Entity Linking with a Paraphrase Flavor

Entity Linking with a Paraphrase Flavor Maria Pershina, Yifan He, Ralph Grishman New York University 719 Broadway, 7th floor, New York, NY 10003, USA {pershina,yhe,grishman}@cs.nyu.edu Abstract The task of Named Entity Linking is to link entity mentions in the document to their correct entries in a knowledge base and to cluster NIL mentions. Ambiguous, misspelled, and incomplete entity mention names are the main challenges in the linking process. We propose a novel approach that combines two state-of-the-art models — for entity disambiguation and for paraphrase detection — to overcome these challenges. We consider name variations as paraphrases of the same entity mention and adopt a paraphrase model for this task. Our approach utilizes a graph-based disambiguation model based on Personalized Page Rank, and then refines and clusters its output using the paraphrase similarity between entity mention strings. It achieves a competitive performance of 80.5% in B3+F clustering score on diagnostic TAC EDL 2014 data. Keywords: disambiguation, linking, pagerank 1. Introduction Entity Linking (EL), such as the EL track at NIST Text Analysis Conference Knowledge Base Population (TAC- KBP), aims to link a given named entity mention from a source document to an existing Knowledge Base (KB) (Ji et al., 2014). In TAC-KBP, all entity mentions linked to the same KB entry are considered to form a cluster, while unlinked entities (NILs) have to be clustered as well. Linking raw entity mentions in a document to real world entities is useful on its own and serves as a valuable compo- nent in larger Knowledge Base Construction systems (May- field, 2014), e.g. the Cold Start track of TAC KBP pro- gram where the goal is to develop an automatic system to construct a KB from scratch. In the Wikification commu- nity (Bunescu and Pas¸ca, 2006) text mentions are linked to Wikipedia, a large and publicly available knowledge base. There are two paradigms to solve the EL problem: lo- cal, non-collective approaches for Entity Linking resolve one mention at a time relying on a context and local fea- tures, while collective approaches try to disambiguate the set of relevant mentions simultaneously assuming that en- tities appearing in the same document should be coherent (Cucerzan, 2007; Kulkarni et al., 2009; Ratinov et al., 2011; Hoffart et al., 2011; Alhelbawy and Gaizauskas, 2014; Per- shina et al., 2015b). Our approach in this paper is based on PPRSim, a state-of-the-art collective model for named entity disambiguation, utilizing the Personalized PageRank (PPR) algorithm (Pershina et al., 2015b). Nevertheless, we still try to capture the local similarity be- tween the entity mention and its candidates in our model. To measure the local similarity, we propose to use an approach which was proven effective for paraphrase de- tection. For this purpose we adopt the state-of-the-art ASOBEK paraphrase model (Eyecioglu and Keller, 2015). It was developed for paraphrase identification in Twitter and was ranked first among 19 teams on the Paraphrase In Twitter (PIT) 2015 task. It uses six simple character and word features and trains an SVM. This universal …

Content truncated to 500 words