Missing Modality Product Completion Benchmark (MMPCBench)
收藏资源简介:
MMPCBench是由格拉斯哥大学联合亚马逊等机构构建的电商多模态补全基准测试,包含内容质量补全和推荐系统两个子任务。数据集基于2024版亚马逊评论数据,涵盖美妆、家居、电子产品等9大品类共9000条商品记录,每条包含图文双模态信息。通过人工构造缺失模态场景,评估MLLM模型在图文互生成任务中的表现,旨在解决电商平台因模态缺失导致的下游推荐性能下降问题。数据构建过程采用五核过滤法确保样本质量,并引入CLIP相似度等指标进行多维度评估。
MMPCBench is an e-commerce multimodal completion benchmark constructed by the University of Glasgow in collaboration with Amazon and other institutions, covering two subtasks: content quality completion and recommender system. Based on the 2024 version of Amazon review data, the dataset includes 9000 product records spanning 9 major categories such as beauty, home goods, and consumer electronics. Each record contains bimodal information of both text and image. By artificially constructing missing modality scenarios, MMPCBench evaluates the performance of Multimodal Large Language Models (MLLMs) in cross-modal text-image generation tasks, aiming to solve the problem of degraded downstream recommendation performance caused by missing modalities on e-commerce platforms. The data construction process employs five-core filtering to ensure sample quality, and introduces metrics including CLIP similarity for multi-dimensional evaluation.




