HomonymSenseNet: A Context-Aware Transformer Framework with Dynamic Sense Memory and Adaptive Negative Sampling for Semantic Representation Learning
VFAST Transactions on Software Engineering · 28.06.2026 · 14 (2) · 382-404
DOI: 10.21015/vtse.v14i2.2478 ↗Kurz gefasst
Fachartikel in der Zeitschrift VFAST Transactions on Software Engineering, online erschienen am 28.06.2026. Verfasst von S Subi und 1 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben.
Abstract (Originalsprache)
NLP systems sometimes experience semantic uncertainty from homonym terms with identical lexical forms but varied context-dependent meanings. Traditional embedding methods and transformer-based models struggle to differentiate semantic senses, causing contextual confusion and poor language understanding performance. HomonymSenseNet, a context-aware transformer framework integrating Dynamic Sense Memory Bank (DSMB), Context-Guided Adaptive Negative Sampling (CANS), Multi-Level Context Fusion, Contrastive Sense Separation Learning, and Rare-Sense Reinforcement Optimisation, addresses these limitations. The approach learns highly discriminative contextual embeddings while maintaining low-frequency semantic senses that are under-represented during training. The publicly accessible SemCor and WiC benchmark datasets were used to evaluate Word2Vec, GloVe, BERT, RoBERTa, and SenseBERT models. HomonymSenseNet outperformed all competing approaches on the SemCor dataset with 96.84% accuracy, 96.53% precision, 96.27% recall, and 96.40% F1-score. The suggested framework has 95.71% accuracy and 0.91 MCC on the WiC dataset. Rare-sense detection excelled with 92.88% recall and 93.15% F1-score. Qualitative t-SNE visualisation showed increased semantic cluster separation, while convergence analysis showed quicker optimisation stability. Experimental results show that HomonymSenseNet is a scalable and effective contextual semantic representation learning and homonym comprehension solution for current NLP applications.
Zitieren
APA
Subi, S., & Shanthini, B. (2026). HomonymSenseNet: A Context-Aware Transformer Framework with Dynamic Sense Memory and Adaptive Negative Sampling for Semantic Representation Learning. VFAST Transactions on Software Engineering, 14(2), 382-404. https://doi.org/10.21015/vtse.v14i2.2478
Vancouver
Subi S, Shanthini B. HomonymSenseNet: A Context-Aware Transformer Framework with Dynamic Sense Memory and Adaptive Negative Sampling for Semantic Representation Learning. VFAST Transactions on Software Engineering. 2026;14(2):382-404. doi:10.21015/vtse.v14i2.2478
BibTeX
@article{Subi2026,
author = {Subi, S and Shanthini, B.},
title = {HomonymSenseNet: A Context-Aware Transformer Framework with Dynamic Sense Memory and Adaptive Negative Sampling for Semantic Representation Learning},
journal = {VFAST Transactions on Software Engineering},
year = {2026},
volume = {14},
number = {2},
pages = {382--404},
doi = {10.21015/vtse.v14i2.2478}
}
EUVIMED