A Hybrid Feature Extraction Method for Nepali COVID-19-Related Tweets Classification

Area of Science:

• Natural Language Processing
• Computational Social Science
• Public Health

Background:

• COVID-19 has caused widespread mortality, with mental health impacts amplified by public fear.
• Social media, particularly Twitter, serves as a crucial platform for expressing public sentiment regarding health crises.
• Sentiment analysis of social media data requires sophisticated methods to interpret both linguistic structure and meaning.

Purpose of the Study:

• To analyze public sentiment towards COVID-19 in the Nepali language using social media data.
• To evaluate the effectiveness of combined syntactical and semantic text representation methods for sentiment analysis.
• To compare the performance of multiple machine learning classifiers using different feature extraction techniques.

Main Methods:

• Utilized TF-IDF and FastText for text representation, combining them into hybrid features.
• Implemented nine machine learning classifiers: Logistic Regression, SVM, Naive Bayes, KNN, Decision Trees, Random Forest, Extreme Tree, AdaBoost, and MLP.
• Evaluated methods on the publicly available Nepali-COVID-19 tweets dataset (NepCov19Tweets) with Positive, Negative, and Neutral categories.

Main Results:

• The hybrid feature extraction method demonstrated superior performance compared to TF-IDF and FastText alone.
• Machine learning classifiers achieved higher accuracy when using hybrid features across all tested algorithms.
• The proposed hybrid approach achieved excellent performance, surpassing existing state-of-the-art methods.

Conclusions:

• Combining syntactical (TF-IDF) and semantic (FastText) features offers a robust approach for COVID-19 sentiment analysis in Nepali.
• The hybrid feature method enhances the accuracy of sentiment classification on social media data.
• This study provides valuable insights into public sentiment during the pandemic, applicable to public health communication and policy.