基于音位的网络盗版文本查重方法

Method for Checking Duplicate Text of Network Piracy Based on Phoneme

传统的文本查重算法是对文本作分词以构建关键词向量,而对于某些特殊应用的网络盗版检测,分词的开销则未必合理和必要。因此,本文提出一种基于汉语音位信息的文本查重方法。文本被表达为声、韵、调三个空间向量,以余弦距离作相似性度量。提出两种相似性判断公式,一种假定三向量独立分布;一种取其线性组合,系数可由音位元素的信息熵算出,通过大文本统计得出信息熵的估计值,以传统的关键词向量/Sim Hash方法做参照产生语料,对其作统计得到模型参数。实验结果表明该方法有一定的精确率和很好的召回率,计算开销低于传统的方法,适合需要过滤大量TN类型文本的场合。

The traditional method checking repetition takes a text as a participle to establish some key vectors,however the piratical cost may not be reasonable or necessary for the discovery of the online copyright violation in some special APP.Therefore this paper proposed a method checking repetition with Chinese phonology.A text was represented by three vectors in spaces of Chinese initial,final and tone and cosine distance was used as a measurement of similarity.Two decision models were proposed.One assumed the three vectors were independent each other,while the other took a linear combination of the three,which needed to calculate the factors using information entropies that could be evaluated by large-corpus counting.Training corpus was generated with the old term-vector/SimHash method being used as a standard and threshold values were calculated.Test results showed the proposed method had a good precision and a very good recall ratio,and computational cost was lowed comparing to traditional methods based on term vectors to be suitable for filtering out a large amount of TN documents.