Context Based Event Detection Model

This dissertation presents a Context based Event Detection model for social text streams. The main motivation behind conducting this research has been because there exists no comprehensive context identification model based on information pertaining to events as the constituent unit for Context Identification. Isolated attempts have been made, whereby event identification based on temporal, location, participants, text tags have been reported in the literature, but none can be classified as a comprehensive Context Identification model.\ud The proposed layered model extracts, classifies, develops linkages and presents the events as a holistic representation of context. The selection of social text streams (emails) for the proposed model poses the challenge of event identification from unstructured short event descriptions as presented in the emails. \ud The event extraction is performed using language's grammatical part of speech triggered Finite State Automata (FSA). For each event attribute, a set of FSA are constructed that are objected towards the recognition of phrases containing the specific event attribute. For the classification of event emails into predefined event classes, a template based graph mining and matching approach is employed. The authors have used a graph representation for emails that significantly supported in reducing the complexity of processing graphs. A set of email graphs is processed for generating the templates representing event classes and the emails are classified into different folders depending upon the template to which they matched. Event algebra is developed for detection and analysis of linkages between events. Finally, a contextual representation of the events is proposed. The usability engineering process is applied to develop the design. \ud Each component of the whole model is evaluated independently. FSA are found to be efficient, accurate, extendable and comprehensive for the event extraction. The template based graph mining and matching approach produces high classification accuracy. The linkage detection and analysis component is evaluated over events in a university domain. \ud The results depict a high degree of accuracy. Future enhancements have been proposed for each layer of the model comprehensively at the end of each chapter and holistically as future directions