Robust Parsing of Utterances i

The rapidly increasing number of spoken-dialogue systems have led to numerous robust parsers for limited domains having been constructed during the last decade. By “parsing” we here mean a mapping from the input utterance to a context-independent semantic representation. By “robust” we mean that the parser will give a reasonable result even on very noisy input. In simple spoken-dialogue applications, parsing can be interleaved with speech recognition (if the language model is grammar-based). The vast majority of launched commercial systems fall into this category. However, spoken-dialogue systems aiming at less system control and more user initiative usually have a statistical n-gram recogniser, due to the difficulties of constructing a grammar with sufficient coverage. In this case (which is what we consider in this article), a separate parser is needed to transform the output from the recogniser into a semantic representation. Generally speaking, these robust parsers have been highly successful. Distinguishing features are: very fast execution, short development times (down to a few person weeks), and performance as good as or better than state-of-the-art parsers based on large-coverage grammars. Such robust parsers are usually based on phrase-spotting, and output slot–filler structures representing the propositional contents of the input utterance. Thus, precision is traded for robustness, execution speed and ease of development. However, propositional slot–filler structures are not sufficient in cases where more user initiative is involved, such as negotiative dialogue (for example, travel planning, appointment scheduling, or apartment browsing). In this paper, starting from a domain model in the form of a relational database, we derive a richer semantic representation formalism suitable for negot evalu recog algor corre fragm et al. doma system