To Link or Not to Link? A Study on End-to-End Tweet Entity Linking

Proceedings of NAACL-HLT 2013, pages 1020–1030, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics To Link or Not to Link? A Study on End-to-End Tweet Entity Linking Stephen Guo Stanford University sdguo@cs.stanford.edu Ming-Wei Chang Emre Kıcıman Microsoft Research {minchang, emrek}@microsoft.com Abstract Information extraction from microblog posts is an important task, as today microblogs cap- ture an unprecedented amount of information and provide a view into the pulse of the world. As the core component of information extrac- tion, we consider the task of Twitter entity linking in this paper. In the current entity linking literature, mention detection and entity disambiguation are fre- quently cast as equally important but distinct problems. However, in our task, we find that mention detection is often the performance bottleneck. The reason is that messages on micro-blogs are short, noisy and informal texts with little context, and often contain phrases with ambiguous meanings. To rigorously address the Twitter entity link- ing problem, we propose a structural SVM algorithm for entity linking that jointly op- timizes mention detection and entity disam- biguation as a single end-to-end task. By com- bining structural learning and a variety of first- order, second-order, and context-sensitive fea- tures, our system is able to outperform exist- ing state-of-the art entity linking systems by 15% F1. 1 Introduction Microblogging services, such as Twitter and Face- book, are today capturing the largest volume ever recorded of fine-grained discussions spanning a huge breadth of topics, from the mundane to the his- toric. The micro-blogging service Twitter reports that it alone captures over 340M short messages, or tweets, per day.1 From such micro-blogging ser- vices’ data streams, researchers have reported min- ing insights about a variety of domains, from elec- tion results (Tumasjan et al., 2010) and democracy movements (Starbird and Palen, 2012) to health is- sues and disease spreading (Paul and Dredze, 2011; Sadilek et al., 2012), as well as tracking prod- uct feedback and sentiment (Asur and Huberman, 2010). A critical step in mining information from a micro-blogging service, such as Twitter, is the iden- tification of entities in tweets. In order to mine the relationship between drugs, symptoms and side- effects, or track the popularity of politicians or sen- timent about social issues, we must first be able to identify the topics and specific entities being dis- cussed. The challenge is that messages on micro- blogs are short, noisy, and informal texts with little context, and often contain phrases with ambiguous meanings. For example, “one day” may be either a set phrase or a reference to a movie. Given such difficulties, current mining and analysis of micro- blogs lists limits its application to certain domains with easy-to-recognize, unambiguous entities in or- der to avoid noise in the extraction results. We begin this paper with a thorough investigation of mention detection and entity disambiguation for social media, focused on the Twitter micro-blogging service. Mention detection is the task of extraction surface form candidates that can link to an entity in the domain of …

Content truncated to 500 words