Mobile Text Misinformation Identification Using Machine Learning

Introduction

More than six million people died because of the COVID-19 as of June 2023 (World Health Organization, 2023). These high casualties put everyone on the alert. People try to find any information that helps them fight the virus. Much information they receive is from their smartphones without doubt in these days because smartphones have become an indispensable device for everyone. One popular function for smartphone users is sending and receiving text messages. Instead of reading newspapers or watching TV news, many mobile users especially younger generations receive their daily news or information via text messages. Other than useful and unbiased information, many of the messages are incorrect or may even be distorted on purpose (van der Linden, 2022). During the pandemic, the problem has become even more serious because the health text misinformation not only gives wrong information, but may also cause fatal results such as advocacy of the ineffectiveness of vaccine. This research tries to mitigate the problem by identifying mobile health text misinformation, so the mobile users can use the findings to better judge the messages they receive and take actions accordingly.

This research proposes a self-reconfigurable system for identifying mobile health text misinformation, which is briefly described as follows. This is a supervised learning system, so before the system is put into use, it needs training by using a set of text messages with known results. The initial parameters of the system are set by heuristics because the keywords of text messages are unknown in advance and have to be speculated. After the training phase, the system starts its testing phase by receiving text messages. Each message will go through a series of steps: preprocessing (including lexical analysis, stopword removal, and stemming), indexing and storage, and testing (classification) by using a dataflow graph. Instead of applying the results immediately, the first round of testing is used to reconfigure the system in order to generate better results later. It is because the initial configuration is usually not desirable as system parameters are unknown in the very beginning. After the first round of testing, better parameters could be found from the test results. The complete steps are repeated with better parameter and more accurate identifications are expected. Experiment results show the accuracy of the proposed method meets the expectation, but still has room for improvement. An explanation for this may be because the short messages do not provide much information and small deviation may cause a great impact on the results. Further refinements are needed before it is put into use.

The rest of this paper is organized as follows. Section 2 shows the background information about this research and related works on misinformation detection. The structure and components of the proposed system are given in Section 3. Section 4 proposes our major method, a self-reconfigurable dataflow graph, for detecting health text misinformation. The experiment results and evaluations are given in Section 5, followed by a conclusion and references.