农产品评价观点抽取和情感识别系统设计实现

来源：计算机技术与发展杂志2023年第08期北京时间：

作者：陈杰;周梓豪*;吴军辉*

单位：同济大学 电子与信息工程学院,上海 201804 Author(s): CHEN Jie;ZHOU Zi-hao* ;WU Jun-hui* School of Electronic & Information Engineering,Tongji University,Shanghai 201804,China 关键词: 观点挖掘;自然语言处理;无监督学习;领域词典;依存句法规则;农产品评价 Keywords: opinion mining;natural language processing;unsupervised learning;domain dictionary;dependency parsing rule;agriculturalproduct review 分类号: TP391. 1 DOI: 10. 3969 / j. issn. 1673-629X. 2023. 08. 017 摘要: 电商平台上的评价数据蕴藏着消费者的情感观点,识别评价情感表达的关键是挖掘其在产品属性方面级别的观点,并判别情感倾向。 先前的有监督学习模型需要相关领域的大量人工标注数据进行训练,耗费较多的人力成本,因此,构建了无监督学习框架的农产品评价观点抽取和情感识别系统。 通过爬虫获取多源电商平台的评价数据,首先通过 LDA模型确定领域主题属性,结合 SO-PMI 算法构建领域情感词典,然后通过 LTP 库的依存句法分析和词嵌入相似度制定方面观点的抽取规则,并提出情感强度值计算方法识别评价的方面情感倾向。 实验证明,该框架的查准率为 85. 08% ,召回率为 78. 50% ,F1 值为 81. 66% ,性能优于传统模型。 根据观点抽取和情感识别结果构建可视化平台,从多个角度挖掘消费者对农产品的偏好。 该系统已实际部署在农资农产品在线服务交易平台的项目中,致力于服务消费者、经销商、电商平台和监管部门四个主体,取得了良好的应用效果。 Abstract: The review data on the e-commerce platform contains sentiment opinions of consumers,and the key to identifying sentimentexpression of reviews is?to extract their opinions at the product attribute level and identify sentiment tendencies. Previous supervisedlearning models require a large amount of manually labeled data in related fields for training, which consumes a lot of labor costs.Therefore, an unsupervised learning framework is constructed for agricultural product review for opinion mining and sentimentrecognition. The review data of the multi - source e - commerce platform is obtained through?the crawler. Firstly, the domain themeattribute is determined by the LDA model, and the domain sentiment dictionary is constructed in combination with?the SO - PMIalgorithm,then the extraction rules of aspect-level opinions are formulated through the dependency syntactic analysis of the LTP libraryand the similarity of word embedding, and the sentiment intensity value calculation method is proposed to identify the aspect - levelsentiment tendency of the review. Experiments show that the accuracy of the proposed framework is 85. 08% ,the recall rate is 78. 50% ,and the F1 value is 81. 66% ,which is better than that of traditional models. According to the opinion mining and sentiment recognitionresults,a visualization platform is built to explore consumers爷 preferences for agricultural products from multiple angles. The system hasbeen actually deployed in the project of the online service trading platform for agricultural materials and agricultural products, and iscommitted to serving consumers,distributors,e-commerce platforms and regulatory departments,which achieved good application results.

摘要：电商平台上的评价数据蕴藏着消费者的情感观点,识别评价情感表达的关键是挖掘其在产品属性方面级别的观点,并判别情感倾向。 先前的有监督学习模型需要相关领域的大量人工标注数据进行训练,耗费较多的人力成本,因此,构建了无监督学习框架的农产品评价观点抽取和情感识别系统。 通过爬虫获取多源电商平台的评价数据,首先通过 LDA模型确定领域主题属性,结合 SO-PMI 算法构建领域情感词典,然后通过 LTP 库的依存句法分析和词嵌入相似度制定方面观点的抽取规则,并提出情感强度值计算方法识别评价的方面情感倾向。 实验证明,该框架的查准率为 85. 08% ,召回率为 78. 50% ,F1 值为 81. 66% ,性能优于传统模型。 根据观点抽取和情感识别结果构建可视化平台,从多个角度挖掘消费者对农产品的偏好。 该系统已实际部署在农资农产品在线服务交易平台的项目中,致力于服务消费者、经销商、电商平台和监管部门四个主体,取得了良好的应用效果。

关键词：观点挖掘;自然语言处理;无监督学习;领域词典;依存句法规则;农产品评价