Research On Social Tagging Mechanism And Its Application On Information Retrieval

Social tagging is a new way for Web users to share, organize and manage resources.With the rapid development of Web 2.0, social tagging systems become more and morepopular nowadays. This dissertation studies the key technologies of social tagging mech-anism and its application in information retrieval. The content of this dissertation in-cludes:(1) Tag recommendation in social tagging systems: Personalized tag recommenda-tion in social tagging systems is to provide a user with a ranked list of tags for a specificresource that best serves the user's needs. Many existing tag recommendation approachesassume that users are independent and identically distributed. This assumption ignoresthe social relations between users, which are increasingly popular nowadays. In this dis-sertation, we investigate the role of social relations in the task of tag recommendationand propose a personalized collaborative filtering algorithm. In addition to the social an-notations made by collaborative users, we inject the social relations between users andthe content similarities between resources into a graph representation of folksonomies.To fully explore the structure of this graph, instead of computing similarities betweenobjects using feature vectors, we exploit the method of random-walk computation of sim-ilarities, which furthermore enable us to model users'tag preferences with the similaritiesbetween users and tags. We combine both the collaborative information and the tag pref-erences to recommend personalized tags to users. We conduct experiments on a datasetcollected from a real-world system. The results of comparative experiments show that theproposed algorithm outperforms state-of-the-art tag recommendation algorithms in termsof prediction quality measured by precision, recall and NDCG.(2) Tag sense disambiguation in social tagging systems: Due to the lack of a unifiedtaxonomy or ontology, tags may be used ambiguously, i.e., the same tag may be used fordifferent meanings. Disambiguating tag senses can benefit many applications leveragingfolksonomies as knowledge sources. In this dissertation, we propose an unsupervised tagsense disambiguation approach. For a target tag, we model all the annotations involvingit with a 3-order tensor to fully explore the multi-type interrelated data. We performspectral clustering over the hypergraph induced from the 3-order tensor to discover the clusters representing the senses of the target tag. We conduct experiments on a datasetcollected from a real-world system. Both the supervised and unsupervised evaluationresults demonstrate the effectiveness of the proposed approach.(3) Ontology learning based on social annotations: The folksonomies built fromthe large-scale social annotations made by collaborative users are perfect data sources forbootstrapping Semantic Web applications. In this dissertation, we develop an ontologyinduction approach to harvest the emergent semantics from the folksonomies. We proposea latent subsumption hierarchy model to uncover the implicit structure of tag space, anddevelop our approach on basis of this model. We identify tag subsumptions with a set-theoretical approach and model the tag space as a tag subsumption graph. While turningthis graph into a concept hierarchy, we address the problem of inconsistent subsumptionsand propose a random walk based tag generality ranking procedure to settle it. We proposean agglomerative hierarchical clustering algorithm utilizing the result of tag generalityranking to generate the concept hierarchy. We conduct experiments on a dataset collectedfrom a real-world system. Both qualitative and quantitative experimental results show acompetitive performance of the proposed approach.(4) Page ranking based on social annotations: With the rapid development of socialtagging systems, large amount of social annotations have been created by large crowdof collaborative users, forming a new dimension of accessing the quality of web pages.In this dissertation, we propose a novel page ranking algorithm for improving web searchperformance. We explore social annotations by effectively combining the language modelof pages and users with the mutual reinforcement between pages and users. First, the usersand pages are modeled by topic models with their corresponding tags. Then, the mutualreinforcement between pages and users is modeled with a bipartite graph. The mutualreinforcement relations are further weighted by the coherence between annotating tagsand language model of pages and users. Finally, the importance of pages and users are si-multaneously calculated in an iterative fashion based on both query relevance and mutualreinforcement. Experimental results show that the proposed algorithm outperforms otherstate-of-the-art algorithms in retrieval performance.