Review: psycholinguistics

This page lists each source on density, depth and locality in sentences, with what it found. It opens with the surveys read first; its eight strands then run from information rate to practice.

Order inside each strand: definitions, seminal work, forward citations.

Surveys read first

Gibson et al. 2019: efficiency as one explanation for lexicon, syntax and processing; vocabulary for strands 1 and 2.
Temperley & Gildea 2018: corpus, simulation and comprehension evidence for shorter dependencies.
McNamara & Magliano 2009: comparison of text comprehension models, construction-integration among them.
Díaz et al. 2002: graph layout problems, with complexity, bounds and heuristics.

1. Information rate

Surprisal of word w_i: s_i = -log2 P(w_i | w_1 ... w_(i-1)); Hale 2001 proposed it as processing cost.
Levy 2008 derived the same cost as the size of the belief update over interpretations.
Linking function f: the map from s_i to processing cost.
Uniform information density (UID): speakers prefer to spread information evenly (Levy & Jaeger 2007; Jaeger 2010).
Entropy rate constancy (ERC): information per sentence, given its context, stays constant through a text (Genzel & Charniak 2002).
Jaeger 2010: in speech corpora, the optional complementizer "that" is kept more often where the clause onset is less predictable.
Genzel & Charniak 2002: out-of-context sentence entropy rises with sentence position, as a constant in-context rate predicts (n-gram models).
Smith & Levy 2013: reading time is linear in surprisal over six orders of magnitude of predictability.
Meister et al. 2021: reading times fit a weakly super-linear f; acceptability shows a clearer penalty for non-uniform surprisal.
Hoover et al. 2023: with better language models, reading time grows super-linearly in surprisal, and so does its variance.
Brothers & Kuperberg 2021: self-paced reading (N = 216), picture naming (N = 36) and an eye-tracking meta-analysis (N = 218) show cost linear in probability.
Wilcox et al. 2023: in 11 languages from 5 families, surprisal predicts reading time in all of them and contextual entropy adds to it in most (significant when pooled), with a linear link.
Xu et al. 2023: in eye-tracking corpora of 7 languages, super-linearity appears in some, and depends strongly on the language model.
Shain et al. 2024: 6 reading datasets, many language models, nonlinear regression: f is linear in surprisal.
Buggy et al. 2026 (preprint): the materials of Brothers & Kuperberg 2021, read with eye-tracking, give a logarithmic effect of probability.
Verma et al. 2023: neural language models do not reproduce clear entropy rate constancy.
Tsipidi et al. 2024: surprisal contours of long discourses follow position in a hierarchical discourse structure.

2. Sentence locality

Dependency length: distance in words between a head and its dependent; D is the sum over a sentence.
Dependency locality theory (DLT): linking a word to a distant head costs in proportion to the new discourse referents in between (Gibson 1998; Gibson 2000).
Yngve depth: constituents a left-to-right producer has promised but not yet built; Yngve 1960 proposed a limit near seven.
Grodner & Gibson 2005: two self-paced reading studies: integration difficulty grows with intervening material.
Demberg & Keller 2008: in an eye-tracking corpus, DLT cost does not predict reading time for words in general, only for nouns.
Lewis & Vasishth 2005: cue-based retrieval with decay and interference, an alternative to counting distance.
Futrell et al. 2015: in 37 languages, D is shorter than random projective reorderings and longer than the projective optimum.
Ferrer-i-Cancho et al. 2022: an optimality score over 93 languages: half reach 70% or more of the possible reduction.
Futrell et al. 2020: locality follows from surprisal computed over a lossy memory of the context.
Sampson 1997: an English corpus shows a bias against left-branching, but no sharp depth limit.

3. Nesting depth

Center-embedding: a clause inside another, with outer material on both sides; its degree counts nested clauses.
Miller & Chomsky 1963: center-embedded sentences are grammatical; their difficulty is a limit of finite memory.
Miller & Isard 1964: self-embedded sentences are harder to learn in free recall (24 listeners, 5 trials; spoken, not written). Embedding effects are small beside the effect of scrambling the words.
Karlsson 2007: 119 multiple center-embeddings in 7 European languages: degree at most 3 in writing, practically none in speech.
Gibson & Thomas 1999: doubly nested relative clauses missing the middle verb phrase are rated as acceptable as complete ones.
Vasishth et al. 2010: in seven experiments, dropping the middle verb eases English reading and hurts German reading.
Hahn et al. 2022: a model trading memory precision against prediction fits nesting difficulty in English, Spanish and German data.

4. Text

Proposition: a predicate with its arguments; the text base is the network of a text's propositions.
Situation model: what the reader builds about the topic, from text base plus knowledge.
Argument overlap: two propositions share an argument; reinstatement search: retrieval from memory when nothing in the buffer overlaps.
Kintsch & Keenan 1973: with sentence length fixed at 16 words, each extra proposition added about 1 s of reading time (about 1.5 s per proposition recalled). The paper does not pit propositions against words as competing predictors. Under restricted reading time, recall no longer rose with propositions presented.
Kintsch & van Dijk 1978: cycles through a limited buffer, coherence by argument overlap, macro-operators that reduce to gist.
Miller & Kintsch 1980: the model applied to readability of short passages. Abstract only: 20 texts and 600 undergraduates, with readability as reading time per proposition recalled. Predictions of readability, reading time and recall reached r = .8 to .9.
Kintsch 1988; Kintsch 1998: construction-integration: bottom-up activation, then integration by spreading activation.
Haviland & Clark 1974: the Given-New Strategy. A target sentence is understood faster (button press, about 140 to 180 ms) when its Given part has a direct antecedent (3 experiments, N = 16, 10, 27). The measure is self-judged comprehension time; the authors warn against taking absolute differences very seriously.
Britton & Gülgöz 1991: the model located inference calls in a 1000-word text; each was repaired into a principled revision.
Gülgöz et al. 1999: revisions that add missing links help learning, most of all for readers short of inference skill, knowledge or motivation (secondary).
Gopen & Swan 1990: old information in the topic position first, new information in the stress position last, subject near its verb.

5. Text tree

Content structure: the hierarchy of idea units; the levels effect: units high in it are recalled better.
Meyer & McConkie 1973; Meyer 1975: higher idea units recalled more often (two passages, 69 undergraduates; secondary).
Meyer et al. 1980: good ninth-grade comprehenders follow top-level structure; signaling helps underachievers.
Hyönä et al. 2002: adult readers who attend to headings have the largest working memory and write the best summaries.
Wu et al. 2016: in eye-tracking, a changed word slowed reading at the top of a three-sentence hierarchy, not at the bottom.
Rhetorical structure theory (RST): a text is a tree of relations between nucleus and satellite spans (Mann & Thompson 1988).
Centering: each utterance has one backward-looking center linking it to the previous one (Grosz et al. 1995).
Gordon et al. 1993: in five reading-time experiments, that center is best realized as a pronoun, usually the subject.
Poesio et al. 2004: in corpora, the pronoun rule holds under most settings; the uniqueness constraint depends on settings.
Entity grid: a sentence-by-entity table of grammatical roles; it ranks orderings and summaries by coherence (Barzilay & Lapata 2008).

6. Reader dependence

Reverse cohesion effect: high-knowledge readers learn more from a low-cohesion text (McNamara et al. 1996b).
McNamara & Kintsch 1996a: high coherence helped recall; high-knowledge readers did better on open questions after low coherence.
O'Reilly & McNamara 2007: with 143 students, the low-cohesion benefit held only for less skilled high-knowledge readers.
Expertise reversal effect: support that helps novices can hurt more experienced learners (Kalyuga et al. 2003).
A 1998 study reviewed in Kalyuga 2003: explanatory notes in a text helped readers with low language competence and slowed more knowledgeable readers.
Davison & Kantor 1982: splitting sentences removes connectives, which shifts the inference load to the reader; some of the best adaptations made sentences longer.

7. Ordering and measurement

Minimum linear arrangement (MinLA): vertices on a line, minimizing the sum of edge lengths; results in state-of-the-art.md.
Flesch 1948: reading ease from sentence length and syllables per word.
Coh-Metrix: Graesser et al. 2004; McNamara et al. 2014; Graesser et al. 2011; open tools in baseline.md.

8. Practice

Kinds: G government guide, L law, T tool, P paper. Practice is convention, untested unless a study is named; adoption figures are measured, descriptive.

EC 2011 (G): average of 20 words per sentence, with a few short ones; keep link words when splitting; the guide calls its advice "hints, not rules".
Style Manual AU 2024 (G): "an average of 15 words and no more than 25 words"; read on the archived page (updated 2024-12-19), which cites no study.
GSA 2026 (G): gives no word count and cites no study; it grounds the advice in the Plain Writing Act of 2010.
NAIC 2025 (L): a chart of US state insurance rules; minimum Flesch score 40 in WV, NH, MT and NV, 45 in FL and NY, 50 in ME and NC.
Vale 2026 (T): averages the listed metrics (Gunning Fog, Coleman-Liau, Flesch-Kincaid, SMOG, ARI); a grade parameter sets the highest acceptable level; owner documentation.
Hemingway 2026 (T): grade score; the Editor Plus default target is grade 9; the page names no formula; owner documentation.
textstat 2026 (T): MIT-licensed readability library, 1.4k GitHub stars on 2026-10-28; its README calls SMOG on texts under 30 sentences statistically invalid.
Redish 2000 (P, commentary, read in full): formulas were built for children's school books and ignore reader differences, layout and retrieval aids.
Redish 2000 reports that about half of 36 studies improved comprehension by improving scores. Revised jury instructions were understood better yet scored worse. A very poor score remains a red flag.
Schriver 2017 (P, secondary): a review of 1940 to 2015 finds formulas often unreliable and invalid; usability testing is the best benchmark; read on an advocacy page.
ISO 2023 (standard, preview read): plain language focuses on how well readers can use a document, not on mechanical measures such as readability formulas (p. v).
Martinez et al. 2022 (P): in about 10 million words, contracts carry more rare words, center-embedding (0.729 vs 0.272 per sentence, OR = 2.56), passives (0.758 vs 0.160, OR = 4.35) and non-standard capitalisation than nine other genres.
Martinez et al. 2022: two experiments pooled to N = 184 (108 retained in the first). Legalese lowered recall (42.4% vs 35.3%) and comprehension (beta = -0.179), also for readers with high reading experience, and the comprehension result replicated with the text in view and a bonus. Center-embedding showed the strongest register effect in an exploratory model with insufficient power per structure. Excerpts were written by a lawyer-author and read by laypeople online, and recall was delayed. The authors link the cost to working memory for long-distance dependencies.