LinBGN-Net: A Unified Linguistic Framework for Multi-Task Text Classification in Telugu Using BERT, GCN, and Naive Bayes
Shaik Asif Hussain, Fahimuddin Shaik, Shaik Karimullah, Gadde Sureshbabu · Forum for Linguistic Studies · 2025
Despite the growing interest in Natural Language Processing (NLP), linguistically grounded multi-task modeling for low-resource languages like Telugu remains underexplored, particularly across semantically complex tasks such as sentiment, emotion, hate speech, sarcasm, and clickbait detection. This study aims to address this gap by conducting a comprehensive linguistic analysis of Telugu texts through the lens of computational modeling. To that end, we propose LinBGN-Net, an innovative ensemble deep learning framework that integrates BERT for contextual embedding, Graph Convolutional Networks (GCN) for structural understanding, and Naive Bayes (NB) for statistical grounding. The model is trained on five well-balanced Telugu datasets (totalling over 250,000 sentences) from Hugging Face, covering diverse label spaces (2 to 5 classes), thus facilitating an in-depth linguistic study across varied text genres and expressions. LinBGN-Net achieves macro-F1 scores of 0.86 (Sentiment), 0.84 (Emotion), 0.93 (Hate Speech), 0.92 (Sarcasm), and 0.95 (Clickbait), outperforming standard baselines such as Naive Bayes, SVM, LSTM, and even standalone BERT by margins of 2–8% across tasks. The results not only demonstrate the effectiveness of LinBGN-Net as a high-performing multi-task model, but also offer valuable linguistic insights into how Telugu expressions of sentiment, emotion, intent, and persuasion can be computationally modeled and understood—contributing significantly to both linguistic research and real-world NLP applications.