多主题下基于 LSTM 语义关联的长文本过滤研究-《计算机技术与发展》

文章信息/Info

Title:: Research on LSTM Semantic Correlation Long Text Filtering Based on Subject Dependence

Author(s):: CAO Chun-ping; WU Ting; School of Optical-electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200082,China

Keywords:: long text filtering; multi-topic; semantic association; LSTM; hierarchical model

摘要:: 现如今互联网上出现了很多评论性文章,这些文章字符数多,且包含较多与主题无关的信息,会影响后续的文本分析任务的性能。因此,针对传统的解决方案不能够对多主题长文本进行建模,以及现有的神经网络无法从相对较长的时间步长中捕获语义关联等问题,文中提出了一种结合单层神经网络和分层长短记忆网络的深度网络模型,并在长文本过滤任务中进行应用。该模型通过词语层 LSTM 网络获得句子内部词语之间的关系并得到具有语义的句向量,然后将句向量输入主题依赖度计算模型和句子层 LSTM 网络模型,进而得到句子与各主题类别的依赖度以及待过滤句子与其他句子之间的关联。最后,在从马蜂窝获取的游记数据集上进行的实验表明,该模型相比 SVM、朴素贝叶斯、LSTM、Bi-LSTM等效果更好。

Abstract:: Nowadays,there are a lot of critical articles on the Internet. These articles have more characters and contain more information irrelevant to the topic,which will affect the performance of subsequent text analysis tasks. The traditional solution cannot model multi-theme long text,and the existing neural network cannot capture the semantic association from a relatively long time step. Therefore,we propose a deep network model combining single-layer neural networks and layered long and short memory networks. The model obtains the relationship between the internal words of the sentence through the word layer LSTM network and obtains the semantic vector. Then the sentence vector is input into the subject dependence calculation model and the sentence layer LSTM network model,in turn,the dependence of the sentence and each topic category and the relationship between the sentence to be filtered and other sentences are obtained.Finally,experiments on the travel data set acquired from the Mafengwo shows that this model is superior to SVM,Naive Bayes,LSTM,Bi-LSTM.