This research aims to make an automatic conversation system which can converse with people naturally. We intend to suggest a model for an intelligent conversation system which represents the human language competence ― human can analyze the communicat ...
This research aims to make an automatic conversation system which can converse with people naturally. We intend to suggest a model for an intelligent conversation system which represents the human language competence ― human can analyze the communication partner's intentions, and construct corresponding sentences properly ― and the human cognitive competence ― human can learn and infer on the basis of encyclopedic knowledge. But for the practical purposes, we limit our research range to the conversation between a human user and an NPC(non player character) in MMORPG games.
Existing conversation systems usually use a methodology that the computer has had conversation patterns between the system and human users, so that the system can output corresponding sentences as a human user inputs a sentence. But formally similar sentences can conceive very different intentions(ex. "안녕하세요 " as a conventional greeting expression vs. "어머님은 안녕하세요 " as a question which asks for specific response) and formally different sentences can conceive very similar intention(for example, both "아르바이트를 하고 싶은데요." and "일 좀 주세요." have same intention of 'asking something').
Besides, if the conversation system doesn't have encyclopedic knowledge, communication through that system can't be continued naturally. Without knowledge, the system may output only fragmentary sentences repeatedly whatever the input sentences are.
These two problems should be solved first in order to improve performance of conversation system. That is, the system can understand intention of human conversation whatever form it has, and can respond to the input sentence properly based on encyclopedic knowledge.
In order to achieve this purpose, we subdivide the conversation system into several modules. Among those modules, we specially pay attention to set up 'speech act-decision module', 'corresponding speech act-decision module' and 'knowledge base-exploring module'.
The speech act-decision module is to catch a user's intention in an input sentence. For example, the sentence "지금 몇 시인지 알려 줄래 " isn't a yes/no question. That sentence asks a specific answer like "10시야." as the sentence "지금 몇 시인지 알려 줘.", which is an imperative sentence. In that sentence, the combination of ‘-어 주다’, ‘-ㄹ래’ and ‘ ’ signals that the sentence isn't simple question but asking. Like this, we survey linguistic forms which signal certain speech act, and arrange collected informations in terms of speech act types. So, the speech act-decision module consists of such information.
The corresponding speech act-decision module is to decide corresponding sentence's speech act type that matches well with the user input sentence. In order to set up this module, we examine verbal corpora which consist of undergraduates' conversation, and then repeatedly adjust information statistically which improves matching between speech act types.
The knowledge base-exploring module presupposes the knowledge base. The knowledge base indicates a archive of encyclopedic knowledge that the computer system has. When a system user inputs a sentence which asks a response based on encyclopedic knowledge like "아르바이트는 어디서 구할 수 있어 ", the computer system has to search for relevant knowledge in order to answer the question properly. It is the knowledge base-exploring module in which a logic how system searches for relevant information in the knowledge base efficiently is recorded.
In order to search for relevant information in the knowledge base efficiently, we subdivide the knowledge base into the frame knowledge base, the semantic web knowledge base, the fact knowledge base, the game-world knowledge base and the thesaurus.
Through realizing all these processes, we have proposed a whole model of an intelligent conversation system.