Exploiting Semantic Technologies For Web-based Applications

The topic of information extraction and management has gained extensive attention in recent years because of huge amount of data produced by social networks, blogs, information centric applications, etc. The data can be in the form of text, images, videos or combination of all. Unstructured or unorganized data is the easiest form of data which can be created in any application scenarios. As a result, semantic technologies have recently been evolved which can effectively process and structure multimodal content on variety of Web-based applications. Semantic technologies incorporate ontologies and semantic relevance measures to extract and manage multimodal content on the Web. This dissertation endeavors to explore semantic technologies for two hot Web-based applications, i.e., automatic Web video classification and disaster management. For Web video classification, our primary focus is to provide visual content-based solution to categorize videos because text information is sometimes noisy, ambiguous or absent. The main contributions of this research are as follows.To solve the problem of automatic Web video classification purely based on visual content-based features, first we utilize Large-Scale Concept Ontology for Multimedia(LSCOM) to extract useful visual information from images/videos and then semantic relevance measures are harnessed with already extracted visual information in a novel way to determine the true video category. In general, three-step framework is proposed for Web video classification. First, we train Content-based category-predictive (CNC) classifier for each category by exploiting visual features to classify Web videos. Second, we refine CNC classifiers by harnessing concept level semantics at keyframe-level through VIREO-374 concept classifiers (LSCOM assisted) into each category, which improves the performance of Web video categorization. Furthermore, the significance of each concept for each category is measured in a novel way by leveraging globally consistent Flickr-enriched context space(FECS), which are named as Category-specific Concept (CSC) classifiers. Later, CSC classifiers are fused with CNC classifiers at a keyframe-level to refine them. Third, the Context-based category-predictive (CXC) classifiers induced from titles and tags are further combined with the refined CNC classifiers to reinforce the video classification performance.Experiments on two large scale Web video datasets, MCG-WEBV and Columbia Consumer Video (CCV) datasets, demonstrate that the proposed approach achieves promising performance.In the next phase of our research, we solve the problem of automatic Web video classification in a different way. First, we leverage external support and create semantic list of category discriminable terms and then utilize semantic relevance measure to map LSCOM assisted high-level concepts to category discriminable terms to recognize true video category.In general, three-step framework is proposed. Firstly, content-based video classification is presented, where initially category classifiers are trained and then semantic relevance is measured between the high-level concepts (LSCOM assisted) from each video and the Category Discriminative Terms (CDTs) through Normalized Google Distance (NGD). CDTs are built through Open Directory Project (ODP) and large scale Web videos. Secondly, context-based video classification is proposed, where the Vector Space Model (VSM) is first employed to compute the similarity between text features from each video and CDTs, and then measure the semantic relevance between text features from each video and CDTs through NGD. Finally,classification results from content and context features based methods are further fused to boost the performance of Web video classification. Experiments on large scale video dataset validate the effectiveness of proposed solution.To further explore the roles of ontologies, we considered another very useful Web-based application, i.e., disaster management. For disaster management, we study the disaster domain in depth through different online sources and develop disaster ontology to automatically extract text information from Web documents in case of an emergency. We propose two-fold use of the ontologies for semantic Disaster Management System (DMS). Firstly, ontologies can be used as background knowledge for effective discovery and selection of resources.Secondly, after population, they can be used for reasoning to support decision making process.We propose disaster ontology by utilizing semantic technologies that covers prominent aspects of disaster management including intrinsic properties of disaster, losses caused, services required, etc. Experiments performed on disaster related Web documents demonstrate that, as compared to conventional crawling, ontology-driven crawling is more efficacious for resource discovery and selection.