In one's life, a person acquires a lot of information and knowledge via natural language. For example, reading is knowledge acquisition via written language, and both hearing and discussion are knowledge acquisition from spoken language. From the viewpoint of knowledge acquisition, dialogue is knowledge transfer process, in which speaker's knowledge (information) is transferred to a hearer. We regard this as an important function of dialogue, however, neither its model nor its implementation has been treated as a main problem in dialogue research. In this research, we focus on the knowledge transfer function of dialogue. For the hearer, knowledge transfer is understanding of utterances, i.e., extraction of information on the utterances. Our final goal is to make a computer program that acquires information and knowledge through dialogue with users.
We studied a model of understanding of utterances in the last year. We believe that hearer's knowledge plays an important role in the process of understanding of utterances. From this point of view, we developed a new model that has the following features.

  • Understanding an utterance is the process of matching between the utterance and hearer's knowledge. Every content word in the utterance has to be connected to an entity in the hearer's knowledge. If every content word is connected to an entity, the utterance is understood and the new information on the utterance is integrated into the hearer's knowledge.
  • The response to the utterance is determined by the result of the matching.

    This year, we extended that model and developed a small prototype system, which can converse with users. The following is an example of conversation.

    {Dialogue 1}
    (The current topic is a room${}_1$.)
    \User{1} heya niha ookina tsukue ga aru (\Utter{11})
    {There is a big desk in the room.}
    \User{1} sono tsukue ha koukada (\Utter{12})
    {The desk is expensive.}

    {Dialogue 2}
    (The current topic is the room${}_1$.)
    \User{2} heya niha koukana tsukue ga aru ka (\Utter{21})
    {Is there an expensive desk in the room?}
    \System hai (\Utter{22})
    \User{2} sore ha chiisai ka (\Utter{23})
    {Is it small?}
    \System iie (\Utter{24})

    In this system, we have concentrated on the implementation of the connecting process of content words to entities in the hearer's knowledge. Two triggers invoke this connecting process. The first trigger is surface {key} forms. The following three forms require that the referents (entities to which the content words are connected) exist already: pronouns, noun phrases marked by {shijisi} (ex. {sono}), and noun phrases marked by {ha}. If the system encounters one of these forms, it searches the current topic and the previous context for the entity to which the form ought to be connected. For example, in \Utter{11} the phrase { heya niha} (in $\phi$ room) invokes the connecting process because of {ha}-marked form, and is connected to the current topic ``room${}_1$''.
    Another trigger is an end of an utterance. When the end of the utterance is recognized, the system tries to integrate the information on the utterance into its knowledge. For the integration, the system tries to connect unconnected content words in the utterance. If the system fails to connect a word to a entity, it creates a new entity and connects the word to it. For example, in \Utter{11} the phrase {ookina tsukue} (big desk) is connected into a new entity ``desk${}_2$''. In \Utter{21} the phrase {koukana tsukue} (expensive desk) is connected into the entity ``desk${}_2$'', because the system already knows that ``the desk${}_2$ is expensive'' from \Utter{12}.
    In this system, interrogative sentences are processed in the almost same manner as declarative sentences. For example, in \Utter{21} the part except {ka}, i.e. {heya niha koukana tsukue ga aru}, is a declarative sentence, which means ``there is an expensive desk in the room.'' This part is processed the same as the declarative sentence. The only difference is the existence of {ka}. {Ka} is regarded as a request of a response, which informs the result of matching between the information on the declarative part and the system's knowledge. If the information is already existed in the system's knowledge, the system answers {hai} (yes): If the information contradicts the system's knowledge, the system answers {iie} (no): If the information is new, the system answers {wakarimasen} (I don't know).

    Keywords: Dialogue, Understanding, Knowledge Acquisition.