基于RoBERTa-BiLSTM-CRF融合模型的在线评论细粒度情感分析

Fine-grained sentiment analysis of online reviews based on RoBERTa-BiLSTM-CRF

在电子商务迅速发展的背景下,在线评论所蕴含的商业价值日益凸显.从在线评论中提取用户关于产品的评价和情感的研究,已经开始从句子级或篇章级的粗粒度情感分析转向属性级的细粒度情感分析.但当前细粒度情感分析方法在情感要素识别任务中存在不能同时解决一词多义、上下文语义信息不全以及标签约束关系缺失等突出问题,且面向属性的情感强度量化方法未充分考虑语法信息.对此,本文提出一种基于RoBERTa-BiLSTM-CRF融合模型的在线评论细粒度情感分析方法,该方法可以有效解决上述问题,更加准确地识别评论中用户评价的产品或服务属性,并结合情感三元组和语法信息有效地量化用户在评论中反馈的情感强度.为了检验所提方法的效果,本文在酒店评论数据、美团外卖评论数据、CLUENER2020等多个领域的数据集上进行对比实验与消融实验.实验结果表明,与已有经典模型相比,本文所提基于RoBERTa-BiLSTM-CRF融合模型的情感要素识别方法在多个数据集上均获得了最佳F1值,且本文所提情感强度量化方法更加精细,能更好地契合人类情感的连续性.此外,消融实验进一步表明融合模型的每个结构都具有重要性.

In the context of the rapid development of e-commerce,the commercial value of online reviews has become increasingly prominent.The methods of the sentiment analysis of users generated online reviews has shifted from coarse-grained sentiment analysis at sentence or paragraph level to fine-grained sentiment analysis at attribution level.However,the current finegrained sentiment analysis methods have limitations in the sentiment factors identification task,such as polysemy,underutilization of context semantics,and ignorance of the constraint condition between labels.Meanwhile,the quantification methods of measuring the attribute-oriented sentiment intensity does not fully consider grammatical information.In this regard,this paper proposes a hybrid fine-grained sentiment analysis method,RoBERTa-BiLSTM-CRF,which can effectively solve the above problems.This method can accurately quantify the attributes-oriented sentiment intensity of products or services evaluated by users in online reviews,and effectively quantify the sentiment strength of user feedback by combining sentiment triplet and grammatical information.In order to evaluate the effect of the proposed method,groups of comparative experiments and ablation experiments were conducted on the datasets from multiple fields,including hotel online reviews,Meituan takeaway online reviews and the CLUENER2020 dataset.The experimental results show that,compared with the state-of-the-art models,the RoBERTaBiLSTM-CRF model used in this paper has obtained the best F1 score in the experiments of sentiment element identification on all datasets,and the quantification methods of measuring the attribute-oriented sentiment intensity proposed in this paper is more accuracy,which can better reflect the continuity of human emotion.In addition,the ablation experiments further demonstrate the importance of each structure of the fusion model.