DESCRIPTIONS 3 { rabbit

It has long been known that knowledge based reasoning is a crucial component of natural language processing (NLP). Yet the complexity involved in representing and using knowledge efficiently has led most NLP work to focus on more tractable aspects of language such as syntax, prosody or semantic construction. The generation of definite descriptions (that is, noun phrases with a definite article such as “the rabbit”) is a case in point. The goal of this sub-task of natural language generation (which is the production of a text satisfying a given communicative goal) is to construct a noun phrase that allows the hearer to uniquely identify its referent in the context of utterance. The standard algorithm for this task (on which most other proposals are based) is presented in (Dale and Reiter, 1995). But neither this algorithm nor the extensions proposed in (Horacek, 1997) and (Krahmer and Theune, 2001) take world knowledge into account when considering the context of utterance. For all these algorithms, the context is a fixed set of positive literals specifying entities and relations between them, which is intended to represent the current (linguistic and situational) context of utterance. Yet many definite descriptions either refer to inferable entities (entities not explicitly mentioned or present in the context of utterance but inferable from it) or refer to contextually salient entities using inferred rather than explicitly mentioned relations (Poesio and Vieira, 1998). For instance, the use of the definite article to refer to the patrons, the waitress, the busboys, etc. in (1) can only be explained by taking into account that world knowledge supports the assumption that these are somehow related to the restaurant.