A Question-Answering System Using Unit Estimation and Probabilistic Near-Terms IR
Masaki Murata, Masao Utiyama, Hitoshi Isahara · 2002
Our question-answering system incorporates several new methods. One is unit estimation, which is useful when the answer is a numerical expression and the question sentence does not include any unit expression. This method can be used to estimate a unit expression for the answer by using a statistical test and corpus data and thus improves the results for such questions. Another new method is called probabilistic near-terms information retrieval. This method enables us to use full-size documents in document retrieval without dividing documents into passages and is useful when the answer is not within the passage where relevant terms occur. We confirmed the effectiveness of the unit estimation by using a statistical test (a T-test) and found that the probabilistic near-terms information retrieval improved the performance in Task-1 and Task-2.