Korea University Question Answering System at TREC 2004

Our QA system consists of two different components. One is for the factoid and list questions, and the other is for the other questions. The components are processed individually, and each result is combined into our submitted run. For the factoid questions, we have tried to find answers by proximity-based named entity search. Given a question, fine-grained named entities for candidate answers are selected, and all the extracted passages containing the named entities and question keywords are scored by a proximity-based measure. List questions are processed in a similar way to the factoid questions, but we empirically give a threshold value to obtain only top n candidate answers. For other questions, relevant phrases consisting of noun phrases and verb phrases are extracted using a dependency relationship to the question target from the initially retrieved sentences. After redundant phrases are eliminated from the answer candidates, final answers are selected using several selection criteria including the term statistics from an encyclopedia. Section 2 summarizes our system for factoid and list questions, and Section 3 for other questions. In Section 4, the TREC evaluation results are analyzed, and Section 5 concludes our work.