NOEM$^{3}$A: a Neuro-symbolic Ontology-Enhanced Method for Multi-intent understanding in Mobile Agents

The article presents NOEM$^{3}$A, a neuro-symbolic framework designed to enhance multi-intent understanding in mobile agents by integrating a lightweight intent ontology with compact language models like TinyLlama and Llama-3.2-3B. This approach employs a token-level decoding prior and Semantic Intent Similarity (SIS) to improve intent prediction accuracy while maintaining low latency and privacy. Experimental results on the MultiWOZ 2.3 dataset demonstrate significant improvements in performance metrics such as exact match and Slot-F1, indicating that symbolic alignment can effectively enhance on-device natural language understanding (NLU).