Introducing a multi-dimensional user model to tailor natural language generation

Abstract

Previous work has shown that it is important to include user modelling in question answering systems in order to tailor the output. In this thesis, we develop a natural language response generation model that handles both definitional and procedural questions, and employs a multi-dimensional user model. The various attributes of the user recorded in the user model include the user's role, domain knowledge, preferences, interests, receptivity and memory capability. We also address how to represent the user's view of a knowledge base to then tailor generation to that user's view, and introduce a representation for knowledge bases with property inheritance based on the Telos system.

We further specify how this multi-dimensional user model influences different stages of McKeown style generation, on determining question type, deciding relevant knowledge, selecting schema, instantiating predicates, and selecting predicates. An algorithm is proposed to describe how the generation process can be tailored according to these influences, and some examples are presented to show that the output is desirable.

We also address the problem of conflicts arising from the variation in response generation suggested by different user model attributes and use various weighting schemes to resolve them. Furthermore, we include a procedure for updating the user model after each interaction which also enables the algorithm to be used over multiple sessions, with up-to-date information about the users.