ARNOR: Attention Regularization based Noise Reduction for Distant Supervision Relation Classification

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1399–1408 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 1399 ARNOR: Attention Regularization based Noise Reduction for Distant Supervision Relation Classification Wei Jia, Dai Dai, Xinyan Xiao and Hua Wu Baidu Inc., Beijing, China {jiawei07, daidai, xiaoxinyan, wu hua} @baidu.com Abstract Distant supervision is widely used in rela- tion classification in order to create large-scale training data by aligning a knowledge base with an unlabeled corpus. However, it also introduces amounts of noisy labels where a contextual sentence actually does not express the labeled relation. In this paper, we propose ARNOR, a novel Attention Regularization based NOise Reduction framework for distant supervision relation classification. ARNOR assumes that a trustable relation label should be explained by the neural attention model. Specifically, our ARNOR framework itera- tively learns an interpretable model and uti- lizes it to select trustable instances. We first introduce attention regularization to force the model to pay attention to the patterns which explain the relation labels, so as to make the model more interpretable. Then, if the learned model can clearly locate the relation patterns of a candidate instance in training set, we will select it as a trustable instance for further train- ing step. According to the experiments on NYT data, our ARNOR framework achieves significant improvements over state-of-the-art methods in both relation classification perfor- mance and noise reduction effect. 1 Introduction Relation Classification (RC) is a fundamental task in natural language processing (NLP) and is par- ticularly important for knowledge base construc- tion. The goal of RC (Zelenko et al., 2003) is to identify the relation type of a given entity pair in a sentence. Generally, a relation should be ex- plicitly expressed by some clue words. See the first sentence in Figure 1. The phrase “was born in” explains the relation type “place of birth” for “Bill Lockyer” and “California”. Such indicating words is called patterns (Hearst, 1992; Hamon and Nazarenko, 2001). knowledge database sentences Relation place_of_birth California Bill Lockyer Entity 1 Entity 2 Bill Lockyer was born in California Bill Lockyer is an attorney general of California place_of_birth place_of_birth Align Figure 1: Two relation instances generated by distant supervision. The bold words “was born in” in s1 is the pattern that explains the relation type “place of birth”. Hence, this instance is correctly labeled. However, the second instance is noisy due to the lack of correspond- ing relation pattern. Before Enity1 Entity 1 Between Entity 2 After Entity 2 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Average Attention Figure 2: Average attention weights of BiLSTM+ATT model across five parts in the sentences on our test set. This model is trained using noisy data generated by dis- tant supervision. It mainly pays attention to the input entity pair and ignores other words which might ex- press the real relation. It also happens in Figure 1. This result comes from the fact that DS method only depends on entities …

Content truncated to 500 words