Atkinson John, Palma Diego
AI Empowered, Santiago, Chile.
Sci Rep. 2025 Apr 25;15(1):14551. doi: 10.1038/s41598-025-87862-3.
Automated Essay Scoring systems have traditionally relied on shallow lexical data, such as word frequency and sentence length, to assess essays. However, these approaches neglect crucial factors like text structure and semantics, resulting in limited evaluations of coherence and quality. To address these limitations, we propose a hybrid approach to AES that combines multiple features from different linguistic levels. By leveraging the complementary nature of these features, our model captures the intricate relationships underlying coherent texts. Through extensive experimentation using standard essay datasets, we demonstrate that our large language model based hybrid model surpasses state-of-the-art methods based on shallow features and pure neural networks. This research represents a significant advancement towards the development of an accurate and effective tool for assessing student writing.
自动作文评分系统传统上依赖于浅层词汇数据，如词频和句子长度，来评估作文。然而，这些方法忽略了诸如文本结构和语义等关键因素，导致对连贯性和质量的评估有限。为了解决这些局限性，我们提出了一种用于自动作文评分的混合方法，该方法结合了来自不同语言层面的多个特征。通过利用这些特征的互补性，我们的模型捕捉到连贯文本背后的复杂关系。通过使用标准作文数据集进行广泛实验，我们证明了基于大语言模型的混合模型优于基于浅层特征和纯神经网络的现有方法。这项研究代表了朝着开发一种准确有效的学生写作评估工具迈出的重要一步。
