CovTiNet: Covid text identification network using attention-based positional embedding feature fusion.

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)
• Computational Linguistics

Background:

• The proliferation of online Covid-19 information, including misinformation and disinformation, necessitates effective identification methods.
• Existing research on Covid text identification (CTI) is limited, particularly for low-resource languages like Bengali.
• Automatic CTI in Bengali is challenging due to linguistic complexity and a lack of NLP resources.

Purpose of the Study:

• To develop an effective deep learning model for identifying Covid-19 related text in Bengali.
• To address the infodemic by enabling better control over misinformation and disinformation.
• To overcome the challenges of CTI in low-resource languages.

Main Methods:

• A novel deep learning network, CovTiNet, was proposed for Bengali Covid text identification.
• CovTiNet utilizes attention-based position embedding for feature fusion and an attention-based Convolutional Neural Network (CNN).
• A new dataset, BCovC, was developed for training and evaluating the model.

Main Results:

• The proposed CovTiNet model achieved a highest accuracy of 96.61±.001% on the BCovC dataset.
• CovTiNet outperformed several established NLP models and baselines, including BERT-M, IndicBERT, and ELECTRA-Bengali.
• The model demonstrated superior performance in the challenging domain of Bengali CTI.

Conclusions:

• CovTiNet offers a robust and accurate solution for identifying Covid-19 related text in Bengali.
• The developed model contributes significantly to NLP research for low-resource languages.
• Effective CTI is crucial for mitigating the societal impact of online misinformation during health crises.