摘要： 在网络社交平台海量的信息文本中含有许多垃圾文本，这些文本的广泛散布影响了人们正常社交。为此，提出一种垃圾文本过滤模型。通过BERT模型提取文本的句编码，采用B-Feature方法对句编码进行特征构造，并根据文本与所得特征之间的联系进一步将该特征构造为特征矩阵，运用BP神经网络分类器对特征矩阵进行处理，检测出垃圾文本并进行过滤。实验结果表明，该模型在长、中、短文本数据集上的准确率较TFIDF-BP模型分别提高7.8%、3.8%和11.7%，在中、短文本数据集上的准确率较朴素贝叶斯模型分别提高2.1%和13.7%，能有效对垃圾文本进行分类和过滤。

关键词: BERT模型, 特征构造, BP神经网络, 垃圾文本过滤, 文本分类, 句编码

Abstract: There are a lot of junk texts in the massive information of online social platforms,which hinder the normal social intercourse of people when they are widely spread.To address the problem,this paper proposes a junk text filtering model.The model uses the BERT model to extract sentence coding of the text.Then the feature of sentence coding is constructed by using the B-Feature method,and the obtained feature is further constructed as a feature matrix based on the relationship between the feature and the text.The feature matrix is processed by using a BP neural network classifier,and junk texts are detected and filtered.Experimental results show that the accuracy rate of the proposed model on text datasets of long,medium,and short length is respectively 7.8%,3.8% and 11.7% higher than that of the TFIDF-BP model,and the accuracy of the proposed model on text datasets of medium and short length is respectively 2.1% and 13.7% higher than that of the naive Bayes model,which can effectively classify and filter junk texts.

Key words: BERT model, feature construction, BP neural network, junk text filtering, text classification, sentence coding

中图分类号: 

• TP391.1