Aspect-Based Sentiment Analysis in Ho Chi Minh City Hotel Reviews Using Aspect-Conditioned BiLSTM With Cross-Aspect Attention and Ordinal Regression
Tam Khoi Tran, Phuoc Dinh Tran · IEEE Access · 2026
The rapid growth of tourism in Ho Chi Minh City has generated a huge volume of Vietnamese hotel reviews, creating an urgent demand for effective aspect-based sentiment analysis (ABSA). This paper introduces a newlarge-scale benchmark dataset of 20,290 manually annotated reviews collected from Agoda, Google Reviews, and Traveloka, covering ten hotel aspects with 5-point ordinal sentiment labels (-2 to +2). The dataset is publicly available at https://github.com/tamkhojj/AC-BiLSTM represents the large and comprehensive Ho Chi Minh city tourism ABSA resource to date.We further propose AC-BiLSTM, a novel model that combines PhoBERT embeddings with aspect-conditioned BiLSTM, multi-head aspect attention, cross-aspect attention, aspect-conditioned gating, and ordinal regression loss. Extensive experiments show that AC-BiLSTM significantly outperforms strong LSTM-based baselines, achieving a macro F1-score of 0.8899, Precision of 0.8921, Recall of 0.8880, and Accuracy of 0.9456. Ablation studies and error analysis confirm the effectiveness and robustness of the proposed components. The dataset are publicly released to facilitate future research on low-resource Vietnamese ABSA.