Towards Zero-shot Question Answering in CPS-IoT: Large Language Models and Knowledge Graphs

Ozan Baris Mulayim, Mario Berges

Abstract:

Natural language provides an intuitive interface for querying data, yet its unstructured nature often makes precise retrieval of information challenging. Knowledge graphs (KGs), with their structured and relational representations, offer a powerful solution to structuring knowledge, while large language models (LLMs) are capable of interpreting user intent through language. This combination of KGs and LLMs has been explored extensively for Knowledge Graph Question Answering (KGQA), primarily for open-domain or encyclopedic knowledge. Domain-specific KGQA, instead, presents significant opportunities for Cyber-Physical Systems (CPS) and the Internet of Things (IoT), where the extraction of structured metadata is essential for automation and scalability of control and analytics applications.
In this work, we evaluate and improve AutoKGQA, a domain-independent KGQA framework that utilizes LLMs to generate structured queries. Through a case study on KGs of sensor data from buildings, we assess its ability to retrieve time series identifiers, which are a requirement for extracting time series data from large sensory databases. Our results demonstrate that while AutoKGQA performs well in certain cases, its domain-agnostic approach leads to systematic failures particularly in complex queries requiring implicit knowledge. We show that domain-specific prompting significantly enhances query accuracy, allowing even smaller LLMs to perform on par with larger ones. These findings highlight the impact of domain-adapted prompting in KGQA (DA-KGQA) and suggest a path toward more efficient, scalable, and interpretable AI-driven metadata retrieval for CPS-IoT applications.