融合TextRank算法的中文短文本相似度计算

doi:10.16180/j.cnki.issn1007-7820.2020.10.009

摘要/Abstract

摘要：

传统的VSM向量空间模型忽略了文本语义,构建的文本特征矩阵具有稀疏性。基于深度学习词向量技术,文中提出一种融合改进TextRank算法的相似度计算方法。该方法利用词向量嵌入的技术来构建文本向量空间,使得构建的向量空间模型具有了语义相关性,同时采用改进的TextRank算法提取文本关键字,增强了文本特征的表达并消除了大量冗余信息,降低了文本特征矩阵的稀疏性,使文本相似度的计算更加高效。不同模型的仿真实验结果表明,融合改进的TextRank算法与Bert词向量技术的方法具有更好的文本相似度计算性能。

关键词: 文本相似度, 提取, TextRank算法, Bert, 词向量技术, 向量空间模型

Abstract:

The traditional VSM vector space model often ignores text semantics, and the constructes text feature matrix is sparse. Based on the word vector technology of deep learning, this paper proposes a similarity calculation method that integrates the improved TextRank algorithm. This method uses the word vector embedding technology to build a text vector space, which makes the vector space model possess the semantic relevance. At the same time, with the improved TextRank algorithm to extract text keywords, the expression of text feature is enhanced and a large amount of redundant information is eliminated. The text characteristic of sparse matrix is reduced, which makes the text similarity computing more efficient. The results of the simulation experiments of different models show that the fusion of the improved TextRank algorithm with Bert word vector technology have better performance of text similarity calculation.

Key words: text similarity, extraction, TextRank slgorithm, Bert, word vector technique, vector space model

中图分类号:

TP391

卢佳伟,陈玮,尹钟. 融合TextRank算法的中文短文本相似度计算[J]. 电子科技, 2020, 33(10): 51-56.

LU Jiawei,CHEN Wei,YIN Zhong. Chinese Short Text Similarity Calculation Based on TextRank Algorithm[J]. Electronic Science and Technology, 2020, 33(10): 51-56.

图/表 9

图1

图2

图3

图4

图5

表1

表2

表3

表4

参考文献 16

[1]	Xu L H, Sun S T, Wang Q. Text similarity The algorithm based on semantic vector space model[C]. Okayama:The Fifteenth International Conference on Computer and Information Science, 2016.
[2]	牛永洁, 田成龙. 融合多因素的TFIDF关键词提取算法研究[J]. 计算机技术与发展, 2019(7):80-83.
	Niu Yongjie, Tian Chenglong. Research on TFIDF keyword extration algorithm based on multiple factors[J]. Computer Technology and Development, 2019(7):80-83.
[3]	Pu H, Fei G, Zhao H, et al. Short text similarity calculation using semantic information[C]. Orlando:International Conference on Big Data Computing & Communications, 2017.
[4]	宋冬云, 郑瑾, 张祖平. 基于混合策略的中文短文本相似度计算[J]. 计算机工程与应用, 2018,54(12):116-120,205.
	Song Dongyun, Zheng Jin, Zhang Zuping. Chinese short text similarity computation based on hybrid strategy[J]. Computer Engineering and Applications, 2018,54(12):116-120,205.
[5]	谷重阳, 徐浩煜, 周晗, 等. 基于词汇语义信息的文本相似度计算[J]. 计算机应用研究, 2018,35(2):391-395.
	Gu Chongyang, Xu Haoyu, Zhou Han, et al. Text similarity computing based on lexical semantic information[J]. Application Research of Computers, 2018,35(2):391-395.
[6]	李小涛, 游树娟, 陈维. 一种基于词义向量模型的词语语义相似度算法[J/OL].(2018-03-12) [2019-06-02] https://doi.org/10.16383/j.aas.c180312.
	Li Xiaotao, You Shujuan, Chen Wei. An algorithm of semantic similarity between words based on word single-meaning embedding model[J/OL].(2018-03-12) [2019-06-02] https://doi.org/10.16383/j.aas.c180312.
[7]	Zhang C, Wang X, Yu S, et al. Research on keyword extraction of Word2vec model in Chinese corpus[C]. Singapore:The Seventeenth International Conference on Computer and Information Science, 2018.
[8]	Wen Y, Yuan H, Zhang P. Research on keyword extraction based on Word2vec weighted TextRank[C]. Wuhan: The Secend IEEE International Conference on Computer and Communications, 2016.
[9]	Devlin J, Chang M W, Lee K, et al. BERT: Pre-training of deep bidirectional transformers for language understanding[EB/OL].(2018-10-11)[2019-06-02] https://arxiv.org/abs/1810.04805.
[10]	周锦章, 崔晓晖. 基于词向量与TextRank的关键词提取方法[J]. 计算机应用研究, 2019,36(4):1051-1054.
	Zhou Jinzhang, Cui Xiaohui. Keyword extraction method based on word vector and TextRank[J]. Application research of computers, 2019,36(4):1051-1054.
[11]	柳林青, 余瀚, 费宁, 等. 一种基于TextRank的单文本关键字提取算法[J]. 计算机应用研究, 2018,35(3):705-710.
	Liu Linqin, Yu Han, Fei Ning, et al. Key-word extracting algorithm from single text based on TextRank[J]. Application Research of Computers, 2018,35(3):705-710.
[12]	王贵新, 郑孝宗, 张浩然, 等. 基于Word2vec的短信向量化算法[J]. 电子科技, 2016,29(4):49-52.
	Wang Guixin, Zheng Xiaozong, Zhang Haoran, et al. An algorithm for vectoring SMS based on Word2vec[J]. Electronic Science and Technology, 2016,29(4):49-52.
[13]	杨飘, 董文永. 基于余弦距基于BERT嵌入的中文命名实体识别方法离的多目标粒子群优化算法[J/OL].(2019-03-17) [2019-06-02] https://doi.org/10.19678/j.issn.1000-3428.0054272.
	Yang Piao, Dong Wenyong.(2019-03-17) [2019-06-02] Chinese NER based on BERT embedding[J]. https://doi.org/10.19678/j.issn.1000-3428.0054272.
[14]	方欣欣, 龚如宾, 李大为. 基于余弦距离的多目标粒子群优化算法[J]. 电子科技, 2016,29(3):48-52,57.
	Fang Xinxin, Gong Rubin, Li Dawei. Multi-objective particle swarm optimization algorithm based on cosine distance[J]. Electronic Science and Technology, 2016,29(3):48-52,57.
[15]	姜猛, 王子牛, 高建瓴. 基于异构数据联合训练的中文分词法[J]. 电子科技, 2019,32(4):29-32,59.
	Jiang Meng, Wang Ziniu, Gao Jianling. Chinese word segmentation based on joint training of heterogeneous data[J]. Electronic Science and Technology, 2019,32(4):29-32,59.
[16]	廖志芳, 周国恩, 李俊锋, 等. 中文短文本语法语义相似度算法[J]. 湖南大学学报(自然科学版), 2016,43(2):135-140.
	Liao Zhifang, Zhou Guoen, Li Junfeng, et al. A Chinese short text similarity algorithm based on semantic and syntax[J]. Journal of Hunan University(Natural Sciences Edition), 2016,43(2):135-140.

	TI+cos	TI+em	TI+mean	TR+cos	TR+em	TR+mean	Imp-TR+cos	Imp-TR+em	Imp-TR+mean
6	0.589	0.378	0.571	0.720	0.347	0.728	0.821	0.491	0.824
7	0.607	0.438	0.619	0.795	0.444	0.810	0.857	0.596	0.863
8	0.630	0.498	0.634	0.848	0.545	0.846	0.871	0.682	0.877
9	0.638	0.552	0.647	0.886	0.671	0.879	0.891	0.775	0.889
10	0.649	0.588	0.652	0.888	0.792	0.894	0.904	0.836	0.903
11	0.661	0.600	0.661	0.885	0.835	0.882	0.892	0.868	0.883
12	0.656	0.619	0.661	0.885	0.871	0.864	0.893	0.903	0.879
13	0.662	0.643	0.661	0.881	0.877	0.853	0.878	0.902	0.845
14	0.660	0.648	0.662	0.878	0.887	0.849	0.865	0.900	0.838
15	0.661	0.659	0.661	0.86	0.881	0.826	0.860	0.897	0.829

	P	R	F₁
TF-IDF+cos	1.000	0.557	0.716
TF-IDF+em	0.660	1.000	0.795
W2v+cos	1.000	0.549	0.709
W2+em	1.000	0.543	0.704
W2v+mean	1.000	0.509	0.675
W2v+cos+TR(10)	0.895	0.882	0.888
W2v+em+TR(14)	0.920	0.856	0.887
W2v+mean+TR(10)	0.930	0.861	0.894
W2v+cos+TI(13)	0.975	0.501	0.662
W2v+em+TI(15)	0.965	0.500	0.659
W2v+mean+TI(14)	0.980	0.500	0.662
W2v+cos+Imp-TR(10)	0.960	0.853	0.904
W2v+em+Imp-TR(12)	0.885	0.922	0.903
W2v+mean+Imp-TR(10)	0.975	0.841	0.903

	P	R	F₁
W2v+cos+TR	0.873	0.867	0.870
W2v+em+TR	0.795	0.774	0.785
W2v+mean+TR	0.634	0.986	0.772
W2v+cos+TI	0.824	0.917	0.868
W2v+em+TI	0.683	0.889	0.772
W2v+mean+TI	0.584	0.989	0.735
W2v+cos+Imp-TR	0.878	0.881	0.880
W2v+em+Imp-TR	0.784	0.853	0.817
W2v+mean+Imp-TR	0.650	0.995	0.787

	P	R	F₁
BERT+cos+TR	0.921	0.917	0.919
BERT+em+TR	0.944	0.785	0.857
BERT+mean+TR	0.967	0.793	0.871
BERT+cos+TI	0.930	0.852	0.889
BERT+em+TI	0.689	0.995	0.814
BERT+mean+TI	0.974	0.824	0.893
BERT+cos+Imp-TR	0.891	0.978	0.932
BERT+em+Imp-TR	0.936	0.867	0.901
BERT+mean+Imp-TR	0.847	0.947	0.894