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

NOEM³A is a lightweight neuro-symbolic framework that enhances compact language models with intent ontologies to improve natural language understanding for mobile agents. By injecting structured symbolic knowledge into both input prompts and output decoding, the method achieves better performance on dialogue understanding tasks while maintaining privacy and low-latency requirements suitable for on-device deployment.

NOEM³A addresses a fundamental challenge in deploying NLU systems on resource-constrained mobile devices: scaling language models is computationally expensive and inefficient when the primary bottleneck is mapping natural language to specific executable intents. Rather than increasing model size, the researchers introduce a hybrid approach that combines neural networks with symbolic knowledge structures, demonstrating that structural constraints can compensate for model capacity limitations.

The technical innovation centers on augmenting inference with ontology-aware mechanisms. The system retrieves relevant intent categories for each query, incorporates candidate action labels directly into prompts, and applies token-level decoding constraints that steer the model toward valid outputs. This dual-layer approach—injecting symbolic structure at both input and output—creates a tighter alignment between natural language semantics and executable actions. The Semantic Intent Similarity metric further refines evaluation by accounting for hierarchical relationships in the ontology rather than treating all mismatches equally.

For developers building mobile applications, this work validates that neuro-symbolic architectures can achieve better accuracy than scaling alone while maintaining privacy and responsiveness critical for local inference. Experiments on MultiWOZ 2.3 show consistent improvements across multiple metrics for both TinyLlama and Llama-3.2-3B. This has immediate implications for voice assistants, smart home systems, and on-device AI products where latency and privacy are non-negotiable constraints. The approach suggests that domain-specific ontologies could become as valuable as model parameters in production NLU systems.

Future development should focus on ontology design methodology and how these techniques scale across diverse domains and languages.

• →Neuro-symbolic integration with intent ontologies improves NLU accuracy on mobile devices without increasing model size.
• →Ontology-augmented decoding constrains outputs to valid action labels, reducing hallucinations and invalid intent predictions.
• →The method achieves measurable improvements on TinyLlama and Llama-3.2-3B models for dialogue understanding tasks.
• →On-device NLU with symbolic constraints enables privacy-preserving AI without cloud-based processing dependencies.
• →Hierarchy-aware semantic similarity metrics better capture intent relationships than lexical matching alone.