Research on the Recognition of Internet Buzzword Features Based on Transformer

Dawei Xu, Yijie She, Zhonghua Tan, Ruiguang Li, Jian Zhao

Abstract: Accurate identification of Internet buzzwords plays an important role in positive Internet opinion guidance. A Transformer-based Internet buzzword feature recognition system was designed to address this problem. The traditional way of crawling data has been improved, a real-time crawling module has been added, and an Internet buzzword corpus has been constructed by itself. The traditional way of crawling data has been improved, a real-time crawling module has been added, and an Internet buzzword corpus has been constructed by itself. Traditional machine learning models suffer from gradient disappearance and gradient explosion, the Transformer model, with its parallel computing and self-attentive mechanism, is a good solution to these problems, and its bi-directional connection allows the parameters of the context to be updated uniformly, thus allowing better aggregation of information and solving the problem of scattered contextual information. Transformation of the position-encoded part of the Transformer model starts with a relative position representation (RPR). It compensates for its inability to obtain relative location information. The experimental results show that the improved Transformer model can achieve an accuracy rate of 90.1%, a recall rate of 92.13%, and an F1 value of 91.16% in recognizing Internet buzzwords.