文脈を考慮した数学的知識へのアクセスに関する研究

本研究的目的是设计一个数学信息访问（MIA）系统，即一个允许人们有效地访问和处理大量数学信息的系统。我们提出了一个数学信息检索（MIR）模块来实现对数学信息的有效检索。然后，我们设计了一个用于文档浏览的数学实体链接（MEL）模块。后一个模块通过将这些数学表达式链接到维基百科中相应的文章，提供了每个文档中包含的数学表达式的信息。在这一年中，我们取得的成就是：(1)在ntir -12 MathIR任务中对我们的MIR模块进行了评估。结果表明，本系统取得了第一名的成绩。总体结果发表在第12届NTCIR会议上。此外，我们的MIR模块关键部件的详细信息在Information Retrieval Journal上发表。(2) MEL模块的初步设计。我们最初的MEL模块是基于MIR的。评估结果表明，初始模块性能不佳。该结果发表在第18届ICADL上。(3) MEL监督学习方法的发展。我们为此任务构建了一个数据集，然后提出了一种测量给定数学表达式在其包含文档中的重要性的方法，最后实现了几个特征（即数学和文本相似性、数学重要性和数学突出性）。结果表明，后一种MEL模型的精度为83.40%，而初始模型的精度为6.22%。这项工作发表在WSDM 2017会议的SWM上。

The objective of this research is to design a mathematical information access (MIA) system, that is a system that allows people to effectively access and process large amounts of mathematical information. We propose a math information retrieval (MIR) module to allow effective search for math information. Then, we design a math entity linking (MEL) module for document browsing. This latter module provides information about math expressions contained in each document by linking these math expressions to their corresponding articles in Wikipedia.In this year, what we have achieved are:(1) The evaluation of our MIR module in the NTCIR-12 MathIR task. The results showed that our system finished at the first place. The overall results were published in the 12th NTCIR conference. In addition, the detail of the key component of our MIR module was published in the Information Retrieval Journal.(2) The initial design of MEL module. Our initial MEL module was based on the MIR. The evaluation showed that the initial module did not perform well. This result was published in the 18th ICADL.(3) The development of a supervised-learning approach for MEL. We constructed a dataset for this task, then proposed a method for measuring the importance of a given math expressions in its containing document, and finally implemented several features (i.e. math and text similarity, math importance, and math prominence). It was shown that the latter MEL module achieved a precision of 83.40%, compared with 6.22% for our initial module. This work was published in the SWM at WSDM 2017 conference.