Content Understanding for Advertising on Weibo: Challenges, Solutions, and Applications

Guest: Chen Zhaoji, senior algorithm expert at Weibo (source: DataFunTalk).

Overview: While algorithm engineers often say "data is king," advertisers rely on content understanding as the foundation for effective ad placement. This talk outlines the role of content understanding in Weibo advertising.

1. Advertising System and Commercialization Overview

Advertising involves three parties—advertisers, the platform, and users—each with distinct profit motives. The platform's core task is to maximize advertiser ROI while minimally disrupting user content consumption.

2. Core Tasks of Ad Delivery

Who is watching: tracking users, collecting historical data, and building user profiles.

What is being watched: understanding the content the user is currently consuming.

What is suitable to show: matching ads to the user's current interests based on the above signals.

3. Non‑Content Scenarios

In feeds such as "following" or "trending," content is diverse and loosely related, so ad placement focuses on user selection rather than strict content relevance.

4. Content Scenarios

In pages, comments, or search results, ads must be contextually relevant, making content understanding critical.

5. Problems Caused by Insufficient Content Understanding

Marketing content regulation: natural posts with overt marketing confuse users; large posting volume and lack of labels make detection hard.

Timing of content display: emotional mismatch between adjacent natural posts and ads can harm user experience; new slang and expressions add difficulty.

Noise in performance evaluation: undiscriminated content leads to artificially high or low click‑through rates, requiring continuous model updates.

Homogenized marketing content: similar ads from different accounts cause user fatigue; defining similarity varies across scenarios.

Mismatch between natural and marketing content: many popular natural posts lack clear commercial signals, making ad targeting challenging.

Long‑tail advertisers lack high‑quality creative assets; automated creative generation is needed but current models lag behind requirements.

6. Building Content Understanding Capability and Business Applications

6.1 Content Understanding Tasks

Word segmentation / entity recognition (people, places, brands).

Sentiment analysis (detecting positive/negative sentiment and handling global vs. brand‑specific negativity).

Similarity judgment (determining if two pieces of content are alike).

Content classification (building a commercial tag taxonomy for text and images).

Specific content identification & generation (keyword spotting, creative pattern detection, intelligent creative generation, celebrity‑brand knowledge graph construction).

6.2 Content‑Based Advertising

Tag construction: create discriminative, appropriately granular tags for content.

Content annotation: two models are used—(1) a fast response model based on public corpora, word vectors, and inverted indexes; (2) a deep BERT‑based model fine‑tuned on Weibo data, cached for low‑latency inference.

Ad delivery based on content: Implicit optimization: within existing audience targeting, further refine ad selection using the content tags of the page the user is viewing. Explicit optimization: expose content tags to advertisers so they can purchase ad slots tied to specific tags, independent of user profile.

7. Celebrity‑Brand Knowledge Graph Construction and Use

The graph links entities (celebrities, works, brands) and relationships (endorsements, participation, sponsorship). Built on Neo4j/CQL, it enables queries such as retrieving brands endorsed by a celebrity (e.g., Zhu Yilong → Wei Quan yogurt) and avoiding competitor ads in related content.

8. Summary and Outlook

Unified content analysis and representation for text, images, and video (segmentation, keyword extraction, vectorization).

Unified vector representations fine‑tuned for specific tasks and fused across modalities.

Basic content understanding services (similarity, classification) built on these vectors.

Flexible support for commercial scenarios such as implicit content optimization and explicit content sales.

For more details, see the accompanying video and related articles linked at the end of the original post.