Description of NTU System at TREC-10 QA Track

Abstract

In the past years, we attended the 250-bytes group. Our main strategy was to measure the similarity score (or the informative score) of each candidate sentence to the question sentence. The similarity score was computed by sums of weights of cooccurred question keywords. To meet the requirement of shorter answering texts proposed in this year, we adapt our system, and experiment on a new strategy that is focused on named entities only. The similarity score is now measured in terms of the distances to the question keywords in the same document. The MRR score is 0.145. Section 2 will deal with our work in the main task. We also attended the list task and the context task this year. In the list task, the algorithm is almost the same as the one in the main task except that we have to avoid duplicate answers and find the new answers at the same time. Positions of the candidates in the answering texts should be considered. We will talk about this in Section 3. In the context task, how to keep the context, and what the answers of the previous questions can help are the main issues. In our strategy, the answers of the first question are kept when answering the subsequent questions, but the answers of the other ones (denoted by question i) are kept only if question i has a co-referential relationship to its previous one. Section 4 will describe this strategy in more detail.