基于情感融合和层次注意力机制的情感分析

Sentiment analysis based on sentiment fusion and hierarchical attention mechanism

针对两个反义词在相似语境下转化成词向量后空间距离相近,容易造成情感信息丢失,循环神经网络等的特征提取方式容易导致网络依赖增强,难以充分提取局部性特征。针对第一个问题,本文提出情感嵌入模块,在词嵌入的过程中加入情感向量与语义信息作为网络的输入层;针对第二个问题,本文提出层次注意力机制,将融合后的词向量切片形成两个子序列,将单词的词向量输入到双向门控循环网络,利用注意力机制对隐藏层进行加权计算,获得子序列文本信息,通过多个网络层获得整个文本序列信息;最后,经过softmax函数输出文本情感极性。在NLPIR微博语料库和NLPCC2014的微博公开数据集进行实验,表明该情感分析模型在准确率上有所提高,证明了模型的有效性。

For two antonyms converted into word vectors in similar contexts,the spatial distances are similar,which is easy to cause loss of emotional information,and feature extraction methods such as recurrent neural networks can easily lead to network dependence enhancement and it is difficult to fully extract local features.For the first question,this paper proposes a sentiment embedding module,which adds emotion vector and semantic information as the input layer of the network in the process of word embedding.Aiming at the second problem,this paper proposes a hierarchical attention mechanism to slice the fused word vector into two subsequences.The word vector of the word is input into the bi-directional gating recurrent network,the attention mechanism is used to perform weighting calculation on the hidden layer to obtain the subsequence text information,and the entire text sequence information is obtained through multiple network layers.Finally,the sentiment polarity of the text is output through the softmax function.Through experiments on the NLPIR Weibo corpus and the NLPCC2014 Weibo public data set,the accuracy of the sentiment analysis model has been improved,which proves the effectiveness of the model.