Psycholinguistics: density, depth and locality in text
The psycholinguistic and discourse-processing evidence covers three questions, from the word to the document tree: how much a text unit should carry, how deep units may nest, and how far apart related units may sit.
Steps: frame · review · state of the art · position · against · baseline · bibliography
Claims
| Claim | Grade | Where |
|---|---|---|
| Per-word reading time is linear in surprisal | measured | state-of-the-art.md §1 |
| If cost is linear in surprisal, spreading information evenly does not lower total reading cost | measured (argument) | state-of-the-art.md §2 |
| Non-uniform surprisal lowers acceptability | supported with caveat | state-of-the-art.md §1 |
| Cost linear in probability, not surprisal | contested | against.md §1 |
| Languages keep dependencies shorter than random baselines | measured | review.md §2 |
| Long dependencies slow reading in controlled sentences | measured | review.md §2 |
| In naturalistic reading, distance cost is weak except at nouns | supported with caveat | against.md §2 |
| Writing has at most degree-3 center-embedding; speech almost none | measured (corpus) | review.md §3 |
| Nesting limits depend on the language's grammar | measured | against.md §3 |
| Reading time tracks propositions more than words | supported with caveat | review.md §4 |
| A missing antecedent costs a bridging inference | measured | review.md §4 |
| Repairing model-found inference gaps improves learning | supported with caveat | state-of-the-art.md §5 |
| Old information first, new information last (topic and stress positions) | convention, untested | against.md §7 |
| Ideas high in the hierarchy are remembered better | measured | review.md §5 |
| The backward-looking center is best realized as a pronoun | measured | review.md §5 |
| Low-cohesion text helps some high-knowledge readers, depending on their skill | measured, contested | against.md §4 |
| MinLA is NP-complete in general and polynomial on trees | measured (theorem) | state-of-the-art.md §6 |
| Ordering concepts under prerequisites is NP-complete in general | supported with caveat | state-of-the-art.md §6 |
| Classic readability formulas predict adult comprehension worse than NLP models | measured | against.md §6 |
| Lab findings transfer to notes that are reread and searched | convention, untested | against.md §5 |
| Contracts carry more center-embedding, rare words and passives than nine other genres; center-embedding hurt recall most | measured (secondary) | review.md §8 |
| For example, US state insurance law fixes a minimum Flesch score of 40 to 50 | convention (statute), descriptive | review.md §8 |
| Government word caps for sentences come with no test in the sources read | convention, untested | position.md |