Citation

Farhan, Yasir Hadi and Tareq, Mustafa and Shannaq, Boumedyen and AlMaqbali, Said and Ali, Oualid (2025) Enhancing Robustness and Scalability in AI-Powered Natural Language Processing: A Novel Hybrid Deep Symbolic Framework. In: 3rd International Conference on Cyber Resilience, ICCR 2025, 3 July 2025 - 4 July 2025, Dubai.

Text 74.pdf - Published Version Restricted to Repository staff only Download (882kB)

Abstract

With transformer-based architectures, such as ChatGPT, and specialized systems like DeepSeek, AI-driven Natural Language Processing (NLP) models have made outstanding strides. Problems such as hallucination, interpretability issues, noise sensitivity, and computational inefficiency at large scale still exist, despite their successes. In this work, we present a deep symbolic NLP framework that combines transformer designs with a knowledge graph and logical inference-based symbolic reasoning. To further enhance semantic understanding and noise robustness, we have incorporated embedding-based semantic augmentation methods inspired by recent advancements in Arabic document retrieval using deep median networks and word embeddings. Based on confidence and context, the adaptive fusion layer dynamically balances neural and symbolic components, taking conceptual inspiration from swarm intelligence applied in vehicular networks. We thoroughly test the system on a variety of benchmark datasets, including SQuAD 2.0, CommonsenseQA, DailyDialog, WMT 2022 English German, and TruthfulQA, and compare it to various baselines, such as BERT-large, RoBERTa, DeBERTa, T5, ChatGPT, and DeepSeek. To enhance the trustworthiness of these kinds of insights the hybrid summarization framework was proposed and tested on the SQuAD 2.0 dataset. The best performance was recorded by the proposed model with the ROUGE-L score of 0.56 and the BERTScore of 0.97, surpassing the results of ChatGPT (0.43/0.89), DeepSeek (0.47/0.91), T5-large (0.49/0.92), and BERT-large (0.44/0.90). This shows categorical betterment in both the accuracy of summaries and semantic relevance.

Item Type:	Conference or Workshop Item (Paper)
Uncontrolled Keywords:	Hybrid neural-symbolic NLP, knowledge graph reasoning
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	18 Mar 2026 07:55
Last Modified:	18 Mar 2026 07:55
URII:	http://shdl.mmu.edu.my/id/eprint/15563

Downloads

Edit (login required)