Study on Information Retrieval based on Similarity Calculation of Intra-Document Structure

本研究的目的是建立基于“内容”的信息检索方法。通过两年的研究，我们得到了以下几个方面的结果：1.基于段落相似度的文档内部结构提取：不仅利用文档内部信息，还利用文档间信息，提高了文档相关部分的检索效率。基于机器学习和抽取模式的多策略命名实体识别器：针对命名实体任务，将这两种策略相结合，提高了命名实体识别的准确率。数字信息提取及其在问答中的应用：我们认为问答是一种理想的上下文检索系统。命名实体对应于4W型问题的答案。另一方面，它是与H型问题相对应的数字表达式。因此，我们提出了一种提取带有上下文的数值表达式的方法，作为问答系统的一部分。

The purpose of this research is establishment of the method for "content"-based information retrieval. In our research, the "content" is regarded as the combination of the following items : a) Logical structure of document annotated by tags, b) Text, and c) Information extracted by the technology of Information Extraction.Through the two year research, we obtained the following results :1. Extraction of structure of intra-documents based on similarity among passages :By using not only intra-document information but also inter-document information, we improve the effectiveness of retrieving relevant portions of document.2. Multi-strategic named entity recognizer based on machine learning and extraction patterns :By combining those two types of strategies for named entity task, we improve the accuracy of recognition of named entities.3. Extraction of numerical information and its application to Question Answering :We consider "Question Answering" is the one of the ideal context retrieval system. Named entities correspond to the answer for the 4W-type questions. On the other hand, it is numerical expressions what corresponds to H-type questions. Therefore, we proposed a method to extract numerical expressions with its context as a part of a QA system.