Abstract
Product defect detection (PDD) is a critical industrial big data application that leverages social media text mining to identify product defects. However, existing PDD approaches face significant challenges: heavy reliance on expert-driven aspect term extraction (ATE) limits scalability and increases costs, and they fail to capture detailed defect information like affected aspects, defect causes, and severity. To address these limitations, this study proposes a novel multi-task learning framework that integrates ATE and aspect-level defect detection. By enhancing ATE with semantic augmentation, this model reduces expert dependence and scales effectively. Experiments on automotive datasets show the proposed method improves the F1 score by 5% (to 0.88) over benchmark models. Furthermore, ablation studies reveal a 36.5% F1 score improvement for the ATE component. The proposed method outperforms benchmark PDD models and provides more fine-grained defect information, enabling a more comprehensive understanding and supporting precise decision-making in quality management.
Funding Information
  • Xi’an Jiaotong-Liverpool University (RDF-20-01-23, TDF-21/22-R24-176)

This publication has 46 references indexed in Scilit: