基于双向LSTM的文本情感倾向分类

Text Sentiment Classification Based on Bilateral LSTM

摘要: 随着互联网普及，在网络上出现了大量带有个人主观性的文本，这些文本含有大量情感相关信息和个人的主观观点，目前通常的卷积网络处理这些文本无法将信息进行关联，处理起来无法达到想要的效果，判断文本情感倾向不够准确，所以本文使用国内百度开发的PaddlePaddle框架，构建双向LSTM (Long Short-Term Memory)网络从众多文本信息和数据中准确而高效地分析出文本中所蕴含的情感，并判断情感极性，对情感倾向做出分类。实验中对美食评论信息进行情感预测，首先利用Embedding来计算出词向量，通过双向LSTM提取特征和融合，借助softmax函数构建分类器，获得文本信息的情感倾向，实验结果较为理想。

Abstract: With the popularity of the Internet, a large number of personally subjective texts appear on the Internet. These text messages contain a large amount of emotional related information and personal subjective opinions. Some of these information are useless and may cause information explosion. At present, the usual convolutional network processing these texts cannot associate the information, and the processing cannot achieve the desired effect, so we use the PaddlePaddle developed by Baidu in China based on the bidirectional LSTM (Long Short-Term Memory) network built by us to obtain a large amount of text information, to analyze the emotion contained in the text accurately and efficiently from the data and judge the polarity of the emotion, classify the emotion, and apply it in practice. The model first uses Embedding to calculate the word vector, then uses the two-way LSTM to extract features and fusion, and finally uses the softmax function to construct a classifier to obtain the emotional tendency of the text information.