Identifying informative tweets during a pandemic via a topic-aware neural language model

Area of Science:

• Computational Linguistics
• Public Health Informatics
• Social Media Analysis

Background:

• Epidemics like COVID-19 generate vast amounts of social media data.
• Analyzing this data is crucial for emergency response but challenging due to non-informative content.
• Existing methods often overlook unlabeled data and topic knowledge, limiting their effectiveness.

Purpose of the Study:

• To develop an automated system for detecting useful information in COVID-19 tweets.
• To address the limitations of existing methods by incorporating topic information and unlabeled data.
• To improve the accuracy and efficiency of social media analysis for public health emergencies.

Main Methods:

• Proposed a novel Topic-Aware BERT (TABERT) model.
• Utilized a topic model to extract latent topics from tweets.
• Integrated topic information with BERT outputs using a flexible framework.
• Employed adversarial training for semi-supervised learning to leverage unlabeled data.

Main Results:

• The TABERT model demonstrated superior performance compared to classic and state-of-the-art baselines.
• Experimental results on a COVID-19 English tweet dataset validated the model's effectiveness.
• The model successfully improved inner representations by utilizing unlabeled data.

Conclusions:

• The Topic-Aware BERT (TABERT) model offers a significant advancement in analyzing social media data during epidemics.
• Incorporating topic modeling and semi-supervised learning enhances the detection of informative tweets.
• This approach can aid emergency response organizations in making better-informed decisions during public health crises.