结合共现网络的方面级情感分析研究

Research on Aspect-Level Sentiment Analysis Combined with Co-Existing Networks

方面级情感分类研究中常用的方法是利用句法结构和神经网络的结合模型,但原始句法结构中都未曾考虑共现关系,这会使依赖关系的构建不完整,导致最终效果不佳。提出融合共现网络的图神经网络,重构句法依赖关系和句法依赖标签,根据词汇与方面词之间的距离剪枝,利用关系图注意力网络融合句法依赖关系和文本,通过多头注意力层,将依赖关系与依赖标签相互作用,得到情感特征,在SemEval2014的Restaurant和Laptop数据集以及Twitter数据集上进行实验,该模型的准确率分别达到78.66%、74.19%和73.97%,均高于其余基准模型。

The common method used in aspect-level emotion classification research is to use a combination model of syntactic structure and neural network,but the co-existing relation is not considered in the original syntactic structure,which makes the construction of dependencies incomplete and leads to poor final results.In this paper,it proposes a graph neural network that fuses the co-occurrence network,first reconstructs syntactic dependencies and syntactic dependency labels,prunes according to the distance between words and aspect words,uses the graph attention network to fuse syntactic dependencies and texts,and then interacts with dependencies with dependency tags through the multi-head attention layer,and finally obtains emotional features,and experiments are conducted on the Restaurant and Laptop in SemEval2014 datasets and Twitter datasets.The accuracy of this model is 78.66%,74.19% and 73.97%,respectively,which are higher than the remaining benchmark models.