Joint Learning of Structural and Textual Features for Web Scale Event Extraction

The web has become the central platform and marketplace for the organization, propagation of events and sale of tickets of any kind. Such events range from concerts, workshops, sport events, professional events to small local events. Individual’s event choices vary tremendously based on preferences and lifestyle. Online users use the web to inform themselves about new events near their location of choice, and potentially use the website to purchase tickets or make a reservation for such events. Event Extraction from the web is a particularly difficult type of information extraction, dealing with the detection of specific types of events and its attributes mentioned in source language data. Traditional research in event extraction focuses on the extraction of political, cultural, or other general interest events from text. Such text is typically editorial news content, e.g (Kuzey, Vreeken, and Weikum 2014), or more lately from social media such as Twitter, e.g. (Ritter, Etzioni, and Clark 2012). This research, however, covers events presented in tables, lists, or most crucially on single pages devoted to that event. This thesis focuses on both the discovery and extraction of such “single event pages”. This work is inspired by a series of works on inducing wrappers for the extraction of specific document types from the web. For example, (Wang et al. 2009) proposes an approach for learning to extract news articles and their basic attributes from a very small training corpus. Though inspired by this work, the approach presented here differs considerably, in scope and techniques used. In scope, I am targeting events, which carry many more attributes than the document types targeted in the above work, and where attributes may occur both in the template structure (as in (Wang et al. 2009)) and in the event descriptions. Further, my approach balances the need for more training data due to the more complex domain with semi-supervised methods for acquiring that training data.