Problem
TV contextual targeting usually stops at genre and keywords. Two dramas such as Breaking Bad and This Is Us feel completely different, and their ads should differ too. An ad that clashes with the viewer's mood (a beach vacation spot after a tense thriller) is at best forgotten and at worst annoying, so the advertiser pays for a weak placement. NBCUniversal's ad sales team needed to describe the emotional register of shows and ads, and match them across thousands of titles.
Example from the article: Yellowstone is tagged with the tones Gritty & Intense and Dark & Brooding, and its top ad matches are a truck brand, an alcohol brand and a workwear brand.
Solution in brief
Build one emotional taxonomy (themes, tones, values) from clustered keyword embeddings. An LLM tags every title from a synopsis and every ad from frames plus audio transcript, and both sides are matched by tags and embeddings. Prompts and models are chosen by testing against human-labeled tags; the winners sit in a config table, so changing them needs no code change.
Questions this case answers
- How to match promotional items to long-form media by shared emotional signals instead of categories and keywords
- How to tag a large catalog with a controlled taxonomy using an LLM and choose the prompt and model against human-labeled samples
- How to match ads to content by emotional context instead of genre and keywords
- How to tag thousands of titles and ad creatives with a standard emotional taxonomy using LLMs
Tools and methods
- Emotional taxonomy — themes, tones, values, built by clustering keyword embeddings
- Foundation models on Databricks — synopses, tagging, matching
- Prompt-variant evaluation — 3 to 4 prompts against 3 to 4 models, scored by F1
- MLflow and Unity Catalog — run lineage, model registry and config tables
- Embedding-based matching — pairs ads with shows
Google or a one-time email code. No password.