Filtering News from Document Streams: Evaluation Aspects and Modeled Stream Utility

Events like hurricanes, earthquakes, or accidents can impact a large number of people. Not only are people in the immediate vicinity of the event affected, but concerns about their well-being are shared by the local government and well-wishers across the world. The latest information about news events could be of use to government and aid agencies in order to make informed decisions on providing necessary support, security and relief. The general public avails of news updates via dedicated news feeds or broadcasts, and lately, via social media services like Facebook or Twitter. Retrieving the latest information about newsworthy events from the world-wide web is thus of importance to a large section of society. As new content on a multitude of topics is continuously being published on the web, specific event related information needs to be filtered from the resulting stream of documents. We present in this thesis, a user-centric evaluation measure for evaluating systems that filter news related information from document streams. Our proposed evaluation measure, Modeled Stream Utility (MSU), models users accessing information from a stream of sentences produced by a news update filtering system. The user model allows for simulating a large number of users with different characteristic stream browsing behavior. Through simulation, MSU estimates the utility of a system for an average user browsing a stream of sentences. Our results show that system performance is sensitive to a user population’s stream browsing behavior and that existing evaluation metrics correspond to very specific types of user behavior. To evaluate systems that filter sentences from a document stream, we need a set of judged sentences. This judged set is a subset of all the sentences returned by all systems, and is typically constructed by pooling together the highest quality sentences, as determined by respective system assigned scores for each sentence. Sentences in the pool are manually assessed and the resulting set of judged sentences is then used to compute system performance metrics. In this thesis, we investigate the effect of including duplicates of judged sentences, into the judged set, on system performance evaluation. We also develop an alternative pooling methodology, that given the MSU user model, selects sentences for pooling based on the probability of a sentences being read by modeled users. Our research lays the foundation for interesting future work for utilizing user-models in different aspects of evaluation of stream filtering systems. The MSU measure enables incorporation of different user models. Furthermore, the applicability of MSU could be extended through calibration based on user behavior.