摘要
汉语二语文本难度自动分级任务是国际中文教育与计算语言学领域中的一个重要主题。本文依据《国际中文教育中文水平等级标准》,提出了基于语法点多样性与复杂性的25个语法点特征并实现了相关特征的自动抽取与计算,在此基础上构建了自动分级模型。实验结果表明,融合语法点特征后多元逻辑回归算法的分级准确率为86.40%,比基于现有语言特征的实验提升了2.4%。进一步研究发现,六级语法点多样性、语法点难度等级均值是区别文章难度级别的关键特征。此外,本文将包含语法点特征在内的207项语言特征融入基于BERT的深度学习模型,取得了87.6%的准确率,超过了基于传统语言特征的方法和基于神经网络的方法。
Automatic Chinese L2 texts grading is an essential topic in international Chinese language education and computational linguistics.This paper proposes 25 grammatical structure features based on the diversity and complexity of grammatical structure,and implements the automatic extraction and calculation of the relevant features,based on which an automatic grading model is constructed.The results show that the accuracy of the model is 86.40%,which is 2.4%higher than that of the existing language features.It was further found that Grade 6 grammar structure diversity and grammar structure difficulty level means were the most relevant features to distinguish the difficulty level of the texts.In addition,the paper incorporated 207 linguistic features including grammar structure features into the BERT-based deep learning model and achieved an accuracy of 87.6%,outperforming both the traditional linguistic feature-based and neural network-based methods.
作者
朱君辉
刘鑫
杨麟儿
王鸿滨
杨尔弘
ZHU Junhui;LIU Xin;YANG Liner;WANG Hongbin;YANG Erhong
出处
《语言文字应用》
CSSCI
北大核心
2022年第3期87-99,共13页
Applied Linguistics
基金
国家语委项目“智能辅助汉语应用文写作研究”(ZDI135-131)
国家社科基金重点项目“汉语作为第二语言学习者阅读素养评价标准的构建与测评研究”(20AYY016)
北京语言大学研究生创新基金项目“国际中文教育阅读文本自动分级”(22YCX114)资助。
