Entity Extraction via Ensemble Semantics

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 238–247, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Entity Extraction via Ensemble Semantics Marco Pennacchiotti Yahoo! Labs Sunnyvale, CA, 94089 pennac@yahoo-inc.com Patrick Pantel Yahoo! Labs Sunnyvale, CA, 94089 ppantel@yahoo-inc.com Abstract Combining information extraction sys- tems yields significantly higher quality re- sources than each system in isolation. In this paper, we generalize such a mixing of sources and features in a framework called Ensemble Semantics. We show very large gains in entity extraction by combining state-of-the-art distributional and pattern- based systems with a large set of fea- tures from a webcrawl, query logs, and Wikipedia. Experimental results on a web- scale extraction of actors, athletes and mu- sicians show significantly higher mean av- erage precision scores (29% gain) com- pared with the current state of the art. 1 Introduction Mounting evidence shows that combining infor- mation sources and information extraction algo- rithms leads to improvements in several tasks such as fact extraction (Pas¸ca et al., 2006), open- domain IE (Talukdar et al., 2008), and entailment rule acquisition (Mirkin et al., 2006). In this paper, we show large gains in entity extraction by com- bining state-of-the-art distributional and pattern- based systems with a large set of features from a 600 million document webcrawl, one year of query logs, and a snapshot of Wikipedia. Further, we generalize such a mixing of sources and fea- tures in a framework called Ensemble Semantics. Distributional and pattern-based extraction al- gorithms capture aspects of paradigmatic and syn- tagmatic dimensions of semantics, respectively, and are believed to be quite complementary. Pas¸ca et al. (2006) showed that filtering facts, extracted by a pattern-based system, according to their ar- guments’ distributional similarity with seed facts yielded large precision gains. Mirkin et al. (2006) showed similar gains on the task of acquiring lex- ical entailment rules by exploring a supervised combination of distributional and pattern-based al- gorithms using an ML-based SVM classifier. This paper builds on the above work, by study- ing the impact of various sources of features exter- nal to distributional and pattern-based algorithms, on the task of entity extraction. Mirkin et al.’s re- sults are corroborated on this task and large and significant gains over this baseline are obtained by incorporating 402 features from a webcrawl, query logs and Wikipedia. We extracted candidate entities for the classes Actors, Athletes and Mu- sicians from a webcrawl using a variant of Pas¸ca et al.’s (2006) pattern-based engine and Pantel et al.’s (2009) distributional extraction system. gradient boosted decision tree is used to learn a re-A gression function over the feature space for rank- ing the candidate entities. Experimental results show 29% gains (19% nominal) in mean average precision over Mirkin et al.’s method and 34% gains (22% nominal) in mean average precision over an unsupervised baseline similar to Pas¸ca et al.’s method. Below we summarize the contribu- tions of this paper: • We explore the hypothesis that although dis- tributional and pattern-based algorithms are complementary, they …

Content truncated to 500 words