核心概念
本文探討如何利用多模態大型語言模型 (MLM),根據廣告內容,主動偵測和校準季節性廣告,以提升廣告投放效率和使用者體驗。
統計資料
使用內部基準測試,多模態大型語言模型在季節性廣告偵測任務中達到了 0.97 的 F1 分數。
關鍵字過濾方法的精確率估計約為 98%。
關鍵字過濾方法的覆蓋率估計約為 10%。
次級關鍵字方法的召回率估計約為 30%。
次級關鍵字方法的精確率約為 10%。
隨機廣告的精確率約為 2%。
人工標記實驗中,精確率為 60%,召回率為 89%,F1 分數為 72%。
引述
"Failing to differentiate and treat seasonal ads leads to missing opportunities on capitalization, and lowered user or advertiser happiness, which needs to be addressed by understanding, detecting, and treating seasonal Ads."
"The ability to proactively detect seasonal advertisements is an important and interesting problem that serves as one of the foundations for seasonal ads treatment."
"We envision MLM as a teacher for knowledge distillation, a machine labeler, and a part of the ensembled and tiered seasonality detection system."