Research of Network Hotspot Events Joint Extracting Based on BERT-CNN-CRF Model for Internet Public Opinion

Abstract

Public emergencies affect the vital interests of a large number of citizens and are extremely concerned because of their complexity and strong influence. Early warning of public opinion is an important part of reducing secondary harm of public emergencies and ensuring social stability. There are problems of insufficient ability to capture the semantics of trigger words and the identification ambiguity of entity boundaries in event elements in the current study of internet hot events extraction. In this paper, we explore the joint extraction of events using sequence annotation, and construct a joint extracting method based on BERT-CNN-CRF model to obtain the input text fusion full-text semantic information vector by using BERT, and extract local contextual semantic features by convolutional neural network (CNN), and obtain joint extraction results by combining conditional random fields (CRF). After crawling several large public social media platforms with over ten thousand of data and comparing three baseline approaches in the latest event extraction methods, our proposed joint BERT-CNN-CRF extraction model has higher accuracy and better efficiency.