卷积神经网络在短文本情感多分类标注应用

情感多分类标注对文本信息的敏感性远高于二分类问题。为了有效利用语义依赖距离和语义多层次进行情感多分类，提出一种多窗口多池化层的卷积神经网络模型。首先使用多窗口的卷积层提取上下文局部语义，然后通过多池化层降低特征维度，同时保留不同层次的语义，由多层次语义构成文本特征向量，最后送入全连接层完成多分类标注。采用斯坦福情感树库数据集验证所提模型的多分类标注效果。实验结果表明，在训练集含短语和未包含短语两种设定下，模型的短文本情感多分类正确率分别达到54.6%和43.5%。

Application of convolutional neural network in multi-category classification for short text sentiment

Deep learning based approaches achieved less for sentiment classification with multiple labels. For this issue, this paper proposes a model called mwmpCNN（multi-windows and multi-pooling Convolutional Neural Network） to grasp the semantic distance and various emotional levels. mwmpCNN assemblies convolution layer with multiple windows to extract local context semantic, and then applies multi-pooling layer to keep multi-level semantic in short text when the feature dimension is reduced. Here, the text feature vector is constructed and reflected by the multi-level semantic, and connection layer is implemented for multi-label classification. This paper evaluates mwmpCNN by the test on Stanford Sentiment Treebank. mwmpCNN exhibits the classification accuracy of 54.6% and 43.5% respectively for the multi-label classification task.