Relation Classification Using Segment-Level Attention-based CNN and Dependency-based RNN

Proceedings of NAACL-HLT 2019, pages 2793–2798 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2793 Relation Classification Using Segment-Level Attention-based CNN and Dependency-based RNN Van-Hien Tran1, Van-Thuy Phi1, Hiroyuki Shindo1,2, and Yuji Matsumoto1,2 1 Nara Institute of Science and Technology 2RIKEN Center for Advanced Intelligence Project (AIP) 1{tran.van hien.ts1,phi.thuy.ph8,shindo,matsu}@is.naist.jp Abstract Recently, relation classification has gained much success by exploiting deep neural net- works. In this paper, we propose a new model effectively combining Segment-level Attention-based Convolutional Neural Net- works (SACNNs) and Dependency-based Re- current Neural Networks (DepRNNs). While SACNNs allow the model to selectively focus on the important information segment from the raw sequence, DepRNNs help to handle the long-distance relations from the shortest de- pendency path of the related entities. Exper- iments on the SemEval-2010 Task 8 dataset show that our model is comparable to the state- of-the-art without using any external lexical features. 1 Introduction Relation classification (RC) is a fundamental task in Natural Language Processing (NLP) that aims to identify semantic relations between pairs of marked entities in given sentences (instances). It has attracted much research effort as it plays a vital role in many NLP applications such as Informa- tion Extraction and Question Answering (Nguyen and Grishman, 2015). Traditional approaches (Kambhatla, 2004; Zhang et al., 2006) usually rely heavily on hand-crafted features and lexical re- sources, or elaborately designed kernels, which are time-consuming and challenging to adapt to novel domains. Recently, neural network (NN) models have dominated the work on RC since they can effectively learn meaningful hidden features without human intervention. However, most previous NN models only ex- ploit one of the following structures to represent relation instances: raw word sequences (Zhou et al., 2016; Wang et al., 2016) and dependency trees (Wen, 2017; Le et al., 2018). While raw se- quences can provide all the information of rela- tion instances, they also add noise to the models from redundant information. While dependency tree structures help the models focus on the con- cise information captured by the shortest depen- dency path (SDP) between two entities, they lose some supplementary context in the raw sequence. It is clear that the raw sequence and SDP highly complement each other. We, therefore, combine them to be more effective in determining the rela- tion without losing any information. While CNNs are able to learn short patterns (local features) (LeCun et al., 1995), RNNs have been effective in learning word sequence informa- tion (long-distance features) (Chung et al., 2014). In this paper, we present a new model combining both CNNs and RNNs, exploiting the information from both the raw sequence and the SDP. Our contributions are summarized as follows: (a) We combine Entity Tag Feature (ETF) (Qin et al., 2016) and Tree-based Position Feature (TPF) (Yang et al., 2016) to improve the seman- tic information between the marked entities in the raw input sentences. (b) We propose Segment-Level Attention-based Convolutional Neural Networks (SACNNs) which automatically pay special attention to the impor- tant text segments from …

Content truncated to 500 words