Exploratory Neural Relation Classification for Domain Knowledge Acquisition

Proceedings of the 27th International Conference on Computational Linguistics, pages 2265–2276 Santa Fe, New Mexico, USA, August 20-26, 2018. 2265 Exploratory Neural Relation Classification for Domain Knowledge Acquisition Yan Fan, Chengyu Wang, Xiaofeng He∗ School of Computer Science and Software Engineering, East China Normal University, Shanghai, China {eileen940531,chywang2013}@gmail.com, xfhe@sei.ecnu.edu.cn Abstract The state-of-the-art methods for relation classification are primarily based on deep neural net- works. This supervised learning method suffers from not only limited training data, but also the large number of low-frequency relations in specific domains. In this paper, we propose an ex- ploratory relation classification method for domain knowledge harvesting. The goal is to learn a classifier on pre-defined relations while discovering new relations expressed in texts. A dynami- cally structured neural network is introduced to classify entity pairs to a continuously expanded relation set. We further propose the similarity sensitive Chinese restaurant process to discover new relations. Experiments conducted on a large corpus show that new relations are discovered with high precision and recall, illustrating the effectiveness of our method. Title and Abstract in Chinese 基于领域知识获取的探索式神经网络关系分类 近几年来，关系分类任务主要利用神经网络模型来自动学习复杂的特征。然而这类监督 式学习方法的局限性在于，首先它需要大量训练数据，其次特定领域存在长尾的低频关 系无法被有效地预定义。在本论文中，我们提出了基于领域知识获取的探索式关系分类 任务。它的目标在于学习预定义关系的分类器，同时从文本中发现新的语义关系。我们 提出了一个动态结构的神经网络，它可以对持续扩充的关系集进行分类。我们进一步提 出了相似度敏感的中餐馆过程算法，用于发现新关系。基于大语料库上的实验证明了该 神经网络的分类效果，同时新发现的关系也有较高的准确率和召回率。 1 Introduction Relation classification assigns semantic relation labels to entity pairs in texts. Besides traditional feature- based (Kambhatla, 2004) and kernel-based (Bunescu and Mooney, 2005) approaches, neural networks (NNs) are introduced to harvest relational facts in recent years. Classical architectures include convolu- tional neural networks (CNNs) (Zeng et al., 2014), recurrent neural networks (RNNs) (Xu et al., 2015), etc. However, above methods are still insufficient for domain knowledge acquisition due to two chal- lenges: (i) most domain entities rarely occur in the corpus, hence pattern-based methods easily suffer from the feature sparsity problem; (ii) a domain knowledge graph tends to be incomplete w.r.t. relation labels and facts (Fan et al., 2017). As a result, unlabeled entity pairs are likely to be wrongly forced into existing relation labels by distant supervision approaches, instead of their true, unknown labels. In this paper we propose a method, Exploratory Relation Classification (ERC), to populate domain knowledge graphs automatically. The goal of ERC is to not only classify entity pairs into a finite set of pre-defined relations, but also simultaneously discover previously unseen relation types and their respective instances from plain texts with high confidence. The resulting numbers of relation types and facts in the domain knowledge graph grow continuously. ∗Corresponding author. This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/ 2266 To solve ERC problem, we propose a Dynamic Structured Neural Network (DSNN). In the network, the convolutional and recurrent network units are employed to model local information such as syntactic and lexical features from the sentence level. For the context sparsity problem (Challenge (i)) an embed- ding layer is introduced to encode corpus-level semantic features of domain entities. The output classes of DSNN can be automatically expanded during the iterative training process, in order to classify entity pairs to both known and newly discovered relations. To explore new relations and corresponding …

Content truncated to 500 words