Live-streaming e-commerce has emerged as a dominant retail channel, offering dynamic, real-time interactions between consumers and sellers. In this context, product attribute value identification (PAVI)—the task of extracting attribute-value pairs (e.g., color, flavor) from unstructured product descriptions plays a crucial role in enhancing search and customized recommendation. However, the spontaneous, informal, and often noisy nature of live-streamed speech poses significant challenges for existing PAVI approaches, which either rely heavily on manual annotations or struggle with scalability in open-domain settings. To address these issues, we propose a weakly-supervised, multi-scale framework for attribute mining in live-streaming e-commerce. Our method leverages a small set of LLM-generated annotations, refined through human verification, to train a quality evaluator and a text generator with multi-scale span corruption. High-confidence outputs from the generator, filtered by the evaluator, are iteratively added to improve the model. In addition, we construct a new dataset comprising over 1,000 h of live-stream content, with manually annotated seed data covering 54 core product attributes. Experimental results demonstrate that our approach outperforms existing baselines. Specifically, in our live-streaming e-commerce context, compared to existing methods, our model achieves improvements of 6.9 in recall and 5.0 in F1 score. The code is available.