Abstract
Any recent application of an expert system requires an effective way to build a very large-scale knowledge base. An on-line massive document, such as a corpus, has a potential possibility to be utilized in automatically preparing such knowledge base. Provided that a natural language interface is available, a user can inquire to the acquired knowledge for a specific topic he wants to know in an interactive way.
This paper first makes investigation on semantic structures required in such a corpus oriented knowledge inquiry system focusing on plan-structure and LFG representation. It then proposes a new knowledge representation framework called a PFT-network. This network was introduced to effectively acquire knowledge from corpus with reduced domain knowledge for both the input corpus and user's utterance. Experimental results of extracting knowledge are given and it is demonstrated that the presented approach suggests the feasibility in knowledge acquisition in view of understanding input sentences so that inter-relationship of knowledge is reflected in the obtained knowledge.
