QA system Metis based on web searching and semantic graph matching

Abstract

We have developed a question-answering system Metis with natural-language interface. Metis generates the answer to a question by comparing the semantic graph of the question sentence with sentences discovered on the Internet as knowledge source. Specifically, we first get a set of semantic frames for the question sentence, as the output from a semantic analysis system, SAGE. Then we extract several keywords from all semantic frames using SVM. After that we search the Web to find knowledge sentences based on the keywords and input each knowledge sentence into SAGE in order to get its semantic graphs similarly. Finally, the similarities between the semantic graph of the question sentence and that of each knowledge sentence are calculated to determine the most reliable knowledge sentence, in which a constituent is chosen as the answer to the question. An experiment to examine the effectiveness of our method showed that 65% of the questions for which suitable knowledge sentences had been found were replied correctly.