Thien Khai Tran
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Thien Khai Tran has a total of 81 co-authors. Here are the top co-authors by number of shared publications:
Tuoi Thi Phan
8 shared publications- Hien Nguyen
Penn State Harrisburg
3 shared publications - An Cong Tran
Can Tho University
2 shared publications Bay Vo
2 shared publicationsDai Phuoc Nguyen
2 shared publications- Dinh Minh Hoa
Ho Chi Minh City University of Foreign Languages and Information Technology
2 shared publications
Latest Publications
- 8 October 2026
- Published by Emerald Publishing in Data Technologies and Applications
- p. 1-24
Evasion-hate: targeted augmentation for hate speech detection under prompt-induced surface-form shifts
Purpose: Online hate speech detection is increasingly used as a decision-support component in content moderation, but most detectors are evaluated on clean held-out text. In practice, users can preserve harmful intent while changing surface form through spelling obfuscation, coded wording or code-switching, creating a robustness gap between benchmark performance and deployment reliability. This paper introduces Evasion-Hate, a targeted, quality-filtered augmentation pipeline for three-way hate/offensive/normal classification in English and Vietnamese. Design/methodology/approach: The pipeline generates synthetic large language model (LLM) candidate rewrites, filters them with checks independent of the downstream classifiers under evaluation and evaluates robustness with paired bootstrap testing, matched quantity controls, competitive augmentation baselines and external functional suites. Findings: Under the main evasion protocol, Evasion-Hate achieves the strongest English robustness result among comparable three-way systems evaluated under our protocol: XLM-R robust Macro-F1 rises from 0.4783 to 0.7582 (+27.99 pp); model-specific PFR is 37.16% for the baseline and 16.66% for Evasion-Hate, while model-specific harmful-only ASR is 56.56% and 11.57%, respectively. Vietnamese models show statistically supported gains on the reconstructed synthetic evaluation sets (PhoBERT-v2 +4.60 pp; ViSoBERT +3.29 pp; both significant by pooled paired bootstrap), although the Vietnamese transform families remain below or near the pre-specified fidelity threshold and should be interpreted diagnostically. Matched quantity controls show the gain is not explained by additional training volume; external HateCheck spelling-variation flagged rate improves by 26.42 pp. Leave-one-transformation-out evaluation shows partial and family-dependent transfer. In English, holding out euphemistic and coded rewriting yields a +18.01 pp gain that remains supported after Holm correction, whereas the +8.48 pp gain for held-out code-switching does not survive multiplicity correction and the character-level family shows little change (+0.31 pp). In Vietnamese, held-out code-switching shows a positive +2.36 pp effect with a bootstrap interval excluding zero, but it does not survive Holm correction; the remaining folds and the pooled evaluation show no supported cross-family transfer. Originality/value: The results establish Evasion-Hate as an auditable protocol for targeted evasion robustness. Cross-family evaluation further identifies where robustness transfers to unseen transformations and where explicit family coverage remains necessary.
- 8 October 2026
- Published by Emerald Publishing in Data Technologies and Applications
- p. 1-24
Evasion-hate: targeted augmentation for hate speech detection under prompt-induced surface-form shifts
Purpose: Online hate speech detection is increasingly used as a decision-support component in content moderation, but most detectors are evaluated on clean held-out text. In practice, users can preserve harmful intent while changing surface form through spelling obfuscation, coded wording or code-switching, creating a robustness gap between benchmark performance and deployment reliability. This paper introduces Evasion-Hate, a targeted, quality-filtered augmentation pipeline for three-way hate/offensive/normal classification in English and Vietnamese. Design/methodology/approach: The pipeline generates synthetic large language model (LLM) candidate rewrites, filters them with checks independent of the downstream classifiers under evaluation and evaluates robustness with paired bootstrap testing, matched quantity controls, competitive augmentation baselines and external functional suites. Findings: Under the main evasion protocol, Evasion-Hate achieves the strongest English robustness result among comparable three-way systems evaluated under our protocol: XLM-R robust Macro-F1 rises from 0.4783 to 0.7582 (+27.99 pp); model-specific PFR is 37.16% for the baseline and 16.66% for Evasion-Hate, while model-specific harmful-only ASR is 56.56% and 11.57%, respectively. Vietnamese models show statistically supported gains on the reconstructed synthetic evaluation sets (PhoBERT-v2 +4.60 pp; ViSoBERT +3.29 pp; both significant by pooled paired bootstrap), although the Vietnamese transform families remain below or near the pre-specified fidelity threshold and should be interpreted diagnostically. Matched quantity controls show the gain is not explained by additional training volume; external HateCheck spelling-variation flagged rate improves by 26.42 pp. Leave-one-transformation-out evaluation shows partial and family-dependent transfer. In English, holding out euphemistic and coded rewriting yields a +18.01 pp gain that remains supported after Holm correction, whereas the +8.48 pp gain for held-out code-switching does not survive multiplicity correction and the character-level family shows little change (+0.31 pp). In Vietnamese, held-out code-switching shows a positive +2.36 pp effect with a bootstrap interval excluding zero, but it does not survive Holm correction; the remaining folds and the pooled evaluation show no supported cross-family transfer. Originality/value: The results establish Evasion-Hate as an auditable protocol for targeted evasion robustness. Cross-family evaluation further identifies where robustness transfers to unseen transformations and where explicit family coverage remains necessary.
- Open Access
- 8 October 2026
- Published by Springer Nature
- p. C1
Correction to: Enhancing Knowledge Base Completion for Long-Tail Entities Through Zero-Shot Learning
- 8 October 2026
- Published by Springer Nature
- p. C1
Open AccessCorrection to: Enhancing Knowledge Base Completion for Long-Tail Entities Through Zero-Shot Learning
- Open Access
- 21 September 2026
- Published by Elsevier in Intelligence-Based Medicine
- Vol. 16
Reasoning-aware multimodal retrieval-augmented generation for evidence-grounded clinical decision support
- 21 September 2026
- Published by Elsevier in Intelligence-Based Medicine
- Vol. 16
Open AccessReasoning-aware multimodal retrieval-augmented generation for evidence-grounded clinical decision support
- 17 August 2026
- Published by Emerald Publishing in Data Technologies and Applications
- Vol. 60 (5) , 621-635
MOSAIC: a structurally robust multimodal framework for fine-grained fake news detection via optimal transport and adaptive expert routing
Purpose: Fine-grained multimodal fake news detection requires reasoning over text–image semantic inconsistency and heterogeneous inputs. Existing methods often strengthen fusion or add contextual information but rarely use cross-modal conflict to control multiple inference stages. This study examines whether token-to-patch discrepancy can guide expert processing, contextual interaction and progressive fusion. Design/methodology/approach: Multimodal optimal transport semantic alignment with integrated conflict-aware experts (MOSAIC) is a conflict-aware framework. It uses Sinkhorn-regularized optimal transport to estimate token-to-patch mismatch, derive conflict cues for expert processing and fusion, and adaptively route heterogeneous inputs. The same architecture is evaluated on complementary benchmarks FineFake and AMG; dataset-specific choices are limited to preprocessing, class-imbalance handling and auxiliary-context availability. Findings: MOSAIC achieves the strongest performance among the compared methods on both benchmarks, improving Macro-F1 by +1.44 points over LLM-MFEFND on FineFake and +2.00 points over MGCA on AMG. Ablation analysis shows a complementary specialization pattern: optimal transport-based conflict modeling contributes most under attribution-level conflict pressure on AMG, whereas adaptive expert routing is most critical under platform-heterogeneous inputs on FineFake. Multi-seed evaluation further confirms stable performance across both settings. Originality/value: This study introduces cross-modal semantic conflict as a reusable architectural control signal rather than a local alignment cue. MOSAIC integrates optimal transport-based conflict modeling, adaptive expert routing and conflict-aware progressive fusion into a unified framework, advancing fine-grained multimodal fake news detection through controlled conflict regulation rather than representation augmentation alone.
- 17 August 2026
- Published by Emerald Publishing in Data Technologies and Applications
- Vol. 60 (5) , 621-635
MOSAIC: a structurally robust multimodal framework for fine-grained fake news detection via optimal transport and adaptive expert routing
Purpose: Fine-grained multimodal fake news detection requires reasoning over text–image semantic inconsistency and heterogeneous inputs. Existing methods often strengthen fusion or add contextual information but rarely use cross-modal conflict to control multiple inference stages. This study examines whether token-to-patch discrepancy can guide expert processing, contextual interaction and progressive fusion. Design/methodology/approach: Multimodal optimal transport semantic alignment with integrated conflict-aware experts (MOSAIC) is a conflict-aware framework. It uses Sinkhorn-regularized optimal transport to estimate token-to-patch mismatch, derive conflict cues for expert processing and fusion, and adaptively route heterogeneous inputs. The same architecture is evaluated on complementary benchmarks FineFake and AMG; dataset-specific choices are limited to preprocessing, class-imbalance handling and auxiliary-context availability. Findings: MOSAIC achieves the strongest performance among the compared methods on both benchmarks, improving Macro-F1 by +1.44 points over LLM-MFEFND on FineFake and +2.00 points over MGCA on AMG. Ablation analysis shows a complementary specialization pattern: optimal transport-based conflict modeling contributes most under attribution-level conflict pressure on AMG, whereas adaptive expert routing is most critical under platform-heterogeneous inputs on FineFake. Multi-seed evaluation further confirms stable performance across both settings. Originality/value: This study introduces cross-modal semantic conflict as a reusable architectural control signal rather than a local alignment cue. MOSAIC integrates optimal transport-based conflict modeling, adaptive expert routing and conflict-aware progressive fusion into a unified framework, advancing fine-grained multimodal fake news detection through controlled conflict regulation rather than representation augmentation alone.
- Open Access
- 8 July 2026
- Published by Springer Nature in Discover Artificial Intelligence
- Vol. 6 (1)
Dominance aware multimodal rumor verification using reliability calibrated multi-expert inference
Modality imbalance undermines multimodal pattern recognition when auxiliary inputs are missing, noisy, or weakly aligned, causing fusion models to collapse toward quasi-unimodal shortcuts and produce unreliable confidence. We propose DOMINO, a dominance-aware framework that mitigates this failure through three mechanisms: (1) a closed-loop dominance controller that regulates auxiliary objectives during training, (2) a missing-safe residual fusion module that enables stable fallback under degraded modalities, and (3) reliability-aware inference via temperature scaling and weighted expert aggregation. Experiments on Weibo and MediaEval2015 demonstrate state-of-the-art performance, achieving 0.9785 macro-F1 on Weibo and 0.9559 macro-F1 on MediaEval2015, while substantially reducing expected calibration error compared with that of strong baselines. These results establish that dominance-aware control and missing-safe fusion are effective principles for robust multimodal recognition under structural modality imbalance.
- 8 July 2026
- Published by Springer Nature in Discover Artificial Intelligence
- Vol. 6 (1)
Open AccessDominance aware multimodal rumor verification using reliability calibrated multi-expert inference
Modality imbalance undermines multimodal pattern recognition when auxiliary inputs are missing, noisy, or weakly aligned, causing fusion models to collapse toward quasi-unimodal shortcuts and produce unreliable confidence. We propose DOMINO, a dominance-aware framework that mitigates this failure through three mechanisms: (1) a closed-loop dominance controller that regulates auxiliary objectives during training, (2) a missing-safe residual fusion module that enables stable fallback under degraded modalities, and (3) reliability-aware inference via temperature scaling and weighted expert aggregation. Experiments on Weibo and MediaEval2015 demonstrate state-of-the-art performance, achieving 0.9785 macro-F1 on Weibo and 0.9559 macro-F1 on MediaEval2015, while substantially reducing expected calibration error compared with that of strong baselines. These results establish that dominance-aware control and missing-safe fusion are effective principles for robust multimodal recognition under structural modality imbalance.
- Open Access
- 29 June 2026
- Published by Elsevier in ICT Express
- Vol. 12 (4) , 933-956
A Systematic Survey of Sentiment Analysis in Low-Resource Southeast Asian Languages: Resources, Methods, and Evaluation Challenges
- 29 June 2026
- Published by Elsevier in ICT Express
- Vol. 12 (4) , 933-956
Open AccessA Systematic Survey of Sentiment Analysis in Low-Resource Southeast Asian Languages: Resources, Methods, and Evaluation Challenges
- 5 April 2026
- Published by Springer Nature
- p. 233-247
Enhancing Fake News Detection Resilience Against Style Attacks: The Content-Style Invariant Detector
- 5 April 2026
- Published by Springer Nature
- p. 233-247
Enhancing Fake News Detection Resilience Against Style Attacks: The Content-Style Invariant Detector
- 18 January 2026
- Published by Springer Nature
- p. 479-487
Towards a Fact-Checking Framework Based on Large Language Models with Structured Knowledge Integration
- 18 January 2026
- Published by Springer Nature
- p. 479-487
Towards a Fact-Checking Framework Based on Large Language Models with Structured Knowledge Integration