Chinese Relation Classification using Long Short Term Memory Networks

Chinese Relation Classification using Long Short Term Memory Networks Linrui Zhang, Dan Moldovan The University of Texas at Dallas 800 West Campbell Road ; MS EC31, Richardson, TX 75080 U.S.A linrui.zhang@utdallas.edu, moldovan@hlt.utdallas.edu Abstract Relation classification is the task to predict semantic relations between pairs of entities in a given text. In this paper, a novel Long Short Term Memory Network (LSTM)-based approach is proposed to extract relations between entities in Chinese text. The shortest dependency path (SDP) between two entities, together with the various selected features in the path, are first extracted, and then used as input of an LSTM model to predict the relation between them. The performance of the system was evaluated on the ACE 2005 Multilingual Training Corpus (Walker et al., 2006), and achieved a state-of-the-art F-measure of 87.87% on six general type relations and 83.40% on eighteen subtype relations in this corpus. Keywords: Long Short Term Memory Networks, Shortest Dependency Path, Chinese Relation Classification 1. Introduction The task of relation classification is to predict semantic re- lations between pairs of entities. Formally, the goal is to predict the semantic relations between a head entity eh and a tail entity et from a given sentence. For example, in the phrase “ÄW¯;ß Russian President”, the relations be- tween “Russia (ORG)” and “President (PER)” are ORG- AFF which indicate that the “President” is affiliated to an or- ganization “Russia” , in ACE annotation ORG-AFF(Russia, President), and to be more specific, “President” is the em- ployee of “Russia”, EMPLOYMENT(Russia, President). In ACE corpus 2005, six general relations such as ORG-AFF and PART-WHOLE, and eighteen subtype relations such as OWNERSHIP, LOCATED, and EMPLOYMENT are defined. The feature-based, kernel-based and deep learning-based methods are the most popular models for relation extrac- tion in the literature. The basic idea of the feature-based approach is to treat relation extraction as a classification problem. Different kinds of features are extracted from text and then fed into a classifier. Such work includes (Kamb- hatla, 2004), (GuoDong et al., 2005) and (Moldovan and Blanco, 2012). However, these approaches suffer from er- ror propagation problems. If the features are not well se- lected, the errors will be added up until the end. (Jiang and Zhai, 2007) systematically analyzed the effectiveness of different features for relation extraction on a large feature space and concluded that just using the basic unit features from each feature space (sequence, syntactic and depen- dency relation) can achieve reasonably good performance, and adding more complex features may not benefit the re- sult. The kernel-based approach is to compute a kernel function to measure the similarity between two data ob- jects. Such work includes (Zelenko et al., 2003), (Culotta and Sorensen, 2004), (Bunescu and Mooney, 2005), (Zhang et al., 2006), (Zhou et al., 2007), (Wang, 2008) and (Plank and Moschitti, 2013). The key issue of the kernel-based approach is the slow training and prediction time, so it is not good enough to process big data. With the development of deep learning, a series of neural …

Content truncated to 500 words