Improvement of accuracy of question-answering based on question reformulation

本研究旨在提高单项式问答的准确率。列表类型的问答是系统要求列举给定问题的所有正确答案的任务。在我们提出的方法中，我们利用现有问答系统给出的答案候选答案的分数分布。首先，我们假设分布是两个分布的混合，即一个正确答案和一个不正确答案。其次，我们还假设这两个分布都是正态分布。在这些假设下，我们提出了一种利用EM算法从混合分布中分离正确答案的方法。另一方面，这两个分布的边界有时可能不清楚，因为问答系统有可能无法产生适当的分数。在这种情况下，我们根据共现…重新计算分数实验结果表明，分数分布的利用对于列表类型的问答是有效的。然而，基于共现的重新计算分数的方法并不有效，因此，我们还研究了一种基于上述发现的方法，即两个分布边界的区别程度指示了系统对给定问题找到的答案候选的适当性。我们将“问题改写”的概念概括如下。我们通过对原始问题应用几种可能的解释来生成一组候选问题。在这种情况下，可以通过选择合适的问题候选者及其答案候选者来实现问题的重构，例如，将问题重构应用到上下文问答中，可能会提高问答的准确性。这一概念也可用于提高跨语言问答的准确性。较少

This research originally aimed at the improvement of accuracy of list-type question answering. The list-type question-answering is the task in which a system requested to enumerate all correct answers to a given question. In our proposed method, we utilize the distribution of the score that an existing question answering system gives to answer candidates. First we assume that the distribution is a mixture of two distributions, i.e., one of the correct answers and one of the incorrect answers. Second we also suppose that each of the two distributions is a normal distribution. Under these assumptions, we proposed a method to separate the correct answers from the mixture of the distributions by using the EM algorithm. On the other hand, the boundary of the two distributions may be occasionally indistinct because there is a possibility that the question-answering system fails to produce appropriate scores. In such a case, we perform a re-calculation of scores according to the co-occurrence … More of each answer candidate and the keywords of the question sentence in the knowledge source.The experimental results show that the utilization of the score distribution is effective for the list-type question-answering. However, the re-calculation of scores based on the co-occurrence does not work effectively.Therefore, we also investigated a method based on the above findings, that is, the degree of distinctness of the two distributions' boundaries indicates the appropriateness of answer candidates that are found by the system for a given question. We generalized the notion of "question reformulation" as follows. We generate a set of question candidates by applying several possible interpretations to the original question. In this situation, the question reformulation can be achieved by selecting appropriate question candidates and their answer candidates.For example, when we applied the question reformulation to contextual question-answering, the accuracy of question answering may be improved. The notion also can be used for the improvement of accuracy of cross-lingual question answering. Less