结合语义扩展和卷积神经网络的中文短文本分类方法

发布时间：2018-12-25 17:06

【摘要】：中文新闻标题通常包含一个或几十个词,由于字符数少、特征稀疏,在分类问题中难以提升正确率。为解决此问题,提出了基于Word Embedding的文本语义扩展方法。首先,将新闻标题扩展为(标题、副标题、主题词)构成的三元组,用标题的同义词结合词性过滤方法构造副标题,对多尺度滑动窗口内的词进行语义组合,提取主题词;然后,针对扩展文本构造卷积神经网络(CNN)分类模型,该模型通过max pooling及随机dropout进行特征过滤及防止过拟合;最后,将标题、副标题拼接为双词表示,与多主题词集分别作为模型的输入。在2017自然语言处理与中文计算评测(NLPCC2017)的新闻标题分类数据集上进行实验。实验结果表明,用三元组扩展结合相应的CNN模型在18个类别新闻标题上分类的正确率为79.42%,比未经扩展的CNN模型提高了9.5%,且主题词扩展加快了模型的收敛速度,验证了三元组扩展方法及所构建CNN分类模型的有效性。
[Abstract]:Chinese news headlines usually contain one or dozens of words. Because of the small number of characters and sparse features, it is difficult to improve the accuracy in classification problems. To solve this problem, a text semantic extension method based on Word Embedding is proposed. Firstly, the news title is extended to a triple (title, subtitle, subject word). The subtitle is constructed by the synonym of the title combined with the method of part of speech filtering, and the words in the multi-scale sliding window are semantically combined to extract the subject word. Then, the (CNN) classification model of convolution neural network is constructed for extended text. The model uses max pooling and random dropout to filter features and prevent over-fitting. Finally, the title and subtitle are spliced into two-word representation, and the multi-subject word set is used as the input of the model. Experiments were carried out on the 2017 Natural language processing and Chinese Computational Evaluation (NLPCC2017) news headline classification data set. The experimental results show that the correct rate of classifying with triple extension and corresponding CNN model on 18 categories of news headlines is 79.42, which is 9.5% higher than that of unexpanded CNN model, and the speed of convergence of the model is accelerated by the expansion of subject words. The validity of the triple extension method and the constructed CNN classification model is verified.
【作者单位】：重庆理工大学计算机科学与工程学院;
【基金】：国家社会科学基金西部项目(17XXW005) 重庆市教委科学技术研究项目(KJ1500903)~~
【分类号】：TP183;TP391.1

【相似文献】