Third International Conference on Spoken Language Processing (ICSLP 94)
This paper proposes a dialog analyzing method that stores and uses some pieces of information in a certain local context, in order to solve actual problems. First, the authors examined English dialog corpora in order to find constraints in natural conversations. It has been found that more than 70 percent of the content of ellipses in the corpora that must be recovered when translating into Japanese, can be extracted from the last sentence uttered by the other person just preceding the current sentence. The authors then implemented a dialog analyzing method on an English-to-Japanese translation system that handles written texts. The method used memorizes and utilizes a local context constructed by both the current sentence and the last sentence of just the previous utterance by the other person. Last, the authors examined the feasibility of the proposed model by translating other English dialog corpora, transcriptions of telephone conversations. It has been found that the model generates appropriate Japanese translations in some linguistic phenomena that are relating to discourse, such as Ellipsis, Polysemy disambiguation, and Selection of Japanese TOPIC/SUBJECT particles W and 'ga.'.
Bibliographic reference. Kamei, Shin-ichiro / Doi, Shin-ichi / Komatsu, Takako / Akamine, Susumu / Iida, Hitoshi / Muraki, Kazunori (1994): "A dialog analysis using information of the previous sentence", In ICSLP-1994, 931-934.