Lexical options in multilingual generation from a knowledge base

Abstract

From the viewpoint of multilingual generation, the common underlying knowledge base should be kept clear of language-specific concepts. This goal presupposes that lexical items of various languages cannot map one-to-one onto concepts all the time. We propose a more flexible way of attaching lexical items to configurations of concepts and roles, and a lexical option finder that determines the set of content words that cover pieces of the message to be expressed, thereby performing the first half of the “chunking” task (dividing the message into separately verbalizable parts). This pool of lexical options will also include synonyms and near-synonyms: items with identical denotation, that is semantic representation in the KB, but different connotational characteristics. From this set, the subsequent steps of the generation process can select the most preferred subset for expressing the message.