Using Large Language Models to Enhance the Reusability of Sensor Data

Abstract

The Internet of Things generates vast data volumes via diverse sensors, yet its potential remains unexploited for innovative data-driven products and services. Limitations arise from sensor-dependent data handling by manufacturers and user companies, hindering third-party access and comprehension. Initiatives like the European Data Act aim to enable high-quality access to sensor-generated data by regulating accuracy, completeness, and relevance while respecting intellectual property rights. Despite data availability, interoperability challenges impede sensor data reusability. For instance, sensor data shared in HTML formats requires an intricate, time-consuming processing to attain reusable formats like JSON or XML. This study introduces a methodology aimed at converting raw sensor data extracted from web portals into structured formats, thereby enhancing data reusability. The approach utilises large language models to derive structured formats from sensor data initially presented in non-interoperable formats. The effectiveness of these language models was assessed through quantitative and qualitative evaluations in a use case involving meteorological data. In the proposed experiments, GPT-4, the best performing LLM tested, demonstrated the feasibility of this methodology, achieving a precision of 93.51% and a recall of 85.33% in converting HTML to JSON/XML, thus confirming its potential in obtaining reusable sensor data.

Leveraging Large Language Models for Sensor Data Retrieval

“…As described in Section 3, the system comprises two primary components: data gathering and data retrieval. Concerning the data gathering component, prior work by the authors [49] evaluated various LLMs, including ChatGPT and Llama 2, to assess their performance in transforming raw sensor data into a structured format. The highest precision and recall achieved were 93.51% and 85.33%, respectively, demonstrating their suitability for the task.…”

“…Subsequently, an LLM (GPT-3.5 Turbo from OpenAI) was employed to transform these datasets into CSV format, as described in Ref. [49]. The average number of tokens in each input file was 98,885.87, so most of the files to be transformed by the LLM had to be split into several files that met the maximum input tokens of the model (16,385 tokens in the case of GPT-3.5 Turbo).…”

Towards a potential paradigm shift in health data collection and analysis: Contemporary challenges of Human-Machine interaction

“…Still, it is too early to make assumptions about LLM performance in the healthcare sphere. The well-known problem of LLM hallucination [39] limits not only the reliability and trustworthiness of the generated results but also creates a relatively new problem of responsibility for LLM-generated mistakes, wrong advice and externally coherent but medically meaningless solutions. GPT-4 demonstrated high interpretative abilities but showed reasoning inconsistencies, while Falcon and LLaMA 2 reached significant accuracy yet with insufficient explanatory reasoning [40].…”