Improvement of Efficiency for Interactive Document Retrieval using Transductive Inference

传统的交互式文档检索系统在检索剩余文档时，只使用用户评价过的文档，而没有积极地利用剩余文档中的信息。然而，近年来在机器学习研究领域提出了使用这些未经用户评估的文档的方法。利用这些方法对未评价的文献进行检索，可以迅速提高检索效率。此外，交互式文档检索系统的设计应考虑到人类的识别负荷。本研究的研究目的是利用人工智能方法，特别是机器学习方法，研究和开发基于转换推理的交互式文档检索方法。为此，本研究将转换推理的概念引入到支持向量机的主动学习框架中。我们提出了一种新的文档选择方法，可以显示出接近用户意愿和系统意愿的文档。我们在计算机上开发了这个方法，并使用大型基准数据集对这个方法进行了评估。通过本研究的成果，用户只需输入关键词即可一次性检索文档。在此输入之后，用户评估显示的文档，并且用户可以看到相关文档。因此，用户可以通过转义来避免生成关键字来检索文档。

In the past, traditional interactive document retrieval system retrieved the remained documents by using only the documents that the user evaluated, and had not used information in the remained documents positively. However, the methods using these documents, which are not evaluated by a user, in recent years is proposed in the machine learning research field. The retrieval efficiency can be expected to be improved rapidly by using these methods based on the not evaluated documents. In addition, the interactive document retrieval system should be designed with considering the recognition load of human being In this research, the research purpose is to research and develop the interactive document retrieval method based on the transductive inference using the artificial intelligence methods, especially the machine learning methods.In such a purpose, the concept of the transductive inference was introduced into the frame of active learning with the support vector machine in this research. We proposed a novel document selection method which can display the documents, which are near the user's desire and the system's desire. And we developed this method on a computer and evaluated this method using large bench mark datasets.By the achievement of this research, the user input the keywords to retrieve documents onetime. After this input, the user evaluates the displayed documents and the user can see the relevant documents. Therefore, the user escapes from making keywords to retrieve the documents.