Network automation is critical for Intelligent Internet of Things (IIoT) systems, such as industrial networks and smart cities, where managing heterogeneous devices at scale poses significant challenges. Intent-Based Networking (IBN) powered by Large Language Models (LLMs) enables natural language intent translation into executable policies, yet struggles with accuracy in few-shot In-Context Learning (ICL) due to inefficient example retrieval for multi-intent scenarios. To address this issue, we propose SubmodICL, a submodular optimization-based method that retrieves contextual examples to maximize query-guided mutual information, thereby ensuring both coverage and representativeness. Integrated with LLMs, SubmodICL significantly improves intent translation performance. We validate our approach through a use case translating flow-table operation intents into API calls, and evaluate LLMs across different parameter scales. Experiments demonstrate that SubmodICL outperforms baselines by 9.121%–12.377% in accuracy for resource-constrained settings (small LLMs, limited examples, or multi-intent tasks). The proposed method provides a generalizable framework for enhancing intent translation in network automation, with potential applicability to IIoT systems.