COVID-19 Informative Tweets Identification Through Word-by-Word Lexicon Replacement Using Pretrained Biomedical Corpus

The coronavirus pandemic has contributed toward the prevalence of numerous fake news and misleading information over the social media especially Twitter. Therefore, a task of identifying whether a tweet is informative or uninformative has caught many researchers’ attentions. The literature showed high dependency on transformers architectures. Yet, since the task requires more concentration on domain-specific terms, there would be a need for lexicon-based expansion. Hence, this paper proposes a word-by-word lexicon replacement method for the task of informative tweet extraction. A pretrained model of medical word embedding has been utilized to perform the replacement. In addition, multiple replacement conditions have been employed. Consequentially, different feature space representations have been applied upon the new tweet document with replaced terms. Lastly, a Logistic Regression (LR) classifier has been used to classify documents into Informative and Uninformative. Using the benchmark dataset of WNUT-2020 at Task 2, results of applying the proposed replacement method showed better performance than without the replacement by obtaining 0.91 of f1-score. This result emphasizes the importance and usefulness of the proposed replacement method in terms of improving the classification accuracy of informative tweet identification.