Classifying untagged video clips into topics using transcript content and temporal proximity to known-topic clips

When a batch of unclassified video clips (e.g. bloopers, outtakes) needs topic assignment and transcripts exist, combine two signals for high-confidence classification:

1. Transcript content matching — keyword/phrase overlap with known topic titles and keeper transcripts (e.g. "poor man brain" → poor-man-brain, "South Park Commons" → spc-podcast-room-tour).

2. Temporal proximity — clips are typically shot sequentially. A blooper clip whose IMG number falls between two keepers of the same topic almost certainly belongs to that topic, even if the transcript is too short or garbled to classify on content alone (e.g. "Just want" — 2 words, but sandwiched between SPC tour keepers → SPC tour).

Workflow:

• Build a sorted list of all clips (keepers + bloopers) with their IMG numbers and known topics from the manifest.
• For each unclassified clip, check transcript first for obvious topic keywords.
• If ambiguous, look at the nearest keepers above and below in IMG order — if they share a topic, assign it.
• For multi-day shoots, day 2 clips may re-record the same topics; match by content since temporal adjacency to day 1 topics breaks down.
• Wordless/no-speech clips adjacent to topic blocks inherit that topic.

Applied this to classify 25 blooper clips across 7 topics with zero ambiguity in the final assignments.