Basic Studies on Automatic Text Summarization as an Aid for Human Intellectual Activities

在这个项目中，我们将重点关注以下两个主题：1。2.减句总结；对多个文献进行汇总，得到如下结果。在句子缩减方面，我们提出了一种删除副词动词短语的方法。这种方法是基于这样一种观察：如果修饰被动词修饰的名词的动词种类有限，那么修饰动词短语很容易被名词联系起来，并可能被删除。这种修饰动词的多样性是用熵来衡量的。我们还提出了一种删除多个修饰短语的方法。修饰短语的可删除程度是通过名词在短语中的重要性和相互信息来估计的。我们开发了一个多文档摘要系统GOLD。以往的经验表明，适当地结合多种摘要技术可以开发出良好的自动摘要系统。因此，我们结合了各种传统和新引入的总结技术来开发GOLD。NTCIR 3的TSC 2评价结果令人满意。在多文档汇总中，要汇总的文档集通常有多个主题。但是，用户不一定对所有主题都感兴趣。因此，需要用户自定义摘要。为了满足这一需求，我们开发了一个具有用户交互的多文档摘要系统。系统建议从文档集中提取的关键词进行汇总，用户从中选择合适的关键词。NTCIR 4的TSC 3评价结果显著，特别是在内容评价方面。作为相关结果，我们提出了一种从单个语料库中获取日语名词消缩写知识的方法。这些知识是有用的，例如，信息检索，词义消歧和总结。

In this project, we put the focus on the following two themes :1.sentence reduction for summarization, 2. multiple document summarization, and the following results were obtained.As for the sentence reduction, we proposed a method for deleting adnominal verb phrases. This method is based on the observation that if the kinds of verbs which modify the noun modified by the verb is limited, then the adnominal verb phrase can be easily associated with by the noun and maybe deleted. Such diversity of modifying verbs is measured by entropy. We also proposed a method of deleting multiple adnominal phrases. The degree of deletability of an adnominal phrase is estimated by the importance of the noun in the phrase and mutual information.We developed a multiple document summarization system GOLD. Previous experiences show that a good automatic summarization system can be developed by combining appropreately a number of summarization techniques. Thus we developed GOLD by combining a variety of summarization techniques both conventional and newly introduced. The evalution results at TSC 2 of NTCIR 3 was satisfactory.In a multiple document summarization, the document set to be summarized usually has multiple topics. However, a user may not necessarily to be interested in all topics. Thus, a user customized summary is needed. To cope with this need, we developed a multiple document summarization system with user interaction. The system suggests keywords extracted from the document set to be summarized and the user choose appropreate keywords among them. The evalution results at TSC 3 of NTCIR 4 was remarkable, in particular, in content evaluation.As a related results, we proposed a method for acquiring knowledge from a single corpus on disabbreviations for Japanese nouns. This knowledge is useful e.g., for information retrieval, word sense disambiguation and summarization.