Automatically Grading Text Difficulty with Multiple Features
Yong Cheng1(),Dekuan Xu1,Xueqiang Lv2
1(School of Chinese Language and Literature, Ludong University, Yantai 264025, China) 2(School of Computer Science, Beijing University of Information Technology, Beijing 100192, China)
[Objective] This paper aims to automatically grade reading difficulty of textual documents. [Methods] We used machine learning method based on multiple features of the texts to decide their difficulty levels automatically. The features, which include word-frequency, structures, topics, and depth, describe the textual contents from different perspectives. [Results] We evaluated our method with the reading comprehension texts for high-school English exams, and achieved an accuracy of 0.88. Our result is better than those of the traditional difficulty classification methods. [Limitations] Due to the high cost of manual annotation, the existing datasets cannot be used to improve our method. [Conclusions] The proposed method increased the effectiveness of machine leanring based data analysis.
school likes happy day nice friends teacher morning eat chinese boy english mother play father lot china afternoon beautiful playing girl homework green friend lunch class tv football breakfast sports
高中
life people time university women study author researchers education college social health experience age person business human public company job american language national brain government body technology family scientists
表2 不同级别下的Top30特征词
序号
筛选特征
分级准确率
序号
筛选特征
分级准确率
13
-fun_dale_chall
0.821
4
-total_polysyllables
0.837
12
-fun_coleman_liau
0.829
14
-fun_flesch
0.838
8
-total_difficult_words
0.830
15
-fun_gunning_fog
0.838
3
-avg_letters
0.834
5
-total_syllables
0.839
11
-fun_readability
0.836
2
-avg_syllables
0.839
9
-fun_smog_index
0.837
10
-fun_kincaid
0.840
7
-total_sentences
0.837
6
-total_words
0.841
1
-avg_sentence_len
0.837
无
0.841
表3 筛选不同结构特征后的性能比较
图4 初中和高中文本的主题分布与相应主题词
状态向量维度
卷积网络窗口数
维度
分级正确率
数目
分级正确率
32
0.867
2
0.870
64
0.880
3
0.869
128
0.869
4
0.871
192
0.871
5
0.880
256
0.873
6
0.878
表4 不同超参数对网络性能的影响
图5 单类型特征在多分类器下的比较结果
特征数目
分级前融合
分级后融合
开发集
测试集
开发集
测试集
单元特征
F
0.850
0.833
0.850
0.833
S
0.843
0.845
0.843
0.845
T
0.815
0.816
0.815
0.816
M
0.880
0.870
0.880
0.870
二元特征 融合
F&S
0.881
0.874
0.875
0.860
F&T
0.866
0.843
0.846
0.837
F&M
0.886
0.877
0.886
0.878
S&T
0.871
0.856
0.859
0.848
S&M
0.883
0.878
0.886
0.879
T&M
0.882
0.876
0.883
0.873
三元特征 融合
F&S&T
0.884
0.877
0.874
0.846
F&S&M
0.887
0.878
0.881
0.875
F&T&M
0.885
0.878
0.878
0.873
S&T&M
0.884
0.877
0.884
0.875
四元特征 融合
F&S&T&M
0.888
0.880
0.888
0.871
表5 多元特征融合实验结果
对比方法
正确率(校验集)
正确率(测试集)
Random Guess
0.494
0.491
FKGL
0.706
0.709
VM_KNN
0.799
0.780
CNN_SC
0.852
0.863
Our Model
0.888
0.880
表6 与现有方法的比较结果
图6 文本分级识别结果
[1]
郭利敏 . 基于卷积神经网络的文献自动分类研究[J]. 图书与情报, 2017(6):96-103. ( Guo Limin . Study of Automatic Classification of Literature Based on Convolution Neural Network[J]. Library and Information, 2017(6):96-103.)
[2]
李慧宗, 胡学钢, 杨恒宇 , 等. 基于LDA的社会化标签综合聚类方法[J]. 情报学报, 2015,34(2):146-155. ( Li Huizong, Hu Xuegang, Yang Hengyu , et al. A Comprehensive Clustering Method for Socialized Label Based on LDA[J]. Journal of the China Society for Scientific and Technical Information, 2015,34(2):146-155.)
[3]
徐彤阳, 尹凯 . 大数据背景下微博语义检索[J]. 情报杂志, 2017,36(12):173-179. ( Xu Tongyang, Yin Kai . Semantic Retrieval of Microblogging in the Background of Large Data[J]. Journal of Intelligence, 2017,36(12):173-179.)
[4]
Bear D, Dole J, Echevarria J , et al. Treasures, A Reading/ Language Arts Program[M]. McGraw-Hill Education, 2009.
[5]
Lester M, Neal S, Royster J , et al. Glencoe Writer’s Choice: Grammar and Composition[M]. McGraw-Hill Education, 2001.
[6]
李欣 . 美国中小学生阅读分级研究[D]. 上海: 华东师范大学, 2016. ( Li Xin . Research on the American Leveled Reading of K-12 Students[D]. Shanghai: East China Normal University, 2016.)
[7]
Kincaid J P, Braby R, Mears J E . Electronic Authoring and Delivery of Technical Information[J]. Journal of Instructional Development, 1988,11(2):8-13.
[8]
Dale E, Chall J S . A Formula for Predicting Readability[J]. Journal of Educational Research Bulletin, 1948,27(2):37-54.
[9]
McLaughlin G H . SMOG Grading: A New Readability Formula[J]. Journal of Reading, 1969,12(8):639-646.
[10]
Graesser A C , McNamara D S, Louwerse M M, et al. Coh-Metrix: Analysis of Text on Cohesion and Language[J]. Journal of Behavior Research Methods, Instruments, & Computers, 2004,36(2):193-202.
[11]
张宁志 . 汉语教材语料难度的定量分析[J]. 世界汉语教学, 2000(3):83-88. ( Zhang Ningzhi . Quantitative Analysis of Corpora Difficulty in Chinese Textbooks[J]. Chinese Teaching in the World, 2000(3):83-88.)
[12]
郭望皓 . 对外汉语文本易读性公式研究[D]. 上海: 上海交通大学, 2009. ( Guo Wanghao . Research on Readability Formula of Chinese Text for Foreign Students[D]. Shanghai: Shanghai Jiao Tong University, 2009.)
[13]
左虹, 朱勇 . 中级欧美留学生汉语文本可读性公式研究[J]. 世界汉语教学, 2014,28(2):263-276. ( Zuo Hong, Zhu Yong . Research on Chinese Readability Formula of Texts for Intermediate Level European and American Students[J]. Chinese Teaching in the World, 2014,28(2):263-276.)
[14]
Salton G, Buckley C . Term-Weighting Approaches in Automatic Text Retrieval[J]. Information Processing & Management, 1988,24(5):513-523.
[15]
Hofmann T . Unsupervised Learning by Probabilistic Latent Semantic Analysis[J]. Machine Learning, 2001,42(1-2):177-196.
[16]
Blei D M, Ng A Y, Jordan M I , et al. Latent Dirichlet Allocation[J]. Journal of Machine Learning Research, 2003,3:993-1022.
[17]
Walker S H, Duncan D B . Estimation of the Probability of an Event as a Function of Several Independent Variables[J]. Biometrika, 1967,54(1-2):167-179.
[18]
Cortes C, Vapnik V . Support-Vector Networks[J]. Machine Learning, 1995,20(3):273-297.
[19]
Ho T K. Random Decision Forests [C]// Proceedings of the 3rd International Conference on Document Analysis and Recognition. IEEE, 1995: 278-282.
[20]
Liu P, Qiu X, Huang X , et al. Recurrent Neural Network for Text Classification with Multi-Task Learning[C]// Proceedings of the 25th International Joint Conferences on Artificial Intelligence. AAAI Press, 2016: 2873-2879.
[21]
Kim Y. Convolutional Neural Networks for Sentence Classification [C]// Proceedings of the 2014 International Conference on Empirical Methods on Natural Language Processing. ACL, 2014: 1746-1751.
[22]
Senter R J, Smith E A . Automated Readability Index[J]. Journal of Competitor New York, 1967,1:1-14.
[23]
Gunning R . The Fog Index After Twenty Years[J]. Journal of Business Communication, 1969,6(2):3-13.
[24]
Coleman M, Liau T L . A Computer Readability Formula Designed for Machine Scoring[J]. Journal of Applied Psychology, 1975,60(2):283-284.
[25]
Graves A, Schmidhuber J . Framewise Phoneme Classification with Bidirectional LSTM and Other Neural Network Architectures[J]. Neural Networks, 2005,18(5):602-610.
[26]
Lai G, Xie Q, Liu H, et al. Race: Large-Scale Reading Comprehension Dataset from Examinations [C]// Proceedings of the 2017 International Conference on Empirical Methods on Natural Language Processing. ACL, 2017: 785-794.
[27]
TensorFlow[CP]. [2018-08-24]..
[28]
蒋晶晶 . CEPT阅读文本易读度分析及词汇检测工具的开发[D]. 长沙: 湖南大学, 2009. ( Jiang Jingjing . Readability Analysis on CEPT Reading Texts and the Development of Lexical Checker[D]. Changsha: Hunan University, 2009.)
[29]
陈炎龙, 张志明 . 基于向量空间模型的英文文本难度判定[J]. 电脑知识与技术, 2010,6(12):2994-2996. ( Chen Yanlong, Zhang Zhiming . The English Text Difficulty Measurement Based Vector Space Model[J]. Computer Knowledge and Technology, 2010,6(12):2994-2996.)
[30]
Maaten L, Hinton G . Visualizing Data Using t-SNE[J]. Journal of Machine Learning Research, 2008,9(11):2579-2605.