Impli-R1:结合多模态推理的隐性属性值抽取方法OA

Multimodal attribute value extraction plays a crucial role in applications such as e-commerce,search engines,and automated product categorization,aiming to identify attribute-value pairs from product descriptions.Traditional approaches mainly focus on explicit attributes whose values are directly presented in text or images,however,with the growing diversity of data,extracting implicit attributes has become an increasingly important research direction.Unlike explicit attributes,implicit attributes are not directly observable and require multimodal reasoning and semantic associations for identification,while exis ting methods still struggle with complex cross-modal reasoning and information fusion.To address these challenges,this paper proposed Impli-R1,which introduced a systematic multimodal reasoning mechanism into implicit attribute value extraction by leveraging high-quality long chain-of-thought reasoning samples and adopting a hybrid training strategy that combined super-vised fine-tuning and reinforcement learning,further enhanced by a group-based reward strategy to improve the granularity and robustness of reasoning.Experimental results show that Impli-R1 significantly outperforms existing methods in terms of Micro-F1 on both the ChiImpAVE and ImplicitAVE benchmark datasets,demonstrating its effectiveness and generalization ability in multimodal implicit attribute value extraction.