Expertise Retrieval in Enterprise Microblogs with Enhanced Models and Brokers

Expert finding in enterprise corpora is motivated by the need to for experts to help solve problems and has become an important research area. The existing schools of research are dominated by probabilistic language models, which are either candidate-based or document-based models. Both of these types of models are typically carried forward both to Web documents and microblogs; ignoring inherent structural differences between them. A key observation is that the thread context provided by microblogs has the unique advantage of propagating expertise evidence through the "post-and-reply" pattern. On the other hand, the short length of microblogs presents a research challenge for identifying expertise. Motivated by these observations, three techniques have been introduced in this dissertation. These are: (a) An approach that enriches the document-candidate association by utilizing the thread structure of microblog; (b) A machine learning-based approach that improves the quality of collected expertise evidence; (c) A filtering technique that utilizes the behavior pattern of experts. With these three techniques, Enhanced Models are built upon existing candidate and document models to improve the expert finding performance. An experiment was conducted to show that Enhanced Models outperform baseline models. It also verifies that the contextual information provided by thread is helpful to expert finding. One important insight is that candidate models perform better than document models in the microblog environment, which is different from the conclusion regarding Web documents. The cause is that the document model is less fault-tolerant to the identification of expertise keywords, which is often poorly calculated due to the short length of documents in microblogs and consequent sparsity and noise. Another insight is microblog noise can make document-centric associations perform worse than candidate-centric associations.;Another social media phenomenon discovered by the research is the "broker" that assists with the expert finding task within the enterprise. The importance is that brokers can help connect people to experts when expert finding systems fail to locate experts. The research contributions related to brokers are: (a) A data analysis on an enterprise microblog dataset to show when experts cannot be found by existing expert finding systems, brokers can be utilized for finding experts; (b) A set of Enhanced Models are introduced for broker finding; (c) An experiment comparing different models of broker finding. The research shows brokers are useful for expert finding and motivates future research.