Uniform multilingual sentence generation using flexible lexico-grammatical resources

It is desirable for a sentence generation architecture to utilize diverse lexical and grammatical forms of expression, especially for multilingual generation and paraphrasing. This work presents a generation architecture capable of generating such variety in a uniform manner. Central to our approach are lexico-grammatical resources which pair elementary semantic structures with their syntactic realization and all lexico-grammatical consequences. Since the resources, contained in the lexicon for the given language, encapsulate all information necessary to produce the realizations of a semantic input, the generation methodology itself is simple and free from exceptional processing. Consistent with the traditional approach to sentence generation, our methodology consists of three parts: the decomposition of the semantic input into non-overlapping pieces to be realized by separate lexico-grammatical resources, the independent realization of the semantic pieces, and putting the realizations together to form a grammatical sentence in the language. The independence of the realization of semantic pieces is an important aspect of the simplicity of the methodology. There exist linguistic phenomena that raise issues with respect to this independence. They include constraints a lexical item might impose on the type of arguments (traditionally called selectional restrictions) and constraints a lexical item might impose on other lexical items with which it co-occurs (resulting in so-called collocations). We investigate the role of these phenomena in generation in a way that has not been done before. We also conclude that they do not pose complications to the independent realization of semantic pieces or to our generation architecture in general.