TREC 2003 Question Answering Track at CAS-ICT
Yi Chang, Hongbo Xu, Shuo Bai · Text REtrieval Conference · 2003
In our system, we make use of Chunk information to analyze the question. A multilevel method is fulfilled to retrieve a candidate Bi-sentence. As to answer selecting, we proposed a voting method. We figure out the performance of each module, and our study shows that 65.54% information has lost in document retrieval and Bi-sentence retrieval. Keyword: Chunk, Multilevel retrieval, voting It is the third time we take part in the TREC-QA track. We undertake the main subtask and submit three runs for evaluation. Our QA system incorporates several useful tools. The first is LT_CHUNK that is developed at University of Edinburgh. We use LT_CHUNK to get the chunks of the sentence and the POS tags of different words. The second is GATE that is developed by University of Sheffield. We use GATE to identify some Named Entities. In our previous QA system, we tried different kinds of elaborate algorithms, but the results were not satisfactory, and we didn’t make it clear what the performance of each step in our system is. So we try to figure them out this year, and our study shows: 65.54% information lost in document retrieval and Bi-sentence retrieval. 2. System Description Our system contains four major modules, namely Question Analyzing Module, Multilevel Bi-sentence Retrieval Module, Entity Recognizing Module and Answer Selecting Module. However, for those definition questions, the Entity Recognizing Module is unnecessary. The system architecture is represented as below.