Semantic Relation Classification via Convolutional Neural Networks with Simple Negative Sampling

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 536–540, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Semantic Relation Classification via Convolutional Neural Networks with Simple Negative Sampling Kun Xu1, Yansong Feng∗1, Songfang Huang2 and Dongyan Zhao1 1ICST, Peking University, Beijing, China 2IBM China Research Lab, Beijing, China {xukun,fengyansong,zhaodongyan}@pku.edu.cn huangsf@cn.ibm.com Abstract Syntactic features play an essential role in identifying relationship in a sentence. Pre- vious neural network models directly work on raw word sequences or constituent parse trees, thus often suffer from irrele- vant information introduced when subjects and objects are in a long distance. In this paper, we propose to learn more robust re- lation representations from shortest depen- dency paths through a convolution neu- ral network. We further take the relation directionality into account and propose a straightforward negative sampling strategy to improve the assignment of subjects and objects. Experimental results show that our method outperforms the state-of-the- art approaches on the SemEval-2010 Task 8 dataset. 1 Introduction The relation extraction (RE) task can be defined as follows: given a sentence S with a pair of nomi- nals e1 and e2, we aim to identify the relationship between e1 and e2. RE is typically investigated in a classification style, where many features have been proposed, e.g., Hendrickx et al. (2010) de- signed 16 types of features including POS, Word- Net, FrameNet, dependency parse features, etc. Among them, syntactic features are considered to bring significant improvements in extraction accu- racy (Bunescu and Mooney, 2005a). Earlier at- tempts to encode syntactic information are mainly kernel-based methods, such as the convolution tree kernel (Qian et al., 2008), subsequence kernel (Bunescu and Mooney, 2005b), and dependency tree kernel (Bunescu and Mooney, 2005a). With the recent success of neural networks in natural language processing, different neural net- work models are proposed to learn syntactic fea- tures from raw sequences of words or constituent parse trees (Zeng et al., 2014; Socher et al., 2012), which have been proved effective, but, often suf- fer from irrelevant subsequences or clauses, espe- cially when subjects and objects are in a longer distance. For example, in the sentence, “The [singer]e1 , who performed three of the nominated songs, also caused a [commotion]e2 on the red carpet”, the who clause is used to modify subject e1, but is unrelated to the Cause-Effect relation- ship between singer and commotion. Incorporat- ing such information into the model will hurt the extraction performance. We therefore propose to learn a more robust relation representation from a convolution neural network (CNN) model that works on the simple dependency path between subjects and objects, which naturally characterizes the relationship between two nominals and avoids negative effects from other irrelevant chunks or clauses. Our second contribution is the introduction of a negative sampling strategy into the CNN mod- els to address the relation directionality, i.e., prop- erly assigning the subject and object within a re- lationship. In the above singer example, (singer, commotion) hold the Cause-Effect relation, while (commotion, …

Content truncated to 500 words