Using Large Languge Models for Processing Sensor Data

Abstract

The wide availability of sensor data stored in multiple formats makes it difficult to reuse in other applications. We consider the problem of extracting sensor data from unstructured and semi-structured texts using Large Language Models. With careful prompt crafting, we have been able to establish a strict JSON structure which can be further processed with automated ease. We establish a workflow that enables the extraction of data using GPT-4, Llama 3, Mistral and Falcon models, and we show that while the closed-source GPT-4 model is generally leading in conversion efficiency, other open-source models can follow this if given appropriate data structures. We define new measures to simplify the comparison, and we present a multi-purpose workflow for sensor data extraction. We observe that some of the smaller models are incapable of correctly extracting data from freeform text but are skilled in processing tabular data. On the other hand, larger models are more robust and avoid conversion mistakes more easily.