A Deep Learning-Based System for Detecting Inappropriate Language in Vietnamese Text

Tóm tắt

With the rapid growth of social media in Vietnam, detecting hate speech in Vietnamese text has become increasingly critical. This paper presents a comprehensive comparison of six deep learning architectures for Vietnamese hate speech detection, evaluated on a large-scale balanced dataset of 300,000 Vietnamese social media comments. We compare lightweight models (BiLSTM, BiGRU, with and without attention, and Pure Transformer) using frozen PhoBERT embeddings against a selectively fine-tuned PhoBERT model. Our fine-tuned PhoBERT achieves state-of-the-art performance with 92.82% F1-score and accuracy, outperforming the best lightweight baseline by 1.08 percentage points. Compared to prior work, our approach demonstrates substantial improvements of 30.12, 21.39, and 24.12 percentage points over ViHSD, VLSP 2019, and ViHateT5 respectively. These gains stem from three factors: (1) larger-scale balanced training data (300,000 samples), (2) Vietnamese-specific pre-training with syllable-level tokenization, and (3) selective fine-tuning of only the top 40% of PhoBERT's layers. We incorporate gradient-based interpretability to visualize token importance scores, enabling explainable predictions. Our analysis reveals that while fine-tuned PhoBERT achieves the highest accuracy, lightweight models offer competitive performance (91.74% F1-score) with significantly lower computational requirements. This work provides both state-of-the-art results and practical insights into accuracy-efficiency trade-offs for Vietnamese hate speech detection systems.