A Distantly-Supervised Relation Extraction Method Based on Selective Gate and Noise Correction

A Distantly-Supervised Relation Extraction Method Based on Selective Gate and Noise Correction Zhuowei Chen1, Yujia Tian1, Lianxi Wang�1,2, Shengyi Jiang1,2 1School of Information Science and Technology, Guangdong University of Foreign Studies, Guangzhou, China, 510006 2Guangzhou Key Laboratory of Multilingual Intelligent Processing, Guangzhou, China, 510006 {20211003051,20211003065,wanglianxi}@gdufs.edu.cn jiangshengyi@163.com Abstract Entity relation extraction, as a core task of information extraction, aims to predict the relation of entity pairs identified by text, and its research results are applied to various fields. To address the problem that current distantly supervised relation extraction (DSRE) methods based on large- scale corpus annotation generate a large amount of noisy data, a DSRE method that incorporates selective gate and noise correction framework is proposed. The selective gate is used to reason- ably select the sentence features in the sentence bag, while the noise correction is used to correct the labels of small classes of samples that are misclassified into large classes during the model training process, to reduce the negative impact of noisy data on relation extraction. The results on the English datasets clearly demonstrate that our proposed method outperforms other base- line models. Moreover, the experimental results on the Chinese dataset indicate that our method surpasses other models, providing further evidence that our proposed method is both robust and effective. 1 Introduction Entity Relation Extraction (RE) is a crucial task in information extraction that aims to identify the rela- tion between entity pairs in text. The findings of RE have practical applications in several fields, such as the construction of knowledge graphs (KG), semantic web annotation, and the development and opti- mization of question-and-answer systems and search engines, which have a significant impact on daily life. However, the task of RE is challenging due to the limited availability of annotated data. To address this challenge, distant supervision has been proposed, which automatically annotates data, significantly increasing the number of annotated samples. However, distant supervision suffers from a strong hypothesis, leading to a large number of noisy labels during data annotation. Training on a dataset with noisy labels can result in model overfitting to the noise, which adversely impacts the model’s performance (Li et al., 2022b; Christou and Tsoumakas, 2021). To mitigate these issues, this paper proposes a novel method for RE that incorporates selective gate and the end-to-end noise correction method. In our model, selective gate is utilized to rationally select sentence features in the sentence bag, while noise correction is used to correct the labels of small classes of samples that are misclassified into larger classes during model training. These techniques reduce the negative impact of noisy data on the distant DSRE model. Additionally, since common word embedding models, such as Word2Vec and Glove, produce static vectors that overlook contextual semantics and the flexible use of multiple-meaning words, this paper introduces a pre-trained language model (PLM) to encode and extract features from sentences. This approach provides richer sentence semantic features, ef- fectively improving prediction accuracy and reducing training time. Experiment results demonstrate that this method significantly outperforms …

Content truncated to 500 words