面向上下文注意力联合学习网络的方面级情感分类模型

针对现有的方面级情感分类模型存在感知方面词能力较弱、泛化能力较差等问题,文中提出面向上下文注意力联合学习网络的方面级情感分类模型(CAJLN).首先,利用双向Transformer的表征编码器(BERT)模型作为编码器,将文本句子预处理成句子、句子对和方面词级输入序列,分别经过BERT单句和句子对分类模型,进行上下文、方面词级和句子对隐藏特征提取.再基于上下文和方面词级隐藏特征,建立上下文和方面词的多种注意力机制,获取方面特定的上下文感知表示.然后,对句子对隐藏特征和方面特定的上下文感知表示进行联合学习.采用Xavier正态分布对权重进行初始化,确保反向传播时参数持续更新,使CAJLN在训练过程中可以学习有用信息.在多个数据集上的仿真实验表明,CAJLN可有效提升短文本情感分类性能.

Context-Oriented Attention Joint Learning Network for Aspect-Level Sentiment Classification

To solve the problems of weak perception for aspect words and generalization ability in the existing models for sentiment classification, a context-oriented attention joint learning network for aspect-level sentiment classification(CAJLN) is proposed. Firstly, the bidirectional encoder representation from transformers(BERT) model is employed as the encoder to preprocess short texts into sentences, sentence pairs and aspect words as input, and their hidden features are extracted through the single sentence and sentence pair classification models, respectively. Then, based on the hidden features of sentences and aspect words, attention mechanisms for sentences and aspect words are established to obtain aspect-specific context-aware representation. Then, the hidden features of sentence pairs and aspect-specific context-aware representations are learned jointly. Xavier normal distribution is utilized to initialize the weights. Thus, the continuous updating of the parameters during the back propagation is ensured, and useful information is learned by CAJLN in the training process. Experiments show that CAJLN effectively improves the performance of sentiment classification for short text on multiple datasets.