Cross-Document Co-Reference Resolution using Sample-Based Clustering with Knowledge Enrichment

Cross-Document Co-Reference Resolution using Sample-Based Clustering with Knowledge Enrichment Sourav Dutta Max Planck Institute for Informatics Saarbr¨ucken, Germany sdutta@mpi-inf.mpg.de Gerhard Weikum Max Planck Institute for Informatics Saarbr¨ucken, Germany weikum@mpi-inf.mpg.de Abstract Identifying and linking named entities across information sources is the basis of knowledge acquisition and at the heart of Web search, rec- ommendations, and analytics. An important problem in this context is cross-document co- reference resolution (CCR): computing equiv- alence classes of textual mentions denoting the same entity, within and across documents. Prior methods employ ranking, clustering, or probabilistic graphical models using syntactic features and distant features from knowledge bases. However, these methods exhibit limita- tions regarding run-time and robustness. This paper presents the CROCS framework for unsupervised CCR, improving the state of the art in two ways. First, we extend the way knowledge bases are harnessed, by con- structing a notion of semantic summaries for intra-document co-reference chains using co- occurring entity mentions belonging to differ- ent chains. Second, we reduce the computa- tional cost by a new algorithm that embeds sample-based bisection, using spectral clus- tering or graph partitioning, in a hierarchi- cal clustering process. This allows scaling up CCR to large corpora. Experiments with three datasets show significant gains in output qual- ity, compared to the best prior methods, and the run-time efficiency of CROCS. 1 Introduction 1.1 Motivation and Problem Statement We are witnessing another revolution in Web search, user recommendations, and data analytics: tran- sitioning from documents and keywords to data, knowledge, and entities. Examples of this mega- trend are the Google Knowledge Graph and its ap- plications, and the IBM Watson technology for deep question answering. To a large extent, these ad- vances have been enabled by the construction of huge knowledge bases (KB’s) such as DBpedia, Yago, or Freebase; the latter forming the core of the Knowledge Graph. Such semantic resources provide huge collections of entities: people, places, compa- nies, celebrities, movies, etc., along with rich knowl- edge about their properties and relationships. Perhaps the most important value-adding com- ponent in this setting is the recognition and dis- ambiguation of named entities in Web and user contents. Named Entity Disambiguation (NED) (see, e.g., (Cucerzan, 2007; Milne & Witten, 2008; Cornolti et al., 2013)) maps a mention string (e.g., a person name like “Bolt” or a noun phrase like “light- ning bolt”) onto its proper entity if present in a KB (e.g., the sprinter Usain Bolt). A related but different task of co-reference reso- lution (CR) (see, e.g., (Haghighi & Klein, 2009; Ng, 2010; Lee et al., 2013)) identifies all mentions in a given text that refer to the same entity, including anaphoras such as “the president’s wife”, “the first lady”, or “she”. This task when extended to process an entire corpus is then known as cross-document co-reference resolution (CCR) (Singh et al., 2011). It takes as input a set of documents with entity men- tions, and computes as output a set of equivalence classes over the entity mentions. This does not in- volve …

Content truncated to 500 words