H-Score

A continuous discriminator variable that scores how “highlight-like” a given text fragment is.

Unlike binary classification (highlight/not-highlight), the h-score provides a similarity metric ranging from -1 to 1, where higher values indicate greater likelihood that content would be selected as a highlight by human experts.

The score enables ranking, instead of hard boundaries, you can ask “what are the top N most highlight-worthy segments in this document?” This is more useful for practical applications where you need to surface the most valuable content, not make binary yes/no decisions.

Related metric: h-coverage, the percentage of a document that users turn into highlight clips. Short videos tend toward 100% h-coverage; long videos cluster near 0%.

Related: 05-atom—highlight-vs-summarization