CETA: A Consensus Enhanced Training Approach for Denoising in Distantly Supervised Relation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2247–2258 October 12–17, 2022. 2247 CETA: A Consensus Enhanced Training Approach for Denoising in Distantly Supervised Relation Extraction Ruri Liu1∗, Shasha Mo1†, Jianwei Niu2, Shengda Fan1 1 School of Cyber Science and Technology, Beihang University, Beijing 100191, China 2 State Key Laboratory of Virtual Reality Technology and Systems, School of Computer Science and Engineering, Beihang University, Beijing 100191, China {ruriliu, moshasha, niujianwei, fanshengda}@buaa.edu.cn Abstract Distantly supervised relation extraction aims to extract relational facts from texts but suf- fers from noisy instances. Existing methods usually select reliable sentences that rely on potential noisy labels, resulting in wrongly se- lecting many noisy training instances or un- derutilizing a large amount of valuable train- ing data. This paper proposes a sentence-level DSRE method beyond typical instance selec- tion approaches by preventing samples from falling into the wrong classification space on the feature space. Specifically, a theorem for denoising and the corresponding implementa- tion, named Consensus Enhanced Training Ap- proach (CETA), are proposed in this paper. By training the model with CETA, samples of dif- ferent classes are separated, and samples of the same class are closely clustered in the feature space. Thus the model can easily establish the robust classification boundary to prevent noisy labels from biasing wrongly labeled samples into the wrong classification space. This pro- cess is achieved by enhancing the classification consensus between two discrepant classifiers and does not depend on any potential noisy la- bels, thus avoiding the above two limitations. Extensive experiments on widely-used bench- marks have demonstrated that CETA signifi- cantly outperforms the previous methods and achieves new state-of-the-art results. 1 Introduction Relation Extraction (RE), which aims to identify the relation between two specific entities in the text, is a fundamental task in natural language pro- cessing. Most supervised RE methods demand large-scale labeled training data, which is diffi- cult to acquire manually. To alleviate the problem, Distant Supervision (DS) is proposed by (Mintz et al., 2009) to automatically generate the labeled *The first two authors contributed equally. †Corresponding authors. < 9:;<; =>?@, ?>AB, CD;AEFG > Distant'Supervision Knowledge'Base Text'Corpus ✓ S1: 9:;<; =>?@ was'born'in'America ✗ S2:'Steven,Jobs,found'Apple'in'America ✗ S3:'Steven,Jobs,died'in'America Bag Figure 1: An example of annotating text corpus by dis- tant supervision. S2 and S3 do not express the relation born but still considered valid instances. text corpus by aligning the plain texts and knowl- edge bases. For example, as shown in Figure 1, [Steve jobs, born, America] is a relational triple in the knowledge base, DS annotates all sentences that contain the entity pair (Steve jobs, America) as valid instances for relation born. However, DS inevitably introduces noisy labels when the sen- tences do not express the labeled relation (e.g., cases S2 and S3 in Figure 1). Hence, investigat- ing a denoising method against noisy labels has become an urgent demand for Distantly Supervised Relation Extraction (DSRE), which aims to train the unbiased RE model under DS-built dataset. To alleviate the noise issue in DSRE, existing studies …

Content truncated to 500 words