Lexical Semantics and Knowledge Representation in Multi l ingual Text Generation

In contrast, a state-of-the-art natural language generation (NLG) system would likely be able to produce only one of them, modulo variations introduced by, for example, passivization, topicalization, or pronominalization. The problem is not simply to choose among paraphrases, but to be able to produce them to start with. In this book, an extended version of his dissertation, Stede provides the theoretical and practical means of endowing a text generator with such capability. He concentrates on tactical planning, namely, on choosing an appropriate verbalization for some content that another component of the NLG system (the text planner) has assembled in response to communicative goals. Stede's approach promotes lexicalization as the central step in this process and the lexicon itself as the link between the languageindependent domain model and the language-specific resources necessary to generate the surface sentence. Stede's ultimate goal is ambitious: he seeks to provide an architecture that can be used as is to perform generation of the same content in different languages--even lexical entries are to be reused when possible. In his view, multilingual generation, including the problem of language divergences, can be seen as a mere extension of the monolingual paraphrase task. Whereas the methods proposed are very compelling with respect to monolingual generation, it is not clear whether the multilingual goal has been fully achieved. I will come back to this point at the end of this review. The book is structured in three main parts: