A topic model-based knowledge graph to detect product defects from social media data

With the help of topic models, social media data offers valuable insights for manufacturers to detect product defects in the after-sales stage. However, topic models struggle with texts mentioning multiple defects or discussing them at a coarse granularity level. Low topic discrimination further limits the application of topic models in defect discovery. To address these problems, we introduce a topic model-based Defect Knowledge Graph (DKG) for accurate defect detection. Firstly, to address the topic-indiscriminative problem, we utilize a topic model named Defect Latent Dirichlet Allocation and an improved Gibbs sampling to extract defect information from multi-source data and construct DKG. Secondly, we establish the Product Component Knowledge Graph (PCKG) to identify multiple defects discussed at coarse granularity levels. Thirdly, with DKG and PCKG, we unveil product defects and related defect information from social media data. Case studies of automobiles and laptops are used for validation. Experimental results show that our method outperforms the state-of-the-art in product defect discovery and provides more comprehensive defect information, which facilitates manufacturers to take prompt remedial actions.