Chain Based RNN for Relation Classification

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1244–1249, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Chain based RNN for Relation Classification Javid Ebrahimi and Dejing Dou Department of Computer and Information Science, University of Oregon Eugene, Oregon 97403, USA {javid,dou}@cs.uoregon.edu Abstract We present a novel approach for relation clas- sification, using a recursive neural network (RNN), based on the shortest path between two entities in a dependency graph. Previ- ous works on RNN are based on constituency- based parsing because phrasal nodes in a parse tree can capture compositionality in a sen- tence. Compared with constituency-based parse trees, dependency graphs can represent relations more compactly. This is particu- larly important in sentences with distant en- tities, where the parse tree spans words that are not relevant to the relation. In such cases RNN cannot be trained effectively in a timely manner. However, due to the lack of phrasal nodes in dependency graphs, applica- tion of RNN is not straightforward. In order to tackle this problem, we utilize dependency constituent units called chains. Our experi- ments on two relation classification datasets show that Chain based RNN provides a shal- lower network, which performs considerably faster and achieves better classification results. 1 Introduction Relation extraction is the task of finding relations between entities in text, which is useful for sev- eral tasks such as information extraction, summa- rization, and question answering (Wu and Weld, 2010). For instance, in the sentence: those “cancers” were caused by radiation “exposures,” the two enti- ties have a cause-effect relation. As reported in de- tail (SaraWagi, 2008), one approach to the problem involves supervised methods where the models rely on lexical, syntactic, and semantic features to clas- sify relations between pairs of entities. The down- side of this approach is that one has to retrain the model for other domains with different target rela- tions. Thus it is not scalable to the web, where thou- sands of (previously-unseen) relations exist (Banko et al., 2007). To address this problem, Open Infor- mation Extraction is proposed, which does not re- quire supervision. In these systems (Banko et al., 2007; Mausam et al., 2012), patterns based on lex- ical, syntactic, POS, and dependency features are extracted. While these patterns give good preci- sion, they suffer from low recall (Banko and Etzioni, 2008). This is because they fail to extract patterns which have not been pre-specified, and thereby are unable to generalize. Recursive Neural Network (RNN) has proven to be highly successful in capturing semantic compo- sitionality in text and has improved the results of several Natural Language Processing tasks (Socher et al., 2012; Socher et al., 2013). Previous ap- plications of Recursive Neural Networks (RNN) to supervised relation extraction (Socher et al., 2012; Hashimoto et al., 2013; Khashabi, 2013) are based on constituency-based parsers. These RNNs may span words that do not contribute to the relation. We investigate the incorporation of dependency parsing into RNN …

Content truncated to 500 words