A Probabilistic Co-Bootstrapping Method for Entity Set Expansion

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2280–2290, Dublin, Ireland, August 23-29 2014. A Probabilistic Co-Bootstrapping Method for Entity Set Expansion Bei Shi, Zhengzhong Zhang Le Sun, Xianpei Han Institute of Software, Chinese Academy of Sciences, Beijing, China State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing, China {shibei, zhenzhong, sunle, xianpei}@nfs.iscas.ac.cn Abstract Entity Set Expansion (ESE) aims at automatically acquiring instances of a specific target category. Unfortunately, traditional ESE methods usually have the expansion boundary problem and the semantic drift problem. To resolve the above two problems, this paper proposes a probabilistic Co-Bootstrapping method, which can accurately determine the expansion boundary using both the positive and the discriminant negative instances, and resolve the semantic drift problem by effectively maintaining and refining the expansion boundary during bootstrapping iterations. Experimental results show that our method can achieve a competitive performance. 1 Introduction Entity Set Expansion (ESE) aims at automatically acquiring instances of a specific target category from text corpus or Web. For example, given the capital seeds {Rome, Beijing, Paris}, an ESE system should extract all other capitals from Web, such as Ottawa, Moscow and London. ESE system has been used in many applications, e.g., dictionary construction (Cohen and Sarawagi, 2004), word sense disambiguation (Pantel and Lin, 2002), query refinement (Hu et al., 2009), and query suggestion (Cao et al., 2008). Due to the limited supervision provided by ESE (in most cases only 3-5 seeds are given), traditional ESE systems usually employ bootstrapping methods (Cucchiarelli and Velardi, 2001; Etzioni et al., 2005; Pasca, 2007; Riloff and Jones, 1999; Wang and Cohen, 2008). That is, the entity set is iteratively expanded through a pattern generation step and an instance extraction step. Figure 1(a) demonstrates a simple bootstrapping process. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ Rome Beijing Paris Milan Tokyo Shanghai London * is the city of at the embassy in * * is the capital of at the hotel in * Chicago Berlin Pattern Generation Instance Extraction Rome Beijing Paris Milan Tokyo Shanghai London Sydney Boston * is the city of at the embassy in * * is the capital of to cities such as * at the hotel in * Chicago Nokia * the official web site New York Pattern Generation Instance Co-Extraction Negative Positive (a) (b) Figure 1: A demo of Bootstrapping (a) and Co-Bootstrapping (b) 2280 However, the traditional bootstrapping methods have two main drawbacks: 1) The expansion boundary problem. That is, using only positive seeds (i.e., some example entities from the category we want to expand), it is difficult to represent which entities we want to expand and which we don’t want. For example, starting from positive seeds {Rome, Beijing, Paris}, we can expand entities at many different levels, e.g., all capitals, all cities, or even all locations. And all these explanations are reasonable. 2) The semantic …

Content truncated to 500 words