Minimally Supervised Rule Learning for the Extraction of Biographic Information from Various Social Domains

Proceedings of Recent Advances in Natural Language Processing, pages 17–24, Hissar, Bulgaria, 12-14 September 2011. Minimally Supervised Rule Learning for the Extraction of Biographic Information from Various Social Domains 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 investigates the application of an existing seed-based minimally super- vised learning algorithm to different so- cial domains exhibiting different proper- ties of the available data. A systematic analysis studies the respective data prop- erties of the three domains including the distribution of the semantic arguments and their combinations. The experimental re- sults confirm that data properties have a strong influence on the performance of the learning system. The main results are in- sights about: (i) the effects of data proper- ties such as redundancy and frequency of argument mentions on coverage and preci- sion (ii) the positive effects of negative ex- amples if used effectively (iii) the different effects of negative examples depending on the domain data properties and (iv) the po- tential of reusing rules from one domain for improving the relation extraction per- formance in another domain. 1 Introduction Domain adaptation is very important for informa- tion extraction (IE) systems. IE systems in the real world are often required to work for new do- mains and new tasks within a limited adaptation or tuning time. Thus, automatic learning of rela- tion extraction rules for a new domain or a new task has been established as a relevant subarea in IE research and development (Muslea, 1999; Tsu- jii, 2000; Uszkoreit, 2011), in particular for min- imally supervised or semi-supervised bootstrap- ping approaches (e.g., (Brin, 1998; Agichtein and Gravano, 2000; Yangarber, 2001; Sudo et al., 2003; Bunescu and Mooney, 2005; McDonald et al., 2005; Greenwood and Stevenson, 2006; Jones, 2005; Xu et al., 2007; Xu, 2007; Kozareva and Hovy, 2010a; Kozareva and Hovy, 2010b)). The advantage of the minimally supervised approaches for IE rule learning is that only initial seed knowl- edge is needed. Therefore the adaptation might be limited to substituting the seed examples. How- ever, different domains/corpora exhibit rather dif- ferent properties of their learning/extraction data with respect to the learning algorithm. Depending on the domain, the need for improving precision by utilizing negative examples may differ. An im- portant research goal is the exploitation of more benign domains for improving extraction in less suitable domains. Xu et al. (2007) and Xu (2007) present a min- imally supervised learning system for relation ex- traction, initialized by a so-called semantic seed, i.e., examples of the target relations. We dub our system DARE for Domain Adaptive Relation Ex- traction. The system supports the domain adapta- tion with a compositional rule representation and a bottom-up rule discovery strategy. In this way, DARE can handle target relations of various com- plexities and arities. Relying on a few examples of a target relation as semantic seed dispenses with the costly acquisition of domain knowledge through experts or specialized resources. In …

Content truncated to 500 words