Disambiguating Entities Referred by Web Endpoints using Tree Ensembles

Disambiguating Entities Referred by Web Endpoints using Tree Ensembles Gitansh Khirbat Jianzhong Qi Rui Zhang Department of Computing and Information Systems The University of Melbourne Australia gkhirbat@student.unimelb.edu.au {jianzhong.qi, rui.zhang}@unimelb.edu.au Abstract This paper describes system details and results of team “EOF” from the Univer- sity of Melbourne in the shared task of ALTA 2016, which addresses the use of cross document coreference resolution to determine whether two URLs refer to the same underlying entity. In our submis- sion, we develop a two stage system which first identifies the underlying entity for a given URL using entity-level features by ranking the entity mentions present in the crawled text with the help of logistic re- gression. This is followed by disambiguat- ing entities present in the given pair of URLs using a tree ensemble model to clas- sify if both URLs refer to the same under- lying entity. Our system achieved a final F1-score of 86.02% on the private leader- board1, which is the best score among all the participating systems. 1 Introduction The exponential expansion of the World Wide Web has resulted in a large data repository, the ma- jority of which is in the form of unstructured natu- ral language text containing ambiguous name enti- ties. A name entity mention may relate to multiple known entities. For example, the entity mention “New York” may refer to the city of New York or the movie New York which was released in 2009. Entity linking (EL) is the process of resolv- ing disambiguity between textual entity men- tions and the correct entity node in the knowl- edge base (KB). EL systems usually rely on semantic resources like Wikipedia as endpoints for disambiguation (Shen et al., 2015), however, 1https://inclass.kaggle.com/c/alta-2016- challenge/leaderboard Chisholm et al. (2016) provide a relaxed def- inition of a KB as any uniform resource loca- tor (URL) which reliably disambiguates linked mentions on the web (Chisholm et al., 2016a). This relaxed definition has motivated the shared task of ALTA 2016 (Chisholm et al., 2016b). The task organizers provided manually selected URL pairs from a heterogenous collection of web- sites including popular social networking websites like LinkedIn, Twitter, ResearchGate; knowledge bases like Wikipedia, IMDB and news websites like NDTV and Economic Times. The participants are asked to classify whether a given pair of URLs refer to the same underlying entity. For example, in Figure 1, URLs in the pair < UA1, UA2 > refer to the same entity “Barack Obama” whereas URLs in the pair < UB1, UB2 > refer to two different entities “Donald Trump” and “Ivanka Trump”. Figure 1: Example of URL pairs Considerable research has been done in the field of EL using existing KB like DBpedia (Auer et al., 2007), YAGO (Suchanek et al., 2007), Freebase (Bollacker et al., 2008) and KnowItAll (Etzioni et al., 2004). Wikipedia has proven to be a great resource in solving EL tasks (Cucerzan, 2007); (Milne and Witten, 2008) where dictionary-based techniques, contextual features and entity refer- ences have been used to train classifiers. …

Content truncated to 500 words