[In-person]An Enhanced Ontology-Based Framework for Sentiment Analysis of Vietnamese Social Media Comments

An Enhanced Ontology-Based Framework for Sentiment Analysis of Vietnamese Social Media Comments
ID:103 Submission ID:42 View Protection:ATTENDEE Updated Time:2026-10-04 23:40:55 Hits:163 In-person

Start Time:2026-10-13 09:15 (Asia/Ho_Chi_Minh)

Duration:15min

Session:[S5] Track 5: Emerging Trends of AI/ML » [S5-1] Track 5: Emerging Trends of AI/ML

Presentation File

Tips: The file permissions under this presentation are only for participants. You have not logged in yet and cannot view it temporarily.

Abstract
Automated sentiment analysis of Vietnamese social media comments presents several distinctive linguistic challenges, including ambiguous compound-word boundaries, implicit negation, polarity shifts across contrastive clauses, and severe class imbalance among fine-grained sentiment levels (N1: Light Negative, N2: Heavy Negative, P1: Light Positive, and P2: Heavy Positive). Although contextual transformer models provide strong predictive performance, they offer limited interpretability and may incorrectly handle polarity changes caused by intensifiers, negations, and contrastive conjunctions. In contrast, rule-based ontologies provide transparent reasoning but are often brittle when processing informal language, spelling errors, slang, and out-of-vocabulary expressions. To address these limitations, this paper proposes the \textbf{Enhanced Ontology-Based Sentiment Analysis Framework (EOSAF)}, which combines deep contextual representations with structured Knowledge Graph (KG) reasoning. The framework comprises: (1) a selectively fine-tuned PhoBERT encoder with hybrid CLS-mean pooling, producing a 768-dimensional contextual representation ($h_{\text{bert}} \in \mathbb{R}^{768}$); (2) an ontology-guided KG reasoning module that models intensifier scaling, negation-window inversion, contrastive-clause analysis, and idiom substitution as a 64-dimensional interpretable feature vector ($h_{\text{kg}} \in \mathbb{R}^{64}$); (3) feature-fusion mechanisms that integrate contextual and symbolic representations; and (4) a learnable Ontology-based Threshold Adjustment layer that calibrates class logits using structured KG evidence. An adaptive offline caching mechanism containing 4,771 domain entities is further incorporated to support efficient local knowledge retrieval during inference. Experiments on real-world Vietnamese social media comments show that the proposed framework achieves an accuracy of \textbf{98.71\%} and an F1-Macro score of \textbf{0.9868}, outperforming existing baselines by up to 12.14\% while preserving interpretable ontology-based reasoning.
Keywords
Vietnamese,Sentiment Analysis,Opinion Mining,Transformer,Classification,Social Media.
Speaker
Thanh Tam Nguyen
Researchers University of Science, Vietnam National University

Submission Author
Thu Hoang Posts and Telecommunications Institute of Technology
Thanh Tam Nguyen University of Science, Vietnam National University
Thi Huyen Trang Le East Asia University of Technology
Binh Cong Nguyen East Asia University of Technology
Trung Kien Ta East Asia University of Technology
Ph.D Nguyen Viet Hung East Asia University of Technology
Comment submit
Verification code Change another
All comments

CONTACT US

Email: asiancomnet2026@usssociety.org

Website & IT Support: hi@aconf.org

Registration Submit Paper