The Impact of Temperature on Extracting Information From Clinical Trial Publications Using Large Language Models

Abstract

Introduction The application of natural language processing (NLP) for extracting data from biomedical research has gained momentum with the advent of large language models (LLMs). However, the effect of different LLM parameters, such as temperature settings, on biomedical text mining remains underexplored and a consensus on what settings can be considered "safe" is missing. This study evaluates the impact of temperature settings on LLM performance for a named entity recognition and a classification task in clinical trial publications. Methods Two datasets were analyzed using GPT-4o and GPT-4o-mini models at nine different temperature settings (0.00-2.00). The models were used to extract the number of randomized participants and classify abstracts as randomized controlled trials (RCTs) and/or as oncology-related. Different performance metrics were calculated for each temperature setting and task. Results Both models provided correctly formatted predictions for more than 98.7% of abstracts across temperatures from 0.00 to 1.50. While the number of correctly formatted predictions started to decrease afterward with the most notable drop between temperatures 1.75 and 2.00, the other performance metrics remained largely stable. Conclusion Temperature settings at or below 1.50 yielded consistent performance across text-mining tasks, with performance declines at higher settings. These findings are aligned with research on different temperature settings for other tasks, suggesting stable performance within a controlled temperature range across various NLP applications.

Ähnliche Arbeiten