基于深度学习方法的中文情感分析

Chinese Sentiment Analysis Based on Deep Learning Methods

针对以文本词向量作为卷积神经网络的输入无法考虑情感特征对文本情感极性的影响、难以突出对类别更具代表性的词且卷积神经网络无法利用文本上下文信息等问题,提出一种基于权重分配的多通道卷积神经网络(WAMCCNN)和双向长短时记忆网络(BILSTM)模型相结合的方法。将文本词向量、情感词向量及词语的特征权重相互结合形成新的特征向量作为卷积网络不同通道的输入,使得模型能够从多方面的特征学习到文本的情感信息且有效利用了每个词语在句子中重要性的信息,获得更多的语义信息。同时,结合BILSTM模型学习到的包含文本上下文信息的全局特征,也解决了卷积神经网络无法利用文本上下文信息的问题。最后在新浪微博评论数据集和京东评论数据集上进行实验,结果表明,该模型分类准确率相比之前的基于深度学习的情感分析模型得到了明显的提升。

Aiming at the problems that the text word vector is used as the input of the convolutional neural network,it cannot consider the influence of emotional features on the sentiment polarity of the text,it is difficult to highlight the words that are more representative of the category,and the convolutional neural network cannot use the text context information.A method based on weight distribution based multi-channel convolutional neural network(WAMCCNN)and bidirectional long-short-term memory network(BILSTM)model is put forward.Combining text word vectors,sentiment word vectors,and feature weights of words with each other to form new feature vectors as inputs to different channels of the convolutional network,allowing the model to learn the emotional information of the text from multiple aspects of the feature and effectively use each word the importance information in the sentence,get more semantic information.At the same time,combining the global features learned by the BILSTM model with text context information,it also solves the problem that the convolutional neural network cannot use the text context information.Finally,experiments are performed on the Sina Weibo review dataset and Jingdong review dataset.The results show that the model classification accuracy has been significantly improved compared to the previous deep learning-based sentiment analysis model.