Researchers tested if sentence skeletons predict narrative coherence — they don't

Researchers built a neural network to measure whether extracted sentence structures (called skeletons) stay consistent across paragraphs as a way to detect incoherent writing. Testing shows that models built on full sentences outperform skeleton-based models for detecting coherence problems, meaning the skeleton approach doesn't work better than existing methods.

This is a null result dressed up as progress. The researchers spent time on a hypothesis that looked plausible in theory — that narrative consistency lives in sentence structure — but the experiment showed it doesn't. That matters because it stops teams from chasing this particular dead-end approach and suggests that coherence detection still requires looking at the whole sentence, not its distilled parts. The field keeps learning that simplifying language into components loses the thing you're trying to measure.