Improving Information Extraction by Acquiring External Evidence with Reinforcement Learning

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2355–2365, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Improving Information Extraction by Acquiring External Evidence with Reinforcement Learning Karthik Narasimhan CSAIL, MIT karthikn@mit.edu Adam Yala CSAIL, MIT adamyala@mit.edu Regina Barzilay CSAIL, MIT regina@csail.mit.edu Abstract Most successful information extraction sys- tems operate with access to a large collec- tion of documents. In this work, we explore the task of acquiring and incorporating exter- nal evidence to improve extraction accuracy in domains where the amount of training data is scarce. This process entails issuing search queries, extraction from new sources and rec- onciliation of extracted values, which are re- peated until sufficient evidence is collected. We approach the problem using a reinforce- ment learning framework where our model learns to select optimal actions based on con- textual information. We employ a deep Q- network, trained to optimize a reward func- tion that reflects extraction accuracy while pe- nalizing extra effort. Our experiments on two databases – of shooting incidents, and food adulteration cases – demonstrate that our system significantly outperforms traditional extractors and a competitive meta-classifier baseline.1 1 Introduction In many realistic domains, information extraction (IE) systems require exceedingly large amounts of annotated data to deliver high performance. In- creases in training data size enable models to han- dle robustly the multitude of linguistic expressions that convey the same semantic relation. Consider, for instance, an IE system that aims to identify en- tities such as the perpetrator and the number of vic- 1Code is available at http://people.csail.mit. edu/karthikn/rl-ie/ ShooterName: Scott Westerhuis NumKilled: 6 A couple and four children found dead in their burning South Dakota home had been shot in an ...apparent murder-suicide, officials said Monday. Scott Westerhuis’s cause of death was "shotgun wound with manner of death as suspected sui- cide," it added in a statement. Figure 1: Sample news article on a shooting case. Note how the article contains both the name of the shooter and the number of people killed but both pieces of informa- tion require complex extraction schemes. tims in a shooting incident (Figure 1). The docu- ment does not explicitly mention the shooter (Scott Westerhuis), but instead refers to him as a suicide victim. Extraction of the number of fatally shot vic- tims is similarly difficult, as the system needs to in- fer that "A couple and four children" means six peo- ple. Even a large annotated training set may not pro- vide sufficient coverage to capture such challenging cases. In this paper, we explore an alternative approach for boosting extraction accuracy, when a large train- ing corpus is not available. Instead, the proposed method utilizes external information sources to re- solve ambiguities inherent in text interpretation. Specifically, our strategy is to find other documents that contain the information sought, expressed in a form that a basic extractor can "understand". For instance, Figure 2 shows two other articles describ- ing the same event, wherein the entities of interest 2355 The …

Content truncated to 500 words