基于事件卷积特征的新闻文本分类

以往的卷积神经网络模型在对文本建模和分类时，通常按顺序提取n-gram卷积特征，忽视了长距离依存关系中的句法结构和语义信息。提出了一种基于事件卷积特征的文本分类方法，利用事件的语义特性弥补之前模型的不足。该方法使用依存关系抽取出文本中的事件集合，通过卷积神经网络进行事件特征提取，并在此基础上进行文本分类。在对中文新闻语料的多分类实验中，该方法较传统的文本分类方法有明显的提高，较使用n-gram的卷积神经网络模型在更为稳定。实验结果说明了模型的有效性以及事件特征的优越性。

Event convolutional feature based news documents classification

In text modeling and classification, previous convolutional neural network (CNN) approaches process on the n-gram features based on the literal order of texts. They neglect the syntactic structure and semantic information over long distance dependencies. This paper proposes a event convolutional feature based model to overcomes the defects by making use of semantic characteristics of events. It finds events from text and applies a CNN to extract features for classification. In Chinese news multi-class classification experiment, the method performs better than traditional ones and is more balanced than n-gram CNN models. The experiment result shows the effectiveness of the model as well as the superiority of the event features.