Toward Socially-Infused Information Extraction: Embedding Authors, Mentions, and Entities

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1452–1461, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Toward Socially-Infused Information Extraction: Embedding Authors, Mentions, and Entities Yi Yang Georgia Institute of Technology Atlanta, GA 30308, USA yiyang@gatech.edu Ming-Wei Chang Microsoft Research Redmond, WA 98052, USA minchang@microsoft.com Jacob Eisenstein Georgia Institute of Technology Atlanta, GA 30308, USA jacobe@gatech.edu Abstract Entity linking is the task of identifying men- tions of entities in text, and linking them to entries in a knowledge base. This task is espe- cially difficult in microblogs, as there is little additional text to provide disambiguating con- text; rather, authors rely on an implicit com- mon ground of shared knowledge with their readers. In this paper, we attempt to cap- ture some of this implicit context by exploit- ing the social network structure in microblogs. We build on the theory of homophily, which implies that socially linked individuals share interests, and are therefore likely to mention the same sorts of entities. We implement this idea by encoding authors, mentions, and en- tities in a continuous vector space, which is constructed so that socially-connected authors have similar vector representations. These vectors are incorporated into a neural struc- tured prediction model, which captures struc- tural constraints that are inherent in the entity linking task. Together, these design decisions yield F1 improvements of 1%-5% on bench- mark datasets, as compared to the previous state-of-the-art. 1 Introduction Entity linking on short texts (e.g., Twitter messages) is of increasing interest, as it is an essential step for many downstream applications, such as market re- search (Asur and Huberman, 2010), topic detection and tracking (Mathioudakis and Koudas, 2010), and question answering (Yih et al., 2015). Tweet entity linking is a particularly difficult problem, because Figure 1: Illustration on leveraging social relations for entity disambiguation. Socially connected users u1 and u2 tend to talk about similar entities (baseball in the example). the short context around an entity mention is often insufficient for entity disambiguation. For example, as shown in Figure 1, the entity mention ‘Giants’ in tweet t1 can refer to the NFL football team New York Giants or the MLB baseball team San Fran- cisco Giants. In this example, it is impossible to disambiguate between these entities solely based on the individual text message. We propose to overcome the difficulty and im- prove the entity disambiguation capability of the entity linking system by employing social network structures. The sociological theory of homophily asserts that socially connected individuals are more likely to have similar behaviors or share similar in- terests (McPherson et al., 2001). This property has been used to improve many natural language pro- cessing tasks such as sentiment analysis (Tan et al., 2011; Yang and Eisenstein, 2015), topic classifica- tion (Hovy, 2015) and user attribute inference (Li et al., 2015). We assume Twitter users will have simi- lar interests in real world entities to their near neigh- bors — an assumption of entity homophily — which 1452 …

Content truncated to 500 words