Analogical Reasoning for Information Retrieval: a Case Study

The objective of this paper aims to propose an efficient method for recognizing and retrieving information from large image databases. We first provide an overview of current research on image information systems. We then propose a similarity calculation method based on multi-source analogy. To illustrate these ideas, a case study of the Dunhuang Frescoes retrieval and analysis system is presented. We demonstrate that calculating the similarity between images can retrieve the Dunhuang Frescoes information. Introduction Digital image processing plays an important role in an information age. Computer processing methods such as image compression, recovery, and segmentation have found applications in a variety of fields such as Geographic Information Systems, remote sensing, medical image archives, multimedia, and digital libraries. As the prevalence and size of multimedia databases increases, automated recognition techniques that can extract useful and comprehensible features become a critical part of a successful information retrieval process. Significant research has been conducted in image database retrieval, focusing on feature vector computation from an image and generating the feature distance based on certain image measurements. Image database retrieval differs from some traditional classification tasks, including face detection and character recognition. Common data mining or knowledge discovery methods are less effective as they generally require large number of examples and few classes (Rowley et al. 1998). The objective of this paper aims to propose an efficient method for recognizing and retrieving information from large image databases. In this study, we borrow some concepts from cognitive science in analogical reasoning to facilitate information retrieval. Analogical reasoning represents an efficient way of using past experience (Gentner 1983, Carbonell 1983). The process raises several important questions. First of all, what are the significant aspects shared by old and new problems? Secondly, how is the past successful experience selected from a possibly large long-term memory? Thirdly, what knowledge is to be transferred from the past experience to the new solutions? Finally, how does the knowledge transformation process occur? Background Much of the research in image retrieval focuses on extracting and retrieving images by comparing features and choosing the one that is most similar. Generally, texture features are used to represent images. One method preserves the spatial information of the images into texture features, keeping the images in a very narrow domain. Primarily used to perform facial image retrieval and recognition, this technique has a relatively higher hit rate and is more mature and practical than other techniques. Recent research has focused on using multi-feature operations to develop new image retrieval methods. The use of multi-feature functions improves matching accuracy and reduces image information loss through feature extraction, therefore overcoming the disadvantage of image retrieval by a single feature. The QBIC™ of IBM is a system that allows users to query large image databases based on visual image content, i.e. properties such as color percentages, color layout, and textures occurring in the images (Flickner et al. 1995). Database Warehousing and Data Mining 94 2002 — Eighth Americas Conference on Information Systems Analogical Generation and Similarity Calculation Analogy is an important aspect of human learning and thinking. Facing a new task, people tend to recall similar situations and adapt one or more previous solutions to fit the new situation. Analogical reasoning occurs when people recall and use information from prior experiences to solve a new problem (Mayer 1992). Research on analogical problem solving is rooted in cognitive psychology (Gentner 1983, Holyoak and Thagard 1989&1995, Reed 1987). When employing analogical reasoning, several important issues need to be addressed. First of all, the similarity between old and new problems must be identified. The common aspects shared by the old and new problems serve as the basis of a similarity measure, which can be used to search for solutions. Secondly, retrieving past experience is a search process in which the solutions to old problems are examined and measured by the similarity measures. The way in which past experience is represented has a major impact on the efficiency of the search. In addition, the knowledge to be transferred is determined by the nature of the problem, the type of the analogical reasoning applied, and the results of past experience retrieval. Finally, transferring knowledge from a past experience to a current situation is a problem-solving process in itself. In this paper, we propose the concept of multi-source analogy, which can release the traditional analogy from the limitation of presumed restriction, allowing its newly added sources to be mapped to a more extended target field. The similarity calculation based on multi-source analogy can thus be applied to a variety of domains. According to the traditional analogical reasoning (AR) theory, AR can be expressed as: ("~" stands for "similar with") (3.1) B b b t t t b t b t = = ⇒ { , }, { , } ~ ~ 1 2 1 2 1 1 2 2 where B and t stand for source and target, respectively. Both B and t consist of two parts. This means that certain similarity between two objects infers further similarity in the other two. In traditional symbolic artificial intelligence systems, the reasoning process is a binary adopt/abandon substitution. In reality, when someone comes upon a new problem, s/he will likely solve the problem based on past experiences. For instance, to create a new piece of upholstered furniture, a designer may think about the most recent fashions, the feeling of the seat of his car, and perhaps the quality and color of his bedroom curtain. He may use some version of these elements in his design. Therefore, the real analogy procedure is multi-source. In addition, he may feel that the color should be more subdued and that the chair should have some interesting trim. . The final design will reflect known features (like the color of the designer's curtains) and new touches (like the trim). This example illustrates that the reasoning process does not follow the binary substitution rule but, rather, is a process influenced by all related sources. The power is like a magnetic field or electronic field– it exists and varies with distance. Our proposed multi-source analogical generation system differs from the traditional AR system in several significant ways: 1) there are more than two sources as analogues; 2) sources effect the result through the power of their gravity field; 3) the sources and their field power establish a reasoning space; 4) the reasoning process is continuous therefore the intermediate area containing potential meaningful objects; 5) the objects relate with all the original sources just in some degree, i.e. the result objects is a hybrid of all sources. These can be formally described as follows: , where Bi is an analogue. T f B B BN = ( , , , ) 1 2 L Let: