Exploiting Structure in Representation of Named Entities using Active Learning

Proceedings of the 27th International Conference on Computational Linguistics, pages 687–699 Santa Fe, New Mexico, USA, August 20-26, 2018. 687 Exploiting Structure in Representation of Named Entities using Active Learning Nikita Bhutani University of Michigan, Ann Arbor nbhutani@umich.edu Kun Qian IBM Research - Almaden qian.kun@ibm.com Yunyao Li IBM Research - Almaden yunyaoli@us.ibm.com H. V. Jagadish University of Michigan, Ann Arbor jag@umich.edu Mauricio A. Hernandez IBM Research - Almaden mahernan@us.ibm.com Mitesh Vasa IBM Research - Almaden mitesh.vasa@us.ibm.com Abstract Fundamental to several knowledge-centric applications is the need to identify named entities from their textual mentions. However, entities lack a unique representation and their mentions can differ greatly. These variations arise in complex ways that cannot be captured using textual similarity metrics. However, entities have underlying structures, typically shared by entities of the same entity type, that can help reason over their name variations. Discovering, learning and manipulating these structures typically requires high manual effort in the form of large amounts of labeled training data and handwritten transformation programs. In this work, we propose an active-learning based framework that drastically reduces the labeled data required to learn the structures of entities. We show that programs for mapping entity mentions to their structures can be automatically generated using human-comprehensible labels. Our experiments show that our framework consistently outperforms both handwritten programs and supervised learning models. We also demonstrate the utility of our framework in relation extraction and entity resolution tasks. 1 Introduction Named entities are atomic objects of reference and reasoning in many cognitive applications and knowledge-centric services like deep question answering, text summarization and analytics. A real-world entity may have a great variety of representations (Gal´arraga et al., 2014; Nakashole et al., 2011). For example, University of California, Santa Cruz could have different string representations or name variations: UCSC, UC Santa Cruz, UC–Santa Cruz. Determining if two representations refer to the same entity is an important primitive in entity resolution and entity linking algorithms that drive these knowledge-centric applications (Shen et al., 2015; Arasu and Kaushik, 2009). Unifying entity representations has been widely studied in record linkage (Christen, 2012), deduplica- tion (Elmagarmid et al., 2007) and reference matching (McCallum et al., 2000). Typically, duplicates are identified using the attributes and/or the contextual information of entities (Zhang et al., 2010; Han et al., 2011; Shen et al., 2015). The string representation or mention of an entity forms key evidence: similar mentions likely refer to the same entity. Although deemed as crucial (Dredze et al., 2010), variations in mentions are typically handled using textual similarity like edit distance and cosine similarity (Zheng et al., 2010; Lehmann et al., 2010; Liu et al., 2013), which can be misleading. Example. Consider the mentions: (a) General Electric Corporation, (b) General Electric China Corporation, (c) GE Corp. Mentions (a) and (b) are textually similar, differing in just one token. However, they refer to different entities, owing to the location detail ‘China’. Conversely, textually dissimilar mentions (a) and (c) refer to the same entity. An entity mention is not merely …

Content truncated to 500 words