Automated sentiment analysis of Vietnamese social media comments faces distinctive challenges, including ambiguous compound-word boundaries, implicit negation, polarity shifts across contrastive clauses, and class imbalance among fine-grained sentiment levels (N1: Light Negative, N2: Heavy Negative, P1: Light Positive, P2: Heavy Positive). Contextual transformer models offer strong predictive performance but limited interpretability, often mishandling polarity changes from intensifiers, negations, and contrastive conjunctions. Rule-based ontologies, by contrast, provide transparent reasoning but remain brittle against informal language, spelling errors, slang, and out-of-vocabulary expressions. To address these limitations, this paper proposes the extbf{Enhanced Ontology-Based Sentiment Analysis Framework (EOSAF)}, combining deep contextual representations with structured Knowledge Graph (KG) reasoning. The framework comprises a selectively fine-tuned PhoBERT encoder with hybrid CLS-mean pooling, an ontology-guided KG reasoning module producing a 64-dimensional interpretable feature vector, a neural feature-fusion layer, and a learnable Ontology-based Threshold Adjustment layer for residual logit calibration, supported by an offline cache of 4,771 domain entities for efficient local knowledge retrieval. On a held-out set of 1,399 Vietnamese comments, EOSAF correctly classifies 1,381 samples, achieving 98.71% accuracy and a Macro-F1 of 0.9868—a 4.59-percentage-point improvement over the standalone PhoBERT baseline while retaining explicit ontology-derived evidence.

Short-video retention;Watch-time predition;Time-series forecasting;Resource-aware machine learning;CPU-aware;QoE