Joint Learning of Local and Global Features for Entity Linking via Neural Networks

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2310–2320, Osaka, Japan, December 11-17 2016. Joint Learning of Local and Global Features for Entity Linking via Neural Networks Thien Huu Nguyen†, Nicolas Fauceglia#, Mariano Rodriguez Muro§, Oktie Hassanzadeh§, Alfio Massimiliano Gliozzo§ and Mohammad Sadoghi‡ † New York University, # Carnegie Mellon University, ‡ Purdue University § IBM T.J. Watson Research Center, Yorktown Heights, New York, USA thien@cs.nyu.edu,fauceglia@cs.cmu.edu {mrodrig,hassanzadeh,gliozzo}.us.ibm.com,msadoghi@purdue.edu Abstract Previous studies have highlighted the necessity for entity linking systems to capture the local entity-mention similarities and the global topical coherence. We introduce a novel framework based on convolutional neural networks and recurrent neural networks to simultaneously model the local and global features for entity linking. The proposed model benefits from the capacity of convolutional neural networks to induce the underlying representations for local contexts and the advantage of recurrent neural networks to adaptively compress variable length sequences of predictions for global constraints. Our evaluation on multiple datasets demonstrates the effec- tiveness of the model and yields the state-of-the-art performance on such datasets. In addition, we examine the entity linking systems on the domain adaptation setting that further demonstrates the cross-domain robustness of the proposed model. 1 Introduction We address the problem of entity linking (EL): mapping entity mentions in documents to their correct entries (called target entities) in some existing knowledge bases (KB) like Wikipedia. For instance, in the sentence “Liverpool suffered an upset first home league defeat of the season.”, an entity linking system should be able to identify the entity mention “Liverpool” as a football club rather than a city in England in the knowledge bases. This is a challenging problem of natural language processing, as the same entity might be presented in various names, and the same entity mention string might refer to different entities in different contexts. Entity linking is a fundamental task for other applications such as information extraction, knowledge base construction etc. In order to tackle the ambiguity in EL, previous studies have first generated a set of target entities in the knowledge bases as the referent candidates for each entity mention in the documents, and then solved a ranking problem to disambiguate the entity mention. The key challenge in this paradigm is the ranking model that computes the relevance of each target entity candidate to the corresponding entity mention using the available context information in both the documents and the knowledge bases. The early approach for the ranking problem in EL has resolved the entity mentions in documents independently (the local approach), utilizing various discrete and hand-designed features/heuristics to measure the local mention-to-entity relatedness for ranking. These features are often specific to each entity mention and candidate entity, covering a wide range of linguistic and/or structured representa- tions such as lexical and part-of-speech tags of context words, dependency paths, topical features, KB infoboxes (Bunescu and Pasca, 2006; Mendes et al., 2011; Cassidy et al., 2011; Ji and Grishman, 2011; Shen et al., 2014) etc. Although the local approach …

Content truncated to 500 words