Prompt-Tuning with Attribute Guidance for Low-Resource Entity Matching

Entity Matching (EM) is a significant task involving the determination of the logical relationship between two entities, such as Same, Different, and Undecidable. Traditional approaches to entity matching (EM) heavily depend on supervised learning, which necessitates a vast collection of high-quality labeled data. This labeling process is both time-consuming and costly, limiting the practical application of these methods. Consequently, there is a pressing demand for low-resource EM solutions that can perform effectively with minimal labeled data. Recently, prompt tuning-based approaches have shown promising results on low-resource entity matching, but they tend to focus solely on entity-level matching, overlooking crucial attribute-level information. Moreover, they lack interpretability and explainability. To address this limitation, this paper introduces PromptAttrib, a comprehensive solution that tackles entity matching challenges through attribute level prompt tuning and logical reasoning. PromptAttrib leverages both entity-level and attribute-level prompts to enhance matching accuracy by incorporating valuable contextual information, and it induces the matching label by fuzzy logic formulas. By considering attributes, the model gains a deeper understanding of the entities, leading to improved matching results. Moreover, PromptAttrib incorporates dropout-based contrastive learning on soft prompts, inspired by the SimCSE technique. This further improves the performance of entity matching. Extensive experiments on real-world datasets demonstrate the efficacy of PromptAttrib.