AIGC Content Analysis Agent for E‑commerce

We present a content analysis Agent designed for Taobao’s e‑commerce ecosystem, leveraging low‑cost, high‑speed AIGC generation to address the challenges of producing effective product‑related content across feeds, search, and detail pages.

The Agent automatically constructs a fine‑grained content feature taxonomy based on production goals, performs multimodal automatic labeling of images, videos and text, and extracts high‑value content features using causal inference and statistical analysis.

Key modules include:

Feature‑system generation: Decomposes user‑defined objectives (e.g., high‑quality AIGC videos) into dimensions such as visual, audio, and copy, and builds a clear tag hierarchy.

Fine‑grained labeling: Combines multimodal models and large‑language models to annotate large volumes of content with detailed tags, reducing manual effort.

Quality‑feature mining: Identifies which tag combinations drive click‑through and conversion rates across different channels and user segments.

Additional capabilities comprise an Auto‑Prompt framework that optimizes prompts without training data, a memory module that reuses historic taxonomies, and an automated model training‑deployment pipeline supporting LoRA fine‑tuning.

Experimental results show >90% labeling accuracy on 30+ tags, a reduction of labeling labor by over 80%, and a double‑digit increase in CTR for image assets. The system also improves video generation efficiency, achieving near‑GPT‑4o performance with 25% of the cost.

Future work will focus on further improving labeling precision, expanding the taxonomy to new content domains, and tightening the closed‑loop from analysis to generation.