h t t p s : / / s o c s c i . u c i . e d u / ~ r f u t r e l l / p a p e r s / f u t r e l l 2 0 1 5 l a r g e s c a l e . p d f

Ingest: Large-scale evidence of dependency length minimization in 37 languages

Type: types/ingest-report.md

Classification

This is a peer-reviewed empirical paper (PNAS 2015, with its Supporting Information and two later published corrections) that tests a quantitative hypothesis against corpus data and counterfactual baselines. Author: Richard Futrell, Kyle Mahowald and Edward Gibson, MIT Brain and Cognitive Sciences. Gibson is the originator of dependency locality theory, so the authors have a stake in the hypothesis they test.

Summary

Dependency length is the distance in words between a syntactic head and its dependent. Dependency length minimization (DLM) is the hypothesis that speakers and grammars prefer orders that keep it short. The usual explanation is processing cost: in a left-corner parser, an unresolved head or dependent must be held in memory until its partner arrives. The paper takes hand-parsed or hand-corrected treebanks of 37 languages from 10 families. For each sentence it compares the summed dependency length of the observed order with 100 random projective reorderings of the same tree. One baseline lets word order vary freely. A more conservative baseline fixes order by relation type. The Supporting Information adds a consistent-head-direction baseline and a fixed-head-position baseline. In every language, observed length grows more slowly with sentence length than in the random baselines. Observed length also stays above the optimal projective ordering. The authors attribute that gap to competing desiderata, such as letting the reader recover structure incrementally. Effect size varies: head-final, case-marking languages (Japanese, Korean, Turkish) minimize less than head-initial ones (Italian, Indonesian, Irish). The authors conjecture that rich morphology makes long dependencies easier to identify. The result is conditional on the tested decomposition. The tree is held fixed and only the linear order varies, so the paper measures minimization through word order. It does not measure restructuring, such as splitting sentences. It also does not separate grammar from usage, and its texts are mostly edited written prose.

Contribution assessment (our opinion)

The problem matters. DLM had been invoked to explain word-order universals, projectivity and parser design, but the evidence covered a few languages or categorical descriptions, and earlier corpus work found only weak evidence in German. This paper turns a widely assumed generalization into a broad, falsifiable test. Observed lengths at chance would have counted strongly against DLM as a universal explanation. The test is credible for three reasons. The baselines hold the tree fixed. The fixed-order and head-direction baselines answer the most obvious rival, that the effect is a by-product of consistent head direction. The analysis models growth with sentence length rather than relying on length-mixed averages. Its demonstrated value is a robust corpus regularity. The memory-cost mechanism is cited from prior work, not tested here, and so are the competing-desiderata account of the gap to optimal and the morphology account of cross-language variance. On my background knowledge, this is a landmark scale-up of earlier work by Gildea and Temperley, Liu and Ferrer i Cancho, not a new hypothesis. The paper's own correction acknowledges Liu's 20-language precedent. Annotation-scheme dependence and reliance on edited text narrow the claim but do not undermine it.

Quotes

Note that the 100 random linearization all have the same underlying dependency structure as the original sentence, just with a potentially different linear order. --- kb/sources/.snapshots/futrell-dependency-length-minimization.md @ sha256:7fe22ec2d681fa769f1250f909ef0c6662023b23a4121db080595a4f44572939 — Results, Free Word Order Baseline

As the figure shows, all languages have average dependency lengths shorter than the random baseline, especially for longer sentences. --- kb/sources/.snapshots/futrell-dependency-length-minimization.md @ sha256:7fe22ec2d681fa769f1250f909ef0c6662023b23a4121db080595a4f44572939 — Results, Fig. 2

whereas observed dependency lengths are always shorter than the random baselines, they are also longer than the minimal baselines (although some languages such as Indonesian come quite close). --- kb/sources/.snapshots/futrell-dependency-length-minimization.md @ sha256:7fe22ec2d681fa769f1250f909ef0c6662023b23a4121db080595a4f44572939 — Discussion

In general, we believe dependency length should not be fully minimized because of other factors and desiderata influencing languages that may conflict with DLM. For example, linearizations should allow the underlying dependency structure to be recovered incrementally, to allow incremental understanding of utterances. --- kb/sources/.snapshots/futrell-dependency-length-minimization.md @ sha256:7fe22ec2d681fa769f1250f909ef0c6662023b23a4121db080595a4f44572939 — Discussion

For example, language users might structure discourse (splitting ideas into multiple sentences), drop optional syntactic elements such as pronouns, or choose favorable word orders. Our work aims to show specifically that languages and language users universally prefer word orders that minimize dependency length, given the dependency tree structures of the sentences they want to express. --- kb/sources/.snapshots/futrell-dependency-length-minimization.md @ sha256:7fe22ec2d681fa769f1250f909ef0c6662023b23a4121db080595a4f44572939 — Correction (omitted ref. 43 statement)

Connections Found

For the proposed writing guidance, this ingest is human-reader background and a bound, not a warrant. It establishes that human languages order words so that linked words sit closer than chance; the link from distance to processing cost is cited from prior work. No KB note yet cites dependency locality or psycholinguistic distance cost, and the ingest does not corroborate any existing claim. The connection that governs the occasion is the transfer test below: guidance for LLM-read prose needs it, and this paper supplies only the human half. Its closest role is a counterpoint-and-bound for Two rewrites exposed a syntax-or-repetition tradeoff (compares-with: both treat distance between linked elements as a cost traded against other desiderata). That note found that flattening syntax either severed attachments or forced restatement. Futrell et al. show that human languages shorten linked distances far below chance but stop short of the optimum. The comparison has one asymmetry. The paper holds the dependency tree fixed and varies only order. The rewrite experiment changed the structure itself by splitting sentences, which is the option the paper excludes. The second connection is Local materialization should outperform distant natural-language declarations (compares-with: whether distance between a dependent element and what governs it raises the cost of correct use). The paper supplies a measured human precedent and a memory-store mechanism at word scale. The note's conjecture concerns LLM readers at section scale, so the paper is not evidence for it. Any use of this source for LLM-facing prose rules, such as the AGENTS.md rule to keep necessary conditions next to their claims, rests on the per-convention test in Human writing structures transfer to LLMs because failure modes overlap. That test asks whether LLMs show the failure the convention prevents for humans. This paper supplies only the human half.

Extractable Value

  1. What the paper establishes about distance and cost -- it establishes a corpus regularity: in all 37 languages, summed head-dependent distance grows more slowly with sentence length than in random reorderings of the same trees. It does not measure processing cost. The cost link (a left-corner parser holds an unresolved head or dependent in memory until its partner arrives, and difficulty is observed at long dependencies) is cited from earlier work. Guidance that cites this paper should claim "languages and their users favor short dependencies", not "this paper shows long dependencies slow readers". [quick-win]
  2. One supported ordering rule: short before long after the head -- the Supporting Information's fixed-head-position baseline shows that, with each dependent's side of the head fixed, dependents are ordered long-before-short before heads and short-before-long after heads in every language. For English, which mostly places complements after the head, the candidate writer rule is: after a verb or noun, put the shorter dependent first and the longer phrase or clause last, and keep a modifier beside the word it modifies. The evidence is that observed text follows this pattern above chance, not that a writer who applies it improves comprehension. [quick-win]
  3. Locality is a pressure, not a target to maximize -- observed lengths sit between random and optimal in every language, and the authors explain the gap by competing needs such as incremental recoverability: a reader should be able to tell at each word which earlier word it attaches to. Guidance should therefore say "reduce distance when the order stays easy to parse", not "minimize distance". The same bound applies to the AGENTS.md rule to keep necessary conditions next to their claims. It matches the rewrite note's finding that forced locality costs repetition. The transfer from human word order to KB prose stays a conjecture. [quick-win]
  4. Order-only versus restructuring is a distinct lever -- the paper measures only what reordering a fixed tree achieves. Its correction notes that users might also shorten dependencies by splitting ideas into multiple sentences or dropping optional elements, which the design does not measure. Writing guidance should name its lever. Reordering within a sentence keeps relations and has the evidence above. Splitting a sentence has no support here and risks severing a qualifier from its claim; the rewrite experiment's failures were of this kind. [experiment]
  5. Explicit marking can buy back distance -- the head-final, case-marking result and the cited German finding suggest that overt marking of a relation lowers the cost of a long dependency. A KB analogue would be explicit restatement or a labelled reference standing in for adjacency, which is the alternative the rewrite note's second pass chose, and it parallels local materialization. This is the authors' post hoc conjecture, not a tested result. [deep-dive]
  6. Methodological template: counterfactual baselines over the same structure -- comparing observed artifacts with random rearrangements of identical content is a reusable design for testing whether a KB's authored order reflects a locality pressure. It would also apply to an LLM-reader experiment that varies distance while holding content fixed, as the local-materialization note proposes. [just-a-reference]

Limitations (our opinion)

The paper establishes a corpus regularity, not its cause. The memory-store mechanism, the incremental-recoverability account of the gap to optimal, and the morphology account of cross-language variance are all cited or conjectured, not tested. A simpler rival is that some ordering conventions shorten dependencies for reasons other than processing cost. The head-direction baselines rule out consistency of head direction, but not every alternative. Results depend on one family of annotation conventions. The authors report robustness to content-head formats, but function-word analysis is unsettled, and different head choices change measured lengths. Distance is counted in words, the analysis is restricted to projective linearizations, and the fixed-order baseline is deterministic and cannot model phenomena such as German clause-dependent order. The texts are mostly edited newspaper, novel and blog prose. Japanese is spoken dialogue, and Latin and Ancient Greek include much poetry, which confounds cross-language comparison with genre. The study cannot separate grammar from speaker choice, so it does not say whether an individual writer's choices minimize. For this KB there are two further gaps: a scale gap, from words within a sentence to conditions across a note, and a reader gap, from human processors to LLMs. Neither is addressed. The paper offers no support for an LLM-facing prose rule until a per-convention transfer test, as in Human writing structures transfer to LLMs because failure modes overlap, finds the same distance cost in models.

Before any writing instruction changes, draft one theory note in kb/theory/ that states the candidate rule "after a head, put the shorter dependent first and keep a modifier beside what it modifies; reduce distance only while the order stays easy to parse" and applies the per-convention test from Human writing structures transfer to LLMs because failure modes overlap. The note should cite this ingest only for the human-corpus half, as a regularity whose cost mechanism is cited rather than tested here. It should record that LLM-reader evidence is still missing.

Relevant Notes: