氏名: 伊藤 雅浩 (m06728)
I propose a method to cope with ill-formed uttarance on an on-line
interpretation model, which processes input words incrementaly.
From both computational and cognitive points of view, there seem to be
a lot of benefits to process dialogue incrementaly.
However, spontaneous speech includes so many disfluencies, but the speakers
often correct disfluencies by themselves.
This phenomena is what is called self-repair.
The trouble is that it is required to make an interpretation again,
when the speaker changed his/her intension or corrected his/her
In order to realize the interpretation of self-repaired utterance
incrementally, I analized syntactic characteristics of Japanese self-repairs.
The analysis gave me an idea to interpret those utterances.
According to my method, the system interprets input sentence by using
three buffers; Noun Phrase buffer, Case buffer and Verb Phrase buffer.
Each buffers has specific mechanisms to detect and correct self-repairs.
As a result of the implementation of the method, I got satisfactory effect,
which means that the system can detect and correct them at approximately
80％ rate in all the self-repairs extracted from human dialogue corpus.