META-DARE: Monitoring the Minimally Supervised ML of Relation Extraction Rules

Proceedings of Recent Advances in Natural Language Processing, pages 378–384, Hissar, Bulgaria, 12-14 September 2011. META-DARE: Monitoring the Minimally Supervised ML of Relation Extraction Rules Hong Li Feiyu Xu German Research Center for Artificial Intelligence (DFKI), LT-Lab Alt-Moabit 91c, 10559 Berlin, Germany {lihong,feiyu,uszkoreit}@dfki.de http://www.dfki.de/lt/ Hans Uszkoreit Abstract This paper demonstrates a web-based online system, called META-DARE1. META-DARE is built to assist researchers to obtain insights into seed-based mini- mally supervised machine learning for re- lation extraction. META-DARE allows re- searchers and students to conduct experi- ments with an existing machine learning system called DARE (Xu et al., 2007). Users can run their own learning experi- ments by constructing initial seed exam- ples and can monitor the learning pro- cess in a very detailed way, namely, via interacting with each node in the learn- ing graph and viewing its content. Fur- thermore, users can study the learned re- lation extraction rules and their applica- tions. META-DARE is also an analysis tool which gives an overview of the whole learning process: the number of iterations, the input and output behaviors of each iter- ation, and the general performance of the extracted instances and their distributions. Moreover, META-DARE provides a very convenient user interface for visualization of the learning graph, the learned rules and the system performance profile. 1 Introduction Seed-based minimally supervised machine learn- ing within a bootstrapping framework has been widely applied to various information extraction tasks (e.g., (Hearst, 1992; Riloff, 1996; Brin, 1998; Agichtein and Gravano, 2000; Sudo et al., 2003; Greenwood and Stevenson, 2006; Blohm and Cimiano, 2007)). The power of this approach is that it needs only a small set of examples of either patterns or relation instances and can learn 1http://dare.dfki.de/ and discover many useful extraction rules and re- lation instances from unannotated texts. Within this framework, Xu et al. (2007) develop a learn- ing approach, called DARE, which learns relation extraction rules for dealing with relations of var- ious complexity by utilizing some relation exam- ples as semantic seed in the initialization and has achieved very promising results for the extraction of complex relations. In the recent years, more and more researchers are interested in understanding the underlying process behind this approach and attempt to identify relevant learning parameters to improve the system performance. Xu (2007) investigates the role of the seed se- lection in connection with the data properties in a careful way with our DARE system. Xu (2007) and Li et al. (2011) describe the applications of DARE system in different domains for different relation extraction tyes, for example, the Nobel- Prize-Winning event, management succession re- lations defined in MUC-6, marriage relationship, etc. Uszkoreit et al. (2009) describe a further em- pirical analysis of the seed construction and its influence on the learning performance and show that size, arity and distinctiveness of the seed ex- amples play various important roles for the learn- ing performance. Thus, the system demonstrated here, called META-DARE, serves as a monitoring and analysis system for conducting various exper- iments …

Content truncated to 500 words