An Intra-lingual Stylistic Comparative Research for Function Evaluation of Large Language Models

The functional evaluation of large language models encompasses aspects such as semantic comprehension, reasoning, text generation, and translation. To provide methodological references for the functional evaluation of different large language models, five models including ERNIE 4.0 Turbo were selected. Initially, the same instruction was issued to these five models via prompt engineering to generate five English texts with similar styles, themes, and word counts. Subsequently, the meta-function theory of systemic functional linguistics was employed to analyze the differences in transitivity processes, mood types, and theme types among these texts, and a comparison was made with a reference text to explore whether stylistic differences exist and their technical motivations. The results revealed that, in general, the generated texts from domestic large language models closely approximate the functional semantic configurations of the original English reference samples, yet there are still hierarchical differences. This is because different models vary in their ability to balance computational power with interpretability, whereas foreign large language models can achieve rapid convergence in small-sample reasoning. Stylistic contrastive research based on semantic configuration features is conducive to comparing the functions of different large language models and optimizing their implementation technologies, thereby fostering customer-friendly technological models.