Aspect-Based Sentiment Analysis for Vietnamese Product Reviews using PhoBERT, BiLSTM, and CRF with Weighted Emission

This paper presents a study of PhoBERT-based sequence labeling approaches for Aspect-Based Sentiment Analysis in Vietnamese product review text. We utilize the UIT-ViSD4SA dataset comprising 11,122 annotated reviews with 35,396 character-level span annotations collected from Vietnamese e-commerce platforms. The research constructs an end-to-end model combining PhoBERT as a contextual encoder, Bidirectional LSTM for sequential pattern learning, and a Conditional Random Field decoder for BIO sequence labeling. To address token-level class imbalance, we introduce weighted CRF emissions with suppressed O-label weights. Our system achieves a macro F1-span of 0.6655 on the validation set after 20 training epochs.