Learning Field Compatibilities to Extract Database Records from Unstructured Text

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 603–611, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Learning Field Compatibilities to Extract Database Records from Unstructured Text Michael Wick, Aron Culotta and Andrew McCallum Department of Computer Science University of Massachusetts Amherst, MA 01003 {mwick, culotta, mccallum}@cs.umass.edu Abstract Named-entity recognition systems extract entities such as people, organizations, and locations from unstructured text. Rather than extract these mentions in isolation, this paper presents a record extraction sys- tem that assembles mentions into records (i.e. database tuples). We construct a probabilistic model of the compatibility between field values, then employ graph partitioning algorithms to cluster fields into cohesive records. We also investigate compatibility functions over sets of fields, rather than simply pairs of fields, to ex- amine how higher representational power can impact performance. We apply our techniques to the task of extracting contact records from faculty and student home- pages, demonstrating a 53% error reduc- tion over baseline approaches. 1 Introduction Information extraction (IE) algorithms populate a database with facts discovered from unstructured text. This database is often used by higher-level tasks such as question answering or knowledge discovery. The richer the structure of the database, the more useful it is to higher-level tasks. A common IE task is named-entity recognition (NER), the problem of locating mentions of en- tities in text, such as people, places, and organi- zations. NER techniques range from regular ex- pressions to finite-state sequence models (Bikel et al., 1999; Grishman, 1997; Sutton and McCallum, 2006). NER can be viewed as method of populat- ing a database with single-tuple records, e.g. PER- SON=Cecil Conner or ORGANIZATION= IBM. We can add richer structure to these single-tuple records by extracting the associations among en- tities. For example, we can populate multi-field records such as a contact record [PERSON=Steve Jobs, JOBTITLE = CEO, COMPANY = Apple, CITY = Cupertino, STATE = CA]. The relational information in these types of records presents a greater opportunity for text analysis. The task of associating together entities is of- ten framed as a binary relation extraction task: Given a pair of entities, label the relation be- tween them (e.g. Steve Jobs LOCATED-IN Cuper- tino). Common approaches to relation extraction include pattern matching (Brin, 1998; Agichtein and Gravano, 2000) and classification (Zelenko et al., 2003; Kambhatla, 2004). However, binary relation extraction alone is not well-suited for the contact record example above, which requires associating together many fields into one record. We refer to this task of piecing together many fields into a single record as record extraction. Consider the task of extracting contact records from personal homepages. An NER system may label all mentions of cities, people, organizations, phone numbers, job titles, etc. on a page, from both semi-structured an unstructured text. Even with a highly accurate NER system, it is not obvi- ous which fields belong to the same record. For example, a single document could contain five names, three phone numbers and only one email. Additionally, …

Content truncated to 500 words