Department of Computer Science University of Crete
This dissertation describes a computational model for detecting analogies between objects and estimating their similarity, based on an analysis of conceptual descriptions of these objects.
The proposed model adopts a representation framework with three general-purpose semantic modeling abstractions (i.e. classification, generalization and attribution), which are adequate for building conceptual models for a wide spectrum of application domains and tasks. Furthermore, their semantics inherently reveal analogies between such conceptual models.
The detection(elaboration) of analogies is achieved by using metrics measuring the distance between conceptual descriptions of objects with respect to each of the prescribed abstractions. The similarity model we propose, consists of three such metrics, namely the classification, generalization and attribution distances.
These metrics distinguish the importance of the specific relations they take into account. In particular, the attribution distance is informed about the importance of the attributes of the involved objects by the so called measures of salience. These measures are estimated from beliefs about three properties of attributes (i.e. the charactericity, abstractness and determinance), which are introduced by our model as predicting the importance of attributes to the elaboration of analogies. The beliefs about the truthness of these properties are measured by functions taking into account specific patterns of representing attributes in conceptual models.
The exact criteria for hypothesizing analogies in the first place and the functions measuring distances and salience ensure the domain independency of the proposed model. Also, they make it operational under non uniform representations of objects(i.e. representations which are not based on specific sets of relations) and tolerant to the absence of explicitly asserted knowledge determining important factors for the detection of analogies. Hence, the proposed similarity model overcomes identified problems of other models for elaborating analogies[Ked88,Hall89] and offers a generally applicable computational mechanism for this elaboration.
The model has been implemented and integrated with the Semantic Index System(i.e. a system for representing scientific knowledge and engineering artifacts). The resulted prototype (SIS/SA) was used in a couple of tasks in the domain of software engineering, namely the analogical reuse of requirements specifications and the integration of requirements viewpoints. These applications demonstrated the applicability of the proposed model in complex tasks involving interdomain analogical reasoning.
SIS/SA was also used in experiments evaluating the consistency of similarity estimates generated by our model with human assessments, as well as, the recall and the time performance of it. Albeit preliminary, these experiments indicated a promising behavior of the similarity model, with respect to the mentioned aspects of evaluation.
Chapter 1: Overview of Analogical Similarity (available in postcript text1.zip )
Chapter 2: Elaboration of Analogies (available in postcript text2.zip)
Chapter 3: Distance and Similarity Functions (available in postcript text3.zip)
Chapter 4: Salience Functions (available in postcript text4.zip)
Chapter 5: Implementation and Evaluation of the Model (available in postcript text5.zip)
Chapter 6: Illustrative Applications (available in postcript text6_1.zip , text6_2.zip , text6_3.zip )
Chapter 7: Comparison with General Models of Analogical Reasoning (available in postcript text7.zip )
Chapter 8: Conclusions and Future Research (available in postcript text8.zip)
Glossary(English-Greek) (available in postcript textG.zip )
Bibliography (available in postcript textL.zip )