Research On The Key Technologies Of Ontology Learning From Instance

As Semantic Web technology applying to real world, more and more large-scale ontology appears on the web,such as conservancy[1], linguistic expression[2], transportation[3] and art[4]. In addition, because the knowledge expert can't design or haven't yet designed more particular ontology architecture at the beginning, the number of instances in one class becomes more and more large. For example, WorldNet [2] has only 14 classes and 140000 instances. Hydrologic Ontology [1] contains 5 classes and 2744 instances. Thus, we propose an OntoOn engine which can learn new classes and new concepts from these instances. With this engine, it becomes more convenience and more effective to help the knowledge expert to re-develop the Ontology. My design goal is to setup a new architecture which is scalable, flexible, self-adaptive and user-oriented for the ontology learning.By survey of the character of the Ontology instances, we propose an design architecture for the Ontology learning from the instances--OntoOn: Architecture of Ontology Learning Engine Based on Code System. This plan contains four steps: 1.Data processing. 2. Ontology Feature Selection. 3. Ontology Structure Construction. 4. Labels for Each Cluster. This dissertation solves the problem of the calculation of similarity between each instance.In this project,Ontology Structure Construction is to construct the ontology structure for instance. For recent discovery, I notice that two instances which link to the same instance by the same property have the similar characteristics. Thus, at the first step, we will compute the similarity between two instances by analysis of the links between instances. In addition, we construct the ontology by cluster these instances using the similarity between these instances. However, such algorithms as SimRank [5], LinkClus [6], ReCoM [7], Fingerprinted SimRank [8], and SimFusion [9] have some problems and are not available for our OntoOn system. SimRank, SimFusion have high accuracy but low efficiency. ReCoM, Fingerprinted SimRank improve the efficiency of SimRank, but the accuracy of similarity is much lower. LinkClus promotes both two sides, but it gets the wrong similarity result in some cases. Based on above-mentioned problem, we proposed a S-SimRank algorithm and the experimental results of ACM Data Set [21] show that S-SimRank outperforms other algorithms.