A Lightweight Neuro-Symbolic Pattern Recognition Framework for Edge Intelligence via NLP-Based Knowledge Graphs

Edge intelligence is a core enabler for next-generation IoT and 6G networks, which demands processing massive heterogeneous data under stringent latency, energy and computational constraints. Deploying heavyweight deep learning models on resource-limited edge devices remains a key challenge, especially for pattern recognition tasks supporting dynamic resource scheduling and intelligent decision-making. To tackle this, we propose a lightweight neuro-symbolic pattern recognition framework for edge intelligence, integrating NLP-based entity/relation recognition with symbolic knowledge graph reasoning. We construct an edge computing domain-specific knowledge graph via two lightweight models: a BERT-BiLSTM-CRF for named entity recognition (NER) and a PCNNATT- BiLSTM for relation extraction, both balanced in accuracy and computational efficiency for edge inference. Extracted structured knowledge forms a neuro-symbolic knowledge graph, enabling lowoverhead semantic reasoning for resource localization, bottleneck analysis and decision support, thus reducing reliance on compute-intensive end-to-end deep models. Experiments show the BERTBiLSTM- CRF achieves a 91.4% F1-score for NER (11.1% higher than baseline BiLSTM-CRF), and the PCNN-ATT-BiLSTM attains 82.1% F1-score for relation extraction (15.3% higher than baseline CNN). Retaining edge-deployment efficiency while enhancing pattern recognition performance, the framework offers an energy-efficient lightweight solution for edge intelligence and new insights for fusing pattern recognition with neuro-symbolic reasoning in intelligent edge computing.