面向在线健康社区的生成式方面级情感分析

Generative Aspect Based Sentiment Analysis for Online Health Communities

[目的/意义]为解决在线健康社区文本中方面实体和评论实体难以对齐的问题,提出了一种基于端到端的生成式方面级情感分析模型BERT-WWM-GPT。[方法/过程]首先,在模型训练阶段通过编码器抽取文本中包含丰富语义信息的特征向量;其次,基于特征向量和标准预测序列在解码器中迭代生成情感三元组,并通过最大似然估计训练模型参数;然后,在模型推理阶段基于文本语义特征向量在解码器生成预测序列;最后,利用规则得到有效的情感三元组表达。[结果/结论]对自建数据集和5份公共数据集进行验证,结果表明BERT-WWM-GPT模型在两个方面级情感分析任务中的F1值分别比基准模型GTS和MuG RoBERTa-large提升了12.25%和7.22%。BERT-WWM-GPT模型能够有效抽取在线健康社区评论中的多重情感三元组,且在其他领域具有优秀的泛化能力。

[Purpose/Significance]The study aims to solve the problem of difficult alignment between aspect entities and comment entities in online health community texts,an end-to-end generative aspect based on sentiment analysis model BERT-WWM-GPT is proposed.[Method/Process]Firstly,in the model training phase,feature vectors containing rich semantic information were extracted from the text using an encoder.Next,based on feature vectors and standard prediction sequences,sentiment triplets were iteratively generated in the decoder,and model parameters were trained through maximum likelihood estimation.Then,in the model inference stage,a prediction sequence was generated in the decoder based on the text semantic feature vector.Finally,used rules to obtain effective sentiment triplet expression.[Result/Conclusion]The results show that the F1 value of BERT-WWM-GPT model is 12.25%and 7.22%higher than that of GTS and MuG RoBERTa-large models in two aspect level sentiment analysis tasks,respectively.BERT-WWM-GPT model can effectively extract multiple affective triples from online health community reviews,and has excellent generalization ability in other fields.