基于引导式主题模型的产品缺陷信息识别方法及应用  

Guided topic model based product defect information detection method and applications

作　　者：施文 毛星鹭 SHI Wen;MAO Xinglu(Business School,Central South University,Changsha 410083,China;Xiangjiang Laboratory,Changsha 410205,China)

机构地区：[1]中南大学商学院,长沙410083 [2]湘江实验室,长沙410205

出　　处：《系统工程理论与实践》2025年第1期290-309,共20页Systems Engineering-Theory & Practice

基　　金：湘江实验室项目(22XJ03025);国家自然科学基金(71971219,72293574);湖南省杰出青年基金(2022JJ10084)。

摘　　要：如何从日益增多的线上投诉中挖掘出潜在的产品缺陷信息已成为制造商和监管部门质量管理工作的重点和难点.为此,本研究提出了一个全新的基于主题模型的产品缺陷信息识别模型.该模型首先利用引导式主题模型的优势,通过缺陷词典事先确定缺陷主题数量,提取出与行业标准内容一致的缺陷主题,并根据这些主题对投诉进行多标签分类;另外,针对投诉附带信息的非结构化特点以及投诉内容缺陷词的稀疏性分布规律,模型融入了产品特征建模以及自适应性词加权两种改进机制,进一步提升了缺陷识别的性能.本研究收集了国内汽车权威平台115,668条线上投诉进行实验,实验结果表明所提模型相比于潜在狄利克雷分配等现有主题模型能生成与行业的实际缺陷分类一致的缺陷主题,还能分析产品的品牌,能源类型等特征对投诉中缺陷主题概率分配的影响.模型在F1指数, AUC指数,汉明损失,排序损失上明显优于对比的其他主题模型.模型的产品特征建模,自适应性词加权和引导式主题模型结构确保了本模型在各评价指标上的精度表现.该研究可帮助制造商和监管部门及时监管产品质量水平,保护消费者合法权益.Extracting potential product defect information from the increasingly abundant online complaints has become a focal point and challenge in quality management for both manufacturers and regulatory agencies.To this end,this study proposes a novel product defect information detection model based on topic modeling.Firstly,the model takes advantage of the guided topic model to determine the number of defect topics through a defect dictionary,extract defect topics consistent with industry standards,and perform multi-label classification of complaints based on these topics.In addition,the model incorporates two improvement mechanisms,namely product feature modeling and adaptive term weighting,to further enhance defect recognition performance by addressing the unstructured nature of complaint information and the sparse distribution of defect words.This study collected 115,668 online complaints from a leading domestic automotive platform for experimentation.Experimental results indicate that compared to existing topic models such as latent Dirichlet allocation,the proposed model can generate defect topics consistent with actual industry defect classifications.Additionally,it can analyze the influence of product brand,energy type,and other features on the probability distribution of defect topics in complaints.The model outperforms other comparative topic models significantly in F1 score,AUC score,Hamming loss,and ranking loss.The product feature modeling,adaptive term weighting,and guided topic model structure of the model ensure its accuracy in all evaluation metrics.This research can help manufacturers and regulatory agencies monitor product quality levels in a timely manner and protect the legitimate rights and interests of consumers.

关 键 词：缺陷信息识别 引导式主题模型 自适应性词加权 产品特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] C931[自动化与计算机技术—计算机科学与技术]