基于深度学习的消费者投诉信息提取仿真
Simulation of Consumer Complaint Information Extraction Based on Deep Learning
孙徕壹 1董微 2李梦寒 2张青川2
作者信息
- 1. 云粒智慧科技有限公司,北京 100037
- 2. 北京工商大学农产品质量安全追溯技术及应用国家工程研究中心,北京 100048;北京工商大学电商与物流学院,北京 100048
- 折叠
摘要
投诉平台的投诉信息中包含了许多消费者对产品及服务问题的反馈,快速、完整且准确地从消费者投诉信息中识别出问题的类别,并提取相关实体,可以为商家及时提升服务、提高用户满意度提供有效辅助.以黑猫投诉平台中的投诉数据作为数据基础,采用pipeline的方法对实体-多标签进行抽取.构建了Bert-BiLstm模型,提取实体之间的多标签类型,同时建立了Bert-BiLstm-CRF模型对实体对进行抽取.实验结果表明,所提出的实体-多标签抽取模型具有较好的性能,与基准模型对比,F1-score评估指标提升了 0.96%～3.22%.
Abstract
The complaint information of the complaint platform contains a lot of consumers· feedback on product and service problems.It can quickly,completely and accurately identify the types of problems from consumer com-plaint information and extract relevant entities,which can provide effective assistance for businesses to timely improve services and improve user satisfaction.Based on the complaint data of the black Cat complaint platform,this paper u-ses the pipeline method to extract entity-relationship.In this paper,the Bert-BiLSTM model was constructed to ex-tract multi-label types between entities,and the Bert-BiLSTM-CRF model was also established to extract entity pairs.Experimental results show that the entity-multi-label extraction model proposed in this paper has good per-formance,and compared with the benchmark model,the F1-Score evaluation index is improved by 0.96%～3.22%.
关键词
多标签分类/实体抽取/双向长短时记忆网络/条件随机场/投诉Key words
Multi-label classification/Entity extraction/BiLSTM/CRF/Complaints引用本文复制引用
基金项目
北京市自然科学基金面上项目(4202014)
教育部人文社会科学研究青年基金(20YJCZH229)
出版年
2024
NETL
NSTL
万方数据