We also describe the applicability of this approach to similar issues which arise in English-Chinese translation. By building language-dependent FB-LTAGs for each language and then linking them via a Synchronous Tree-Adjoining Grammar (STAG), we are able to elegantly model the speciic and language-dependent syntactic and semantic distinctions necessary to lter the choice of target lexical items.