融合多种类型语法信息的属性级情感分析模型

Aspect-Based Sentiment Analysis Model with Multiple Grammatical Information

属性级情感分析(ABSA)的目标是识别出句子中属性的情感倾向。现有的方法大多使用注意力机制隐性地建模属性与上下文中情感表达的关系,而忽略了使用语法信息。一方面,属性的情感倾向与句子中的情感表达有紧密的联系,利用句子的句法结构可以更直接地对两者建模;另一方面,由于现有的基准数据集较小,模型无法充分学习通用语法知识,这使得它们难以处理复杂的句型和情感表达。针对以上问题,提出一种利用多种类型语法信息的神经网络模型。该模型采用基于依存句法树的图卷积神经网络(GCN),并利用句法结构信息直接匹配属性与其对应情感表达,缓解冗余信息对分类的干扰。同时,使用预训练模型BERT具有多种类型的语法信息的中间层表示作为指导信息,给予模型更多的语法知识。每一层GCN的输入结合上一层GCN的输出和BERT中间层指导信息。最后将属性在最后一层GCN的表示作为特征进行情感倾向分类。通过在SemEval 2014 Task4 Restaurant、Laptop和Twitter数据集上的实验结果表明,提出模型的分类效果超越了很多基准模型。

The aim of aspect based sentiment analysis(ABSA)is to classify the sentiment polarity towards a particular aspect in a sentence.Existing approaches usually apply attention mechanism to modeling the connection between aspects and opinion expression in an implicit way.However,they ignore the potentially useful grammatical information.On one hand,the sentiment polarity of an aspect is closely related to opinion expression,and syntactic information helps to better model the relation of them.On the other hand,models are hard to learn general grammatical knowledge when trained on existing small benchmark datasets,resulting in difficulty to handle complex sentence patterns and opinion expressions.To address the problem,this paper proposes a neural network that combines various kinds of grammatical information to enhance the accuracy.This paper employs dependency tree based graph convolutional networks(GCN)to match aspects and their corresponding opinion expression directly using syntactic information and eliminate useless information.This paper also uses middle layers of BERT as guiding information to enhance the model,which contains various kinds of contextual and grammatical information.The input of each GCN layer fuses the output of the preceding GCN layer and BERT(bidirectional encoder representations from transformers)middle layers.Finally,the aspect representation of the last layer GCN is used as the feature to identify sentiment polarity.Experiments on SemEval 2014 Task4 Restaurant,Laptop and Twitter show the model outperforms other baselines.