Bidirectional Long Short-Term Memory Networks for Relation Classification

Bidirectional Long Short-Term Memory Networks for Relation Classification Shu Zhang1, Dequan Zheng2, Xinchen Hu2 and Ming Yang1 1 Fujitsu Research and Development Center, Beijing, China {zhangshu, yangming}@cn.fujitsu.com 2 School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China {dqzheng, xchu}@mtlab.hit.edu.cn Abstract Relation classification is an important se- mantic processing, which has achieved great attention in recent years. The main challenge is the fact that important infor- mation can appear at any position in the sentence. Therefore, we propose bidirec- tional long short-term memory networks (BLSTM) to model the sentence with complete, sequential information about all words. At the same time, we also use fea- tures derived from the lexical resources such as WordNet or NLP systems such as dependency parser and named entity rec- ognizers (NER). The experimental results on SemEval-2010 show that BLSTM- based method only with word embeddings as input features is sufficient to achieve state-of-the-art performance, and import- ing more features could further improve the performance. 1 Introduction The automatic classification of semantic relations is an important task, which could offer useful in- formation for many applications, such as question answering, information extraction, the construc- tion and completion of semantic or relational knowledge base. In this work, we focus on the classification of semantic relations between pairs of nominals (Hendrickx et al., 2010). Given a sentence S with annotated pairs of nominal e1 and e2, the task is to classify which of the following nine semantic re- lations holds between the nominals: Cause-Effect, Instrument-Agency, Product-Producer, Content- Container, Entity-Origin, Entity-Destination, Component-Whole, Member-Collection, Mes- sage-Topic, or Other if it does not belongs to any of the nine annotated relations. For example, News and commotion are con- nected in a Cause-Effect relation in the sentence “The news brought about a commotion in the of- fice.” In this instance, the relation between news and commotion could be inferred by the meaning of the two nominals and the context of “brought about” around them. Therefore, how to grasp and represent the lexical and context information are the key research points for semantic relation clas- sification. Supervised methods with carefully handcrafted features from lexical and semantic resources have achieved high performance (Hendrickx et al., 2010; Rink and Harabagiu, 2010). However, the selection of features and the effective integration of knowledge sources into relation classification seem to be difficult. Recently, deep neural networks has been ap- plied with the aim of reducing the number of handcrafted features, and getting effective fea- tures from lexical and sentence level (Socher et al., 2012; Zeng et al., 2014; Yu et al., 2014). Different from previous work, we propose bi- directional long short-term memory networks (BLSTM) to solve the relation classification. For every word in a given sentence, BLSTM has com- plete, sequential information about all words be- fore and after it. Long distance relationship may be solved in some extent in this networks. At the same time, we also use features derived from the lexical resources such as WordNet or NLP tools such as dependency …

Content truncated to 500 words