A Hybrid LSTM-CNN Approach Using Multilingual BERT for Sentiment Analysis of GERD Tweets

Abstract

Analyzing public sentiment through platforms like Twitter is a common approach for understanding opinions on political matters. This study introduces a deep learning sentiment analysis model that integrates Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) to assess attitudes toward the Grand Ethiopian Renaissance Dam (GERD). LSTM is utilized to capture long-range dependencies in text, while CNN identifies significant local patterns. An initial dataset of 30,000 unlabeled tweets was collected in 2024 G.C., out of which 17,064 were labeled as positive, negative, or neutral. The labeled tweets were divided into 13,112 for training and the remaining for testing. The hybrid LSTM-CNN model demonstrated superior performance compared to the standalone models, delivering more accurate and balanced sentiment classification. A major feature of this study is the analysis of tweets written in Amharic, Arabic, and English. The model was trained over 35 epochs with a batch size of 46 and a learning rate of 0.001. Using multilingual BERT (mBERT) embeddings notably enhanced the model’s performance, with training and testing accuracies reaching 95.3% and 92%, respectively. The hybrid model also achieved a precision, recall, and F1-score of 90%. In a focused analysis of Arabic tweets, 3,710 were negative, 9,793 positive, and 4,814 neutral. These results emphasize the influence of linguistic diversity and class distribution on classification performance. While mBERT showed strong results, addressing class imbalance and expanding language-specific features remains crucial for further improvements.