SAVER: A Benchmark for Structured Understanding of E-Commerce Advertisement Videos

Producing high-quality e-commerce advertisement videos is costly. This motivates reusing the creative structure of proven, high-performing advertisements to promote new products. Such reuse requires understanding not only where the original product appears, but also how it is connected to demonstrations, spoken claims, on-screen text, and the overall commercial message. Existing advertisement benchmarks mainly evaluate answers to individual questions, without assessing whether models can produce a consistent, structured representation of an entire video for this purpose. We introduce SAVER, a benchmark for structured evaluation and representation of e-commerce advertisement videos, containing 6,224 videos across 37 top-level and 283 second-level industry categories. Each video is paired with a StructuredScript annotation that organizes global context, detailed content, and commercial information. Shared entity identities and timestamps connect content across shots and modalities, while evidence references connect commercial interpretations to supporting observations. We further introduce Qwen-SAVER, a Qwen2.5-Omni-7B model fine-tuned on five script-generation tasks. The model first describes the video content and then uses these descriptions to infer commercial information. We evaluate Qwen-SAVER alongside eight general-purpose baselines on 500 test videos. Compared with the untuned model, Qwen-SAVER improves the schema-conformance rate and commercial-claim annotation-correspondence micro-F1 by 81.8 and 40.6 percentage points, respectively. Finally, a qualitative study explores how field-level modifications to a StructuredScript can guide product-replacement video generation, illustrating its potential for advertisement reuse.