Fisheye Matching: Viewpoint-Sensitive Feature Generation Based on Concept Structure

In this paper, we propose an extended Vector Space Model (VSM) called Fisheye Matching method, which generates features related to the users’ viewpoints based on an electronic dictionary. We have also developed a GUI system employing Fisheye Matching method, which can help users order the vast collection of documents in several ways employing the users’ viewpoint information extracted by the Fisheye Matching method. When a user gets some ideas from papers or articles, he may organize his thoughts by relating incoming information with knowledge which already exists in his mind. This process is getting harder for him in proportion to the volume of information which he considers, and it is useful to illustrate a concept structure on a paper or on a display, which leads to reducing his confusion. We assert that ordering documents while reading is an effective way of dealing with the vast collection of documents, and the systems which assist such processes should be able to find relations among documents based on users’ viewpoints/interests. From this point of view, we have proposed an extended VSM called Fisheye Matching method [1] to perform vector matching on the vector space sensitive to users’ viewpoints. Each feature in the Fisheye Matching method is generated as a set of words which belong to the same concept (meaning) in a dictionary. Choosing concepts (features) appropriate for the users’ viewpoints from a dictionary, a vector space is constructed so that the matching results can be expected to be more preferable for them. We have also proposed the algorithm to find concepts appropriate for the users’ viewpoints from training sets of documents. Furthermore, each concept in a dictionary has a heading information, which can be presented explicitly to the users as a kind of their viewpoint information. Some experiments on document retrieval tasks have been performed, and Table 1 shows parts of concepts found and used as features during the experiments. 1 We have used the EDR electronic dictionary developed by Japan Electronic Dictionary Research Institute, Ltd. : http://www.iijnet.or.jp/edr/ Table 1. Examples of concepts extracted as features in the case of retrieving documents about medical topics ID Heading Term Group Words (stemmed) 3f98b3 value of health diseas sickne health . . . 444506 component of living body protei immuno choles dna 30f6da internal organs eye heart lung knee . . . 3f969e disease syndro aids cancer cold . . . 44479c medical supplies drug medici laxati acid . . . 30f6f7 medical instruments bandag cathet glasse . . . Fig. 1. Fview: GUI system for document ordering support From both this table and its precision property, it was confirmed that the Fisheye Matching method can not only retrieve documents in which the users take interest, but also supply them with useful information on their viewpoints. We have also developed the GUI system which assists users to order documents with the Fisheye Matching method (Fig. 1). By using the Fisheye Matching method, the system can extract the users’ viewpoint information from the diagrams produced by them, which information cannot be used by the system only to retrieve documents suited for the users’ current interests, but also to indicate the similarity among documents of which they may not be aware. Furthermore, the users’ viewpoint information are also presented to them as a list of heading information of extracted concepts. Several students actually used this system, and it has been confirmed that an effective assist for users can be achieved.