摘要
跨提示中文作文自动评分任务是真实教育环境中广泛应用的方法,然而,在使用预训练模型实现的过程中主要涉及两个问题,一是不同提示提取的文本特征差异,限制模型在其他提示上的泛化能力;二是传统BERT预训练模型的分词方法未能充分考虑中文分词的需求。针对这些问题,一方面,使用基于BERT模型的多尺度文本表示方法,从文章、段落和标记三个不同尺度来提取文本特征,增强提取通用文本特征的能力;另一方面,使用中文BERT-wwm模型来增强模型对中文语义的理解能力,该模型采用全词掩码的预训练方式,解决BERT模型难以理解复杂中文语法和逻辑的问题。实验结果表明,在自制数据集中,模型的预测得分与真实人工评分具有较高的一致性,达到0.736,并在文本特征高度差异化的题材中表现良好。
Automated essay scoring across different prompts in Chinese compositions is a widely applied method in real educational environments.However,in the process of using pretrained models,two main challenges arise:first,the variation in text features extracted from different prompts limits the model's ability to generalize across other prompts;second,traditional BERT pretrained models'tokenization methods inadequately address the needs of Chinese word segmentation.To tackle these issues,on one hand,a multi-scale text representation method based on BERT models is employed.This method extracts text features from articles,paragraphs,and tokens to enhance the model's capability to capture universal text features.On the other hand,the Chinese BERT-wwm model is used to enhance the model's understanding of Chinese semantics.This model adopts a full-word masking pretraining approach to address the issues of traditional BERT models struggling with understanding complex Chinese grammar and logic.Experimental results demonstrate promising outcomes.On our proprietary dataset,the model achieves high consistency with human-scored evaluations,reaching 0.736,and performs well in subjects with highly differentiated text features.
作者
赵国良
陈亮
王珺琳
ZHAO Guoliang;CHEN Liang;WANG Junlin(Shenyang Ligong University,Shenyang 110159,China)
出处
《通信与信息技术》
2025年第1期114-117,共4页
Communication & Information Technology
