Entity Linking on Microblogs with Spatial and Temporal Signals

Entity Linking on Microblogs with Spatial and Temporal Signals Yuan Fang ∗ † University of Illinois at Urbana-Champaign 201 N. Goodwin Avenue Urbana, IL 61801, USA fang2@illinois.edu Ming-Wei Chang Microsoft Research 1 Microsoft Way Redmond, WA 98052, USA minchang@microsoft.com Abstract Microblogs present an excellent opportunity for monitoring and analyzing world happen- ings. Given that words are often ambiguous, entity linking becomes a crucial step towards understanding microblogs. In this paper, we re-examine the problem of entity linking on microblogs. We first observe that spatiotem- poral (i.e., spatial and temporal) signals play a key role, but they are not utilized in exist- ing approaches. Thus, we propose a novel en- tity linking framework that incorporates spa- tiotemporal signals through a weakly super- vised process. Using entity annotations1 on real-world data, our experiments show that the spatiotemporal model improves F1 by more than 10 points over existing systems. Finally, we present a qualitative study to visualize the effectiveness of our approach. 1 Introduction Microblogging services provide an immense plat- form for intelligence gathering, such as market re- search (Asur and Huberman, 2010), disaster mon- itoring (Sadilek et al., 2012) and political analy- sis (Tumasjan et al., 2010). Extracting entities from microblogs is an essential step for many such ap- plications. Suppose that a marketing firm is inter- ested in the sentiment about some product on Twit- ter. However, any sentiment analysis is potentially ∗Work done during an internship at Microsoft Research. †Also affiliated with Agency for Science, Technology and Research, 1 Fusionopolis Way, Singapore 138632. 1Can be downloaded at http://research.microsoft.com/en- us/downloads/84ac9d88-c353-4059-97a4-87d129db0464/. misleading if we cannot correctly retrieve the tweets mentioning the target product. To retrieve tweets for a given entity (e.g., a prod- uct or an organization), a straightforward approach is to formulate a keyword query. However, simple keyword matching is largely ineffective, since key- words are often ambiguous. For instance, “spurs” can refer to two distinct sports teams (San Antonio Spurs, which is a basketball team in the US, and Tot- tenham Hotspur F.C., which is a soccer team in the UK), besides being a non-entity verb or noun. Thus, retrieving tweets via keyword queries inevitably mixes different entities. Given the ambiguity of keywords, in this paper, we study the task of entity linking (Bunescu and Pasca, 2006) on microblogs. As input, we are given a short message (e.g., a tweet) and an entity database (e.g., Wikipedia where each article is an entity). As output, we map every surface form (e.g., “spurs”) in the message to an entity (e.g., San Antonio Spurs) or to ∅ (i.e., a non-entity). This task is particularly challenging for microblogs due to their short, noisy and colloquial nature. Fortunately, we also observe two new opportunities, which are often missing in traditional data. First, microblogs usually embed rich meta-data, most notably spatiotemporal (i.e., spatial and tem- poral) signals. Specifically, all tweets are associ- ated with a timestamp, and many can be mapped to a location. We observe that entity priors often change across time and location. Consider …

Content truncated to 500 words