融合Bi-LSTM和条件随机场的在线学习情感分析方法  

Sentiment analysis method for online learning reviews integrating Bi-LSTM and conditional random field

作　　者：周燕[1] ZHOU Yan(College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]华南农业大学数学与信息学院,广东广州510642

出　　处：《厦门大学学报（自然科学版）》2023年第4期687-694,共8页Journal of Xiamen University：Natural Science

基　　金：广东省教学改革项目(JG22026,JG22029);教育部产学合作协同育人项目(22097133283536);广东省课程思政示范课程(202022019)。

摘　　要：为改善文本评论的细粒度属性识别和情感分析的准确度,提出基于双向长短期记忆(Bi-LSTM)和条件随机场(CRF)的属性级情感分析框架.将评论句的属性项提取和情感极性分析建模为序列标注问题,提出新的标注方案,在完成属性项提取的同时确定情感极性.结合词性(POS)嵌入和词嵌入作为神经网络输入,并融合Bi-LSTM和CRF网络,利用Bi-LSTM高效捕捉两个方向的词语关联,并将结果输入CRF网络以得到特征函数与输出标签之间的条件分布,实现高质量特征提取和准确标签分配.实验结果表明,结合所提新标注方案后,Bi-LSTM和CRF网络具有互补性,融合网络性能显著优于单一网络.此外,所提方案在公开数据集上取得了与当前先进方法大致相当的性能,且在外部知识库不可用的在线学习评论数据集上,所提方法的情感分析准确度优于当前其他先进的深度学习方法和学习评论分析方法,具有较好的应用价值.To improve the accuracy of attribute item recognition and sentiment analysis in large-scale online text reviews,here we propose an aspect-level sentiment analysis framework based on bidirectional long short-term memory(Bi-LSTM)and conditional random fields(CRF).The attribute term extraction and sentiment polarity analysis of review sentences are modeled as sequence labeling problems,and a new tagging scheme is developed to determine the sentiment polarity while the attribute term extraction is being completed.The part-of-speech(POS)embedding and word embedding are combined as the input of the neural network,the Bi-LSTM and CRF network are fused,and the Bi-LSTM is used to efficiently capture the word associations in both directions.These results become the input into the CRF network to obtain the conditional distributions between the feature functions and the output labels,so that high-quality feature extractions and accurate label assignments are achieved.Experimental results show that,based on the proposed new tagging scheme,the Bi-LSTM and CRF networks behave complementarily,and the fusion network performs significantly better than the single network does.In addition,the proposed method achieves roughly comparable performance to state-of-the-art methods on public datasets,and the sentiment analysis accuracy of the proposed method outperforms other advanced generic methods and student feedback analysis methods on online learning review dataset in which external knowledge bases are unavailable.All these results demonstrate that the proposed method secures certain application values.

关 键 词：双向长短期记忆网络 条件随机场 情感分析 特征函数 词嵌入 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]