This study examines the relationship between price discovery and commodity-related media sentiment in the Chinese futures market through a tone-based natural language processing framework. We integrate a customized Chinese BERT model into a BiLSTM network with an Attention mechanism, enhancing semantic representation through contextual embeddings and salient feature emphasis. Analyzing over 850,000 news articles, we find sentiment provides significant incremental predictive power for commodity portfolio returns. We finally design a long-short trading strategy that consistently outperforms several benchmark approaches. The results emphasize that deep semantic sentiment analysis offers substantive predictive value for asset returns.