Pairwise Relation Classification with Mirror Instances and a Combined Convolutional Neural Network

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2366–2377, Osaka, Japan, December 11-17 2016. Pairwise Relation Classification with Mirror Instances and a Combined Convolutional Neural Network Jianfei Yu School of Information Systems Singapore Management University jfyu.2014@phdis.smu.edu.sg Jing Jiang School of Information Systems Singapore Management University jingjiang@smu.edu.sg Abstract Relation classification is the task of classifying the semantic relations between entity pairs in text. Observing that existing work has not fully explored using different representations for relation instances, especially in order to better handle the asymmetry of relation types, in this paper, we propose a neural network based method for relation classification that combines the raw sequence and the shortest dependency path representations of relation instances and uses mirror instances to perform pairwise relation classification. We evaluate our proposed models on two widely used datasets: SemEval-2010 Task 8 and ACE-2005. The empirical results show that our combined model together with mirror instances achieves the state-of-the-art results on both datasets. 1 Introduction Relation classification is a very important task for many Natural Language Processing (NLP) applica- tions including question answering (Yao and Van Durme, 2014), knowledge base population (Socher et al., 2013) and opinion mining (Kobayashi et al., 2007). The goal of relation classification is to auto- matically identify the semantic relation between a pair of entities in free text. For example, a relation classification system should be able to capture the Cause-Effect relation between the entities pressure and burst in the sentence “The burst has been caused by water hammer pressure.” Like any classification task, a key research question of relation classification is the identification of a good feature representation for each relation instance. Traditional approaches focus on either com- bining many manually designed features (Zhou et al., 2005; Jiang and Zhai, 2007; Li and Ji, 2014) or leveraging various kernels to implicitly explore a large feature space (Bunescu and Mooney, 2005; Zhang et al., 2006; Qian et al., 2008; Nguyen et al., 2009), but both approaches suffer from their poor generalization ability on unseen words, and fail to achieve very satisfactory performance (Nguyen et al., 2015). Recently, with the advances of deep learning in NLP, neural networks (NNs) have exhib- ited their advantages in dealing with unseen words through pre-trained word embeddings and capturing meaningful hidden representations. Different NN architectures, including Convolutional Neural Net- work (CNN) (Zeng et al., 2014), Recursive Neural Network (ReNN) (Socher et al., 2012) and Recurrent Neural Network (RNN) (Xu et al., 2015b), have been applied to relation classification. However, most existing NN-based approaches only exploit one of the following structures to represent relation instances: raw word sequences (Zeng et al., 2014; dos Santos et al., 2015), constituency parse trees (Socher et al., 2012; Hashimoto et al., 2013) and dependency parse trees (Xu et al., 2015a; Xu et al., 2015b; Miwa and Bansal, 2016). For the models based on raw sequence, despite maintaining all the information in relation instances, they cannot well handle long-distance relations. For the models based on constituency …

Content truncated to 500 words