Impact of Large–Model Temperature on Generative Entity and Attribute Extraction Results

This paper investigates the impact of temperature parameters in large language models on entity extraction and attribute extraction results. Through experimental research, this study analyzes the performance of large models in entity extraction and attribute extraction tasks under various temperature settings, with a primary focus on the impact of temperature parameters on the accuracy, completeness, and consistency of data extraction. The results show that, overall, general large models perform better in entity extraction tasks than in attribute extraction. The possible reason is that entity extraction tasks have clear entity boundaries and concentrated data distribution, which are highly compatible with the model's feature capture capabilities. In contrast, attribute extraction is characterized by complex semantics, deep reasoning requirements, obvious long-tail effects of data, and a lack of targeted model architecture, making it difficult for the model to extract attributes accurately. In the experimental results of entity extraction using large models, temperature has a minimal effect on the accuracy and completeness of entity extraction, although fluctuations are observed. In the experimental results of attribute extraction by large models, when the temperature of the large model is 0, the accuracy and completeness perform best. With the increase of the large model temperature, the overall trend of accuracy first rises and then falls. The study on the performance of different model parameters at different temperatures reflects the synergistic effect between “version iteration optimization” and “parameter scale dividend” in large language models. In terms of versions, Llama3.1 exhibits better accuracy and completeness in entity extraction and attribute extraction tasks at low temperatures <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${(T &lt; 1)}$</tex>. Under high temperatures <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${(T\geq 1)}$</tex>, Llama3-70b achieves the best accuracy in entity extraction tasks, while Llama3.2 exhibits better accuracy and completeness in attribute extraction tasks. In terms of model parameters, overall, large-parameter models have stronger entity representation capabilities and deeper network structures. With deep context modeling and a sharp logit distribution, the output results are more accurate, complete, and stable. In terms of multi-round extraction consistency, the results at low temperatures exhibit high similarity and good consistency but a weak ability to capture dynamic information. In contrast, the results at high temperatures show low similarity and poor consistency.