Topic Audiolization: A Model for Rumor Detection Inspired by Lie Detection Technology

The psychological factors influence the frequency and intensity of the lie's audio. In addition, the changes could be associated with each word in a segment of the liar's output. We consider a single comment as the audio signal of a word and the entire propagation process as the audio signal of a paragraph. Thus, rumor detection based on propagation features is transformed into lie detection based on audio signals. Firstly, we propose a Topic2Audio method to transform the topic space into an audio-like signal. This method quantifies topic propagation features and maps them to the audio-like signal's amplitude, frequency, and offset. Secondly, we retrieved audio features using Fourier transforms and Mel spectrum algorithms. Finally, we propose a rumor detection model based on topic audiolization (TARD). The model can detect 'rumor' signals using audio classification techniques by transforming the topic space into audio -like signals. Experiments conducted on three datasets, i.e., Weibo, Politifact, and Gossipcop, show performance improvements at +2.2%, +1.9%, and +1.3% when compared with the state-of-the-art method, respectively. Experiments show that our model can identify rumors effectively.

Automated summary

This study proposes a rumor detection model based on topic audiolization, which transforms the topic space into audio-like signals. Experimental results show that the model achieves significant performance improvements in rumor identification.