EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Neuro-Symbolic Graph Transformers with Ontology-Guided Attention for Context-Aware Social Discourse Analysis

Wyssem Fathallah, Abdelweheb Gueddes, Mourad Mars, Mohammed J. AlGhamdi, Mohamed Ali Mahjoub, Hedi Sakli

Technologies · 06.11.2026 · 14 (11) · 734

DOI: 10.3390/technologies14110734 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Technologies, online erschienen am 06.11.2026. Verfasst von Wyssem Fathallah und 5 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Umm al-Qura University.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Social media messages are short and noisy: the same tokens can denote a mention, a follow, or a reply depending on the surrounding user graph. Pre-trained language models frequently propose relations that a formal domain ontology forbids, whereas graph embeddings largely ignore message wording. This paper presents HyKGA (Hybrid Knowledge-Graph Attention), a single-pass neuro-symbolic encoder that (i) injects k-hop, time-decayed graph neighbors and multi-scale centrality encodings at tokenization under a strictly causal, target-edge-excised protocol, (ii) compiles the type axioms and the irreflexivity and cardinality constraints of a Description Logic ontology into two pre-Softmax masks, on self-attention and on the relation classifier, and (iii) aligns text and graph representations with an InfoNCE contrastive objective. On the Twitter Social Alignment Benchmark (TSAB), HyKGA reaches an F1 score of 0.91 and a filtered link-prediction MRR of 0.574, 5 points above QLoRA-fine-tuned LLaMA-3-8B and a late-fusion BERT–GraphSAGE hybrid, with an Ontological Violation Rate of 2.0% on relational projections (1.6% on active positive predictions and 1.0% across all thresholded predictions) against 14–35% (10.2–15.8% on positive predictions and 6.7–9.9% across all thresholded predictions) for neural baselines. Controlled ablations show that pre-Softmax masking adds 3.8 F1 points over post hoc ontology filtering. Gains of 3 to 4 macro-F1 points carry over to RumourEval, PHEME and Twitter15 as task transfer. When 20% of the axioms are removed, F1 degrades gracefully to 0.88. End-to-end latency is 18.7 ms per message, 7.6 to 11.7 times lower than retrieval-augmented pipelines on the same GPU.

Förderung

  • Umm al-Qura University

Zitieren

APA

Fathallah, W., Gueddes, A., Mars, M., AlGhamdi, M. J., Mahjoub, M. A., & Sakli, H. (2026). Neuro-Symbolic Graph Transformers with Ontology-Guided Attention for Context-Aware Social Discourse Analysis. Technologies, 14(11), 734. https://doi.org/10.3390/technologies14110734

Vancouver

Fathallah W, Gueddes A, Mars M, AlGhamdi MJ, Mahjoub MA, Sakli H. Neuro-Symbolic Graph Transformers with Ontology-Guided Attention for Context-Aware Social Discourse Analysis. Technologies. 2026;14(11):734. doi:10.3390/technologies14110734

BibTeX
@article{Fathallah2026,
  author = {Fathallah, Wyssem and Gueddes, Abdelweheb and Mars, Mourad and AlGhamdi, Mohammed J. and Mahjoub, Mohamed Ali and Sakli, Hedi},
  title = {Neuro-Symbolic Graph Transformers with Ontology-Guided Attention for Context-Aware Social Discourse Analysis},
  journal = {Technologies},
  year = {2026},
  volume = {14},
  number = {11},
  pages = {734},
  doi = {10.3390/technologies14110734}
}
Kostenlos starten →