Representations of discourse

Abstract: "This work is concerned with empirical studies of cognitive and computational aspects of discourse representations. A more specific aim is to contribute to the development of natural language interfaces for interaction with computers, especially the development of representations making possible a continuous interactive dialogue between user and system. General issues concerning the relationship between human cognitive and computational aspects of discourse representations are studied through an empirical and theoretical analysis of a psychological theory of discourse coherence, the theory of mental models. The analysis suggests that there are principled limits to what workers in computational linguistics can learn from psychological work on discourse processing. As far as the theory of mental models as a psychological theory of discourse is concerned, the effect of previous background knowledge of the domain of discourse on the processing of the types of texts often used in previous work is demonstrated. It is argued that this demonstration does not invalidate any of the basic assumptions of the theory, but should rather be seen as a modification or clarification. An attempt is also made to study the possible existence of different cognitive strategies used by different subjects and in different tasks. While some supporting evidence for this can be seen, it is argued that the results obtained are not conclusive on this issue. Another set of studies use the so-called Wizard-of-Oz method, i.e. dialogues with simulated natural language interfaces. Here the focus of the analysis is on the dialogue structure, and on the use of referring and co-referring expressions in the dialogues. The basic result of the dialogue analysis is that it is possible to describe these kinds of dialogues using a dialogue grammar, the LINDA-model, the basic feature of which is the partitioning of dialogues in a number of initiative-response (IR) units. The study of referring expressions also shows a lack of some of the complexities encountered in human dialogues. The results point to the possibility of using computationally simpler methods than what has hitherto been assumed, both for the dialogue management and for the resolution of anaphoric references."