Improved Pattern Learning for Bootstrapped Entity Extraction

Proceedings of the Eighteenth Conference on Computational Language Learning, pages 98–108, Baltimore, Maryland USA, June 26-27 2014. c⃝2014 Association for Computational Linguistics Improved Pattern Learning for Bootstrapped Entity Extraction Sonal Gupta Christopher D. Manning Department of Computer Science Stanford University {sonal, manning}@cs.stanford.edu Abstract Bootstrapped pattern learning for entity extraction usually starts with seed entities and iteratively learns patterns and entities from unlabeled text. Patterns are scored by their ability to extract more positive en- tities and less negative entities. A prob- lem is that due to the lack of labeled data, unlabeled entities are either assumed to be negative or are ignored by the existing pat- tern scoring measures. In this paper, we improve pattern scoring by predicting the labels of unlabeled entities. We use var- ious unsupervised features based on con- trasting domain-specific and general text, and exploiting distributional similarity and edit distances to learned entities. Our system outperforms existing pattern scor- ing algorithms for extracting drug-and- treatment entities from four medical fo- rums. 1 Introduction This paper considers the problem of building ef- fective entity extractors for custom entity types from specialized domain corpora. We approach the problem by learning rules bootstrapped us- ing seed sets of entities. Though entity extrac- tion using machine learning is common in aca- demic research, rule-based systems dominate in commercial use (Chiticariu et al., 2013), mainly because rules are effective, interpretable, and are easy to customize by non-experts to cope with er- rors. They also have been shown to perform bet- ter than state-of-the-art machine learning methods on some specialized domains (Nallapati and Man- ning, 2008; Gupta and Manning, 2014a). In ad- dition, building supervised machine learning sys- tems for a reasonably large domain-specific cor- pus would require hand-labeling sufficient data to Seed dictionary for class ‘animals’: {dog} Text: I own a cat named Fluffy. I run with my pet dog. I also nap with my pet cat. I own a car. Pattern 1: my pet X Extractions = positive : {dog}, unlabeled : {cat} Pattern 2: own a X Extractions = positive : {dog}, unlabeled : {car} Figure 1: An example pattern learning system for the class ‘animals’ from the text. Pattern 1 and 2 are candidate patterns. Text matched with the pat- terns is shown in italics and the extracted entities are shown in bold. train a model, which can be costly and time con- suming. Bootstrapped machine-learned rules can make extraction easier and more efficient on such a corpus. In a bootstrapped rule-based entity learning system, seed dictionaries and/or patterns provide weak supervision to label data. The system itera- tively learns new entities belonging to a specific class from unlabeled text (Riloff, 1996; Collins and Singer, 1999). Rules are typically defined by creating patterns around the entities, such as lexico-syntactic surface word patterns (Hearst, 1992) and dependency tree patterns (Yangarber et al., 2000). Patterns are scored by their abil- ity to extract more positive entities and less neg- ative entities. Top ranked patterns are used to extract candidate entities …

Content truncated to 500 words