h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . c o g n i t i o n . 2 0 0 6 . 0 9 . 0 1 1

Ingest: Minimization of dependency length in written English (Temperley 2007)

Type: types/ingest-report.md

Classification

This is a peer-reviewed empirical paper in Cognition 105:300-333. It derives predictions from a processing theory and tests them with eleven statistical comparisons over an annotated corpus. It is a production-side corpus study. It runs no reading-time, comprehension, or rewrite experiment. Author: David Temperley, Eastman School of Music, University of Rochester. He is known for computational models of music and language cognition. The paper was accepted under Jacques Mehler's editorship.

Summary

Temperley asks whether written English shows the preference that Gibson's Dependency Locality Theory (DLT) predicts. DLT holds that integrating a word with its head costs more as the distance between them grows. Temperley measures distance in words and treats a phrase's complexity as the sum of all its head-dependent distances. From this he derives four Dependency Length Minimization Rules (DLMRs):

  1. Branch consistently.
  2. Keep left-branching constituents short in a mainly right-branching language.
  3. When the order is free, put shorter dependents closer to the head.
  4. When the order is fixed, keep the closer dependents short.

He tests them on the Wall Street Journal portion of the Penn Treebank, about 1 million words from 1989, and reruns the tests on the roughly 400,000-word Brown portion. Most predictions hold in the WSJ data:

  • Subjects are much longer when quotation verbs are inverted (9.47 vs 2.16 words).
  • Direct objects are longer than subjects (5.80 vs 3.13). The gap survives a hand-coded check limited to new-entity indefinites (8.95 vs 5.95).
  • Adverbial clauses after the main clause are longer than those before it (11.41 vs 9.76).
  • Subjects of relative clauses are shorter inside subject NPs than inside object NPs (1.41 vs 1.61).
  • Of two post-verbal adjuncts, the second is longer (5.96 vs 3.04). The same holds for the first two of three adjuncts, a case that separates DLT from Hawkins's EIC theory.
  • Subjects are shorter after a fronted adjunct, but objects are not. This rules out a simple general avoidance of long sentences.

One prediction fails. Two fronted adjuncts show a small, non-significant short-long order where DLT predicts long-short. The same pattern appears more strongly in the Brown data. The paper concludes that DLT and a rival "short-first" principle together cover the data. It argues that written word order more plausibly serves comprehension than production, but it does not test that attribution.

Contribution assessment (our opinion)

When it was published, the question was important. Comprehension evidence for distance-based processing cost was strong, but production evidence was mostly Hawkins's constituent-recognition work. This paper turned a comprehension theory into specific, refutable predictions about free syntactic choices. Tests 9 and 10 are a good control design: the predicted effect appears for subjects and is absent for objects in the same contexts. Test 11 is a well-chosen test that separates DLT from EIC. The method is simple, though. It compares mean constituent lengths with one-way F-tests and has no multivariate control for discourse status, semantic class, or genre. Only one effect has a hand-coded given/new control, and that control covers 100 cases per group. My tentative judgment is that the paper's lasting value is to show that dependency-length minimization is a measurable pressure on production. It also shows plainly where that pressure fails to predict orderings (fronted adjuncts, left-branching conjuncts, Turkish). From background knowledge, it was an early step toward later multilingual dependency-length studies. I cannot verify its priority or literature coverage from this capture.

Quotes

Test 1: Subject NPs in quotation sentences, in S–V order and V–S order Result: Average length of subject NP: in S–V order, 2.16 words; in V–S order, 9.47 words. --- kb/sources/.snapshots/temperley-2007-dependency-length-written-english.md @ sha256:f6ae5fff8a98c76918267cfafe3f9b27f7a7cd87745a8d1535e0e886a1c7da9a — Test 1, quotation-verb inversion

Test 7: Clauses with two postmodifying adjuncts Result: Average length of first adjunct = 3.04; average length of second adjunct = 5.96. (F(1, 5226) = 675.6, p < .0005) The prediction is confirmed; in clauses with two postmodifying adjuncts, the second adjunct tends to be significantly longer than the first. --- kb/sources/.snapshots/temperley-2007-dependency-length-written-english.md @ sha256:f6ae5fff8a98c76918267cfafe3f9b27f7a7cd87745a8d1535e0e886a1c7da9a — Test 7, two post-verbal adjuncts

Test 8: Clauses with two premodifying adjuncts Result: Average length of first adjunct = 3.15, average length of second adjunct = 3.48. (F(1, 1224) = 3.4, p = .06, n.s.) In this case, the prediction—that the first adjunct will be longer than the second—is not confirmed; on average, the second adjunct is slightly longer. --- kb/sources/.snapshots/temperley-2007-dependency-length-written-english.md @ sha256:f6ae5fff8a98c76918267cfafe3f9b27f7a7cd87745a8d1535e0e886a1c7da9a — Test 8, two fronted adjuncts

The dependency-length view predicts a long-short ordering here, but this was not found; rather, a short-long preference was found (though it is small and statistically non-significant), as would be predicted by the short-first principle. --- kb/sources/.snapshots/temperley-2007-dependency-length-written-english.md @ sha256:f6ae5fff8a98c76918267cfafe3f9b27f7a7cd87745a8d1535e0e886a1c7da9a — Discussion of Test 8

Would we expect to find dependency-length minimization effects in speech? --- kb/sources/.snapshots/temperley-2007-dependency-length-written-english.md @ sha256:f6ae5fff8a98c76918267cfafe3f9b27f7a7cd87745a8d1535e0e886a1c7da9a — Discussion; the habit suggestion below is framed for speech production

it may be that strategies for dependency-length minimization such as those discussed above—the avoidance of long subject NPs, for example—are acquired by speakers as general habits, so that elaborate length calculations are not required on every occasion. --- kb/sources/.snapshots/temperley-2007-dependency-length-written-english.md @ sha256:f6ae5fff8a98c76918267cfafe3f9b27f7a7cd87745a8d1535e0e886a1c7da9a — Discussion, speech production

Connections Found

For the occasion, this source is the operational anchor among the dependency-locality captures. It names specific constructions where skilled writers shorten dependencies, so it is the closest available material to word-order rules a writer could apply. It does not directly support the claim that distance causes processing cost. That link is imported from Gibson's comprehension work, which the paper cites but does not test.

It bounds Two rewrites exposed a syntax-or-repetition tradeoff. That note found two ways to lower parsing load in mature prose: splitting, which lost relations, and restating, which added 4.9% more words. Temperley documents a third move in human writing. Reordering constituents keeps the syntactic relations and shortens dependencies without adding words. The paper measures what writers chose, not whether the result was easier to understand. Reordering is therefore an untested candidate for the note's tradeoff, not a refutation of it.

It is evidence for the human side of Human writing structures transfer to LLMs because failure modes overlap. At sentence scale, a writing convention (dependency-shortening order) tracks a named human processing cost (activation decay over intervening words). The note's per-convention transfer test then needs an LLM-side check that this source cannot supply.

It compares with Local materialization should outperform distant natural-language declarations on one axis only: distance between a dependent element and what it depends on as a cost. Temperley gives a human mechanism at sentence scale. The note is a document-scale conjecture about LLM readers and explicitly does not assume a monotonic decline with distance. The comparison keeps the scales and the readers apart. It does not support the note.

No other tracked KB note addresses sentence-level syntactic processing cost.

Extractable Value

  1. Candidate word-order rules with their corpus support -- the writer-facing rules the paper supports, each as "skilled writers do this", not "this improves reading":
  2. put the heavier of two post-head phrases last (Tests 7 and 11; the heavy-NP and dative cases are discussed but not tested here);
  3. keep pre-verbal subjects short, and move elaboration such as appositives and non-restrictive relative clauses to object position or a later mention (Tests 2 and 3, and the Ortega example);
  4. put long adverbial clauses after the main clause (Tests 4 and 5; the effect is modest, about 1.5 words);
  5. keep the subject short after a fronted adjunct (Test 9);
  6. avoid relative clauses nested inside subjects, which create center-embedding (Test 6, which does not replicate in Brown).

These rules preserve relations, so they fit the gap that the syntax-or-repetition note leaves open better than splitting does. Their comprehension benefit is inferred from DLT, and their benefit for LLM readers is untested. [experiment] 2. What the source establishes about distance and cost -- it establishes a production correlation. Written English shows the constituent-length asymmetries that a distance-cost theory predicts, at very large N, and the Brown data mostly agree. It does not establish that distance causes processing cost. Guidance written from this paper should cite comprehension evidence, such as Gibson's DLT studies, for the cost claim and cite this paper only for "writers already do this". [just-a-reference] 3. A rule the evidence does not support: "long before short" for fronted material -- DLT predicts that the longer of two fronted adjuncts should come first. Writers do the reverse, weakly in the WSJ data and strongly in Brown. Guidance that follows dependency-length arithmetic mechanically would recommend an order skilled writers avoid. A "short-first" default plus the DLT-specific rules for subject length covers the observed choices better. [quick-win] 4. Two separable metrics -- dependency length can be counted in words or in new discourse referents. Complexity can be the sum of distances or the peak integration cost. The paper uses words and the sum, while Gibson uses referents and the peak. A writing rule or lint check must choose one, and the choices can rank sentences differently. [just-a-reference] 5. Habit as the production mechanism -- the paper suggests that speakers acquire dependency-minimizing strategies as habits, for example avoiding long subjects, instead of calculating lengths. The suggestion is framed for speech production; applying it to writing is an extension. This matches how guidance could be stated: as a few construction-level heuristics, not as an instruction to compute dependency lengths. [just-a-reference]

Limitations (our opinion)

  • Production evidence, not processing evidence. The tests count what writers produced. The step from "writers prefer shorter dependencies" to "shorter dependencies are easier to read" rests on cited comprehension work and on the author's argument that writing lacks speech's time pressure. No reader was measured, and the paper says the production-versus-comprehension question is unresolved. For the occasion, this means the paper can justify a rule as conventional skilled practice but not as a measured readability gain.
  • No evidence about LLM readers. The proposed cost mechanism is activation decay in human working memory. Nothing here shows that a transformer pays a cost that grows with intervening words. Human writing structures transfer to LLMs because failure modes overlap requires that check for each convention. Guidance for mixed human and agent readers can rely on the human side only.
  • Simple statistics and partial confound control. Each test compares mean lengths between two groups. There is no regression on discourse status, definiteness, semantic type, or genre. Given/new is controlled only in Test 3, on 100 hand-classified cases per group coded by one person. Argument versus adjunct is approximated from Treebank function tags. The author names discourse factors as an open alternative. Very large N makes small differences significant. Test 6's 0.2-word difference has p < .05, and statistical significance does not show that a rule is worth stating.
  • Narrow, edited genre. The WSJ data is 1989 house-edited news prose. The Brown rerun widens the genre, but the paper does not give its significance tests. Brown fails to replicate Test 6. As extracted from the table, Brown object NPs in Test 10 are longer with a fronted adjunct (5.09 vs 4.54), which the text does not discuss. Commonplace's technical, list-heavy, heavily qualified prose is not represented.
  • The rival explanation is not ruled out. A short-first principle explains the adjunct and conjunct orderings and the failed fronted-adjunct prediction. DLT is needed only for short left-branching constituents and for the Test 9 and 10 asymmetry. The DLT-versus-EIC comparison is "far from resolved" in the author's words, and the Turkish test rests on 87 cases.
  • Heavy-NP shift and the dative alternation are not tested here. The paper explains them as predicted by DLMR 3 and cites other authors for the data. Do not cite this paper as corpus evidence for those two constructions.

Record one entry in the "Ideas to test later" section of kb/reports/retained/asd-ste100-inspired-rewrite.md. The entry would propose a constituent-reordering pass over one KB note. The pass would apply heavy phrases last, short subjects, long adverbial clauses after the main clause, and short subjects after fronted adjuncts, and it would measure meaning preservation and word count against the two earlier rewrites. Do this before any dependency-length rule enters a writing instruction. This source supports the rules only as observed human practice, and their value for agent readers has no test yet.

Relevant Notes: