基于胶囊网络的跨语言方面级情感分析

Cross-language Aspect Level Sentiment Analysis Based on Capsule Network

跨语言情感分析的目的是利用数据资源丰富的源语言帮助资源较少的目标语言进行情感分析。针对中文文本标注语料较少和不同方面项的不同情感极性特征重叠影响文本情感分析准确率的问题,提出一种基于胶囊网络的跨语言方面级情感分类方法 BBCapNet,该方法利用BERT模型学习源语言的语义特征训练词向量作为嵌入层,然后利用BiLSTM学习上下文信息,利用胶囊网络(Capsule Network)获取文本中局部信息和整体情感极性间的关系,从而提取不同方面项的情感特征,最后使用归一化指数函数(Softmax)进行分类。通过与其他主流方法进行对比论证,论证结果表明,该方法在跨语言方面级情感分类效果上有显著提升。

The purpose of cross-language sentiment analysis is to use source languages with abundant data resources to assist target languages with limited resources in sentiment analysis.A cross-language aspect level sentiment classification method BBCapNet based on Capsule Networks is proposed to address the issue of limited corpus for Chinese text annotation and overlapping sentiment polarity features of different aspects,which affects the accuracy of text sentiment analysis.This method uses the BERT model to learn semantic features of the source language and train word vectors as embedding layers,and then uses BiLSTM to learn contextual information,The Capsule Network is used to obtain the relationship between the local information and the overall sentiment polarity in the text,so as to extract the sentiment characteristics of different aspects.Finally,the normalized exponential function(softmax)is used for classification.By comparing with other mainstream methods,the results show that this method has a significant improvement in cross-language aspect level sentiment classification.