Defining the Structure of Discourse Segments for Interpretation of Demonstratives

Kazuhiro Mogi

Tohoku University

In this paper I will present a formalism that specifies the range of antecedents of demonstratives in Japanese, focusing on the interpretation of the structure of discourse segments influencing identification of demonstratives.

Grosz and Sidner (1986) proposed a model for capturing discourse structures. Grosz and Sidner's theory (GST hereafter) uses the hierarchical relation of discourse segment purposes to identify discourse segments, but this approach has no clear definition of the relation between discourse segment purposes, and then fails to capture the structural relation between discourse segments. In order to overcome this difficulty, I employed Rhetorical Structure Theory (RST hereafter) proposed by Mann and Thompson (1988) because it allows us to describe the structure precisely, and modified it with a focus?stack formalism. The focus?stuck is used to narrow down the range for demonstratives to refer, incrementally updating the information associated with discourse segments. The referring domain is stacked up with the push action and the pop action removes the domain from the stack.

In RST, the relations between parts of discourse are described on the basis of the rhetorical relations, while the internal structures of discourse are described by the nucleus schema, which represents more informative elements of discourse, and by the satellite schema, which supports the nucleus.

First of all, I defined the segment to introduce the focus?stack formalism to RST. Secondly, I proposed the focus?stack actions based on RST. The actions consist of the immediate pop, the delay pop and the satellite pop. The immediate pop pops the referring domain from the stack at the end of a discourse segment. The delay pop works as the immediate pop does except that the pop action is delayed. The satellite pop pops only satellite elements that the segment has at the end of a discourse segment.

I examined whether the antecedents of demonstratives can refer to by these three pop actions. The result is that the best referred action is the satellite pop, the second is the delay pop, and the third the immediate pop action. As we can see this result, the satellite pop is the most effective action when we take into account the remained information on the stack. According to these results, we can conclude that the informative elements in discourse segments are best referred to by demonstratives.