Plato: A Selective Context Model for Entity Resolution

Plato: A Selective Context Model for Entity Resolution Nevena Lazic, Amarnag Subramanya, Michael Ringgaard, Fernando Pereira Google Inc., Mountain View CA 94043, USA {nevena, asubram, ringgaard, pereira}@google.com Abstract We present Plato, a probabilistic model for en- tity resolution that includes a novel approach for handling noisy or uninformative features, and supplements labeled training data derived from Wikipedia with a very large unlabeled text corpus. Training and inference in the pro- posed model can easily be distributed across many servers, allowing it to scale to over 107 entities. We evaluate Plato on three standard datasets for entity resolution. Our approach achieves the best results to-date on TAC KBP 2011 and is highly competitive on both the CoNLL 2003 and TAC KBP 2012 datasets. 1 Introduction Given a document collection and a knowledge base (KB) of entities, entity resolution, also known as en- tity disambiguation or entity linking, is the process of mapping each entity mention in a document to the corresponding entity record in the KB (Bunescu and Pasca, 2006; Cucerzan, 2007; Dredze et al., 2010; Hachey et al., 2013). Entity resolution is challenging because referring expressions are often ambiguous on their own and can only be disambiguated by their surrounding con- text. Consider the name Newcastle; it can refer to the city of Newcastle upon Tyne in UK, to the foot- ball (soccer for US readers) club Newcastle United F.C., to a popular beverage (Newcastle Brown Ale), or to several other entities. The ambiguity can only be resolved with appropriate context. Another com- plicating factor is that no KB is complete, and so a name in a document may refer to an entity that is missing from the KB. This problem is commonly called NIL detection. In this paper we present a probabilistic model for entity resolution. Our system, hereafter referred to as Plato, is designed to be resilient to irrelevant fea- tures and can be seen as a selective extension of the na¨ıve Bayes model. Specifically, we assume that most of the context features of a mention are irrele- vant to its disambiguation. This contrasts with the na¨ıve Bayes assumption that all features are gen- erated from a class-conditional distribution and are thus all relevant to the class assignment. Our empir- ical results support this modeling choice. We train Plato in a semi-supervised manner, starting with la- beled data derived from Wikipedia, and continuing with a very large unlabeled corpus of Web docu- ments. The use of unlabeled data enables us to ob- tain a better estimate of feature distributions and dis- cover new features that are not present in the (la- beled) training data. Plato scales up easily to very large KBs with millions of entities and includes NIL detection as a natural by-product of inference. We named our system after the Greek philosopher be- cause the system’s inference of real underlying en- tities from imperfect evidence reminds us of Plato’s Theory of Forms. Previous entity resolution studies (Milne and Wit- ten, 2008; Cucerzan, 2007; Ratinov et al., …

Content truncated to 500 words