The generation gap : The problem of expressibility in text planning

This thesis identifies and provides a solution for a particular problem in natural language generation: the problem of ensuring the expressibility of a text plan. Natural language generation is the process of going from a representation of a situation to a textual expression of some relevant portion of that situation in a natural language. Generation systems must have a principled way of ensuring that the message composed by the text planner is expressible in language, that is, that there are linguistic resources (words, syntactic structures) available for the linguistic component to realize the elements of the plan, and their composition is in accordance with the rules of composition in the language. I have addressed the problem of expressibility by designing a level of representation, the Text Structure, which is used by the text planner in composing the utterance. This intermediate level of representation bridges the generation gap between the representation of the world in the application program and the linguistic resources provided by the language. The terms and expressions in the Text Structure are abstractions over the concrete resources of language (the words, morphological markings, syntactic structures, etc. that actually appear in a stream of text). These abstract linguistic resources group together the expressible combinations of concrete linguistic resources. I have identified three kinds of information that are essential to an abstract linguistic representation: the constituency, the semantic category of the constituent (e.g. event, property), and the structural relations among the constituents (e.g. argument, adjunct). By providing the planner with a set of abstract resources, rather than letting it choose from the individual features that make them up, it is prevented from choosing a set of features that is not realizable. These abstractions can further constrain composition by defining what kinds of constituents can be extended and how semantic categories can compose. Text Structure is implemented in the Spokesman Generation System, which produces text for a variety of application programs. I describe in detail the structures in Spokesman\u27s text planner and walk through an example of the generation of a biographical paragraph from the Main Street simulation program