Naoyoshi TAMURA
Department of Electrical and Computer Eng., Yokohama National University
Tokiwadai 156, Hodogaya-ku, Yokohama 240, Japan
e-mail: tam@tamlab.dnj.ynu.ac.jp
I. Purpose of the Research Our research subject in this year is to construct a model for rhetorical structure analysis on dialogue processing, which proceeds incrementally paying attention to the coherence preservation. There is no doubt that deep (in some case domain specific) knowledge and analysis are indispensable, it is no less important to understand the scheme or the structure of the argument in order not only to understand what is written but to generate responses properly so that meaning of each sentence is expanded on the structure. The subject includes several subproblems:
1) Incremental Rhetorical Structure Analysis
When a sentence is read into the system as the most primitive element of rhetorical structure, we think it is reasonable that the system will build the structure incrementally according to the sentence and the state of system. Moreover in such mechanism the system manages the intermediate structure on a stack just like bottom up parsers. 2) Extension of the Rhetorical Structure Theory
Original RST was proposed to represent the rhetorical structure on a written text. We try to apply the theory to spoken dialogue to confirm the theory is applicable to dialogue. 3) Coherency on Dialogue
Every two adjoining sentences in a text have a relationship in some point of view, i.e. coherency. We will investigate the coherency on spoken dialogue. 4) Utilization of the Surface Characteristics
We will investigate usable surface information of sentences. For example, the attitude of the author (speaker) is found in the predicate part of sentences. When such types of sentence are recognized, that makes structuring of sentences easier. \medskip II. Implementation and Experiments In our model, the analyzer has a stack, which can store pieces of rhetorical structure. The action of the analyzer is similar to that of LR(1) parser. That is, if connection between top two elements on the stack is stronger than that between the top element and input sentence, the analyzer acts reduce operation. Otherwise shift operation. The connection is decided according to coherency, predicate pattern, etc. We call the model the Partial Tree Stack Model of Rhetorical Structure. \medskip III. Results We implemented the model and experimented with the system for 100 newspaper articles. Though not all part of our model are implemented, the system generates a tree relative to the rhetorical structure of input text. The evaluation of our system is measured with the distance between original paragraphs on the tree.