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 2 0 d e p e n d e n c y . p d f

Ingest: Dependency Locality as an Explanatory Principle for Word Order

Type: types/ingest-report.md

Classification

A peer-reviewed empirical linguistics paper (Language 96(2), 2020). It tests a stated hypothesis against constructed null baselines on large parsed corpora and introduces a corpus method for separating grammar from usage. It also reviews prior theory and experimental evidence.

Author: Richard Futrell (UC Irvine), Roger P. Levy and Edward Gibson (MIT). Gibson originated the dependency locality theory of sentence processing, and Levy originated surprisal theory. The authors are central figures in this literature and advocates of the efficiency hypothesis they test.

Summary

The paper tests dependency-length minimization (DLM): language users prefer word orders in which words linked by a head–dependent relation are close in linear order. It uses Universal Dependencies 2.1 corpora for 53 languages, mostly written text (news, blogs, fiction, nonfiction, Wikipedia). Dependency length is measured as intervening words plus one. The dependency formalism is held fixed (UD with function-word-head transformations, replicated with original content-head parses and with function words removed). First, observed dependency length grows more slowly with sentence length than in six random-reordering baselines built from independently motivated constraints (projectivity, consistent head direction, fixed word order). The difference holds at p < 0.001 in regression for all languages and baselines; sign tests are significant in 49–53 of 53 languages depending on the baseline. The authors bound this result: whatever constraints exist reduce dependency length beyond these baselines, and unrun constraints could still explain it. Second, they fit probabilistic linearization models of grammatical order from the corpora. Observed orders are shorter than sampled grammatical reorderings, which they read as DLM in usage. Grammatical reorderings are shorter than the projective baseline, which they read as DLM in grammar. That split depends on the linearization model, which captures only order among a head's immediate dependents and only projective orders. Third, dependency length varies across languages and correlates with head-finality; the authors offer given-before-new ordering as a tentative explanation and leave the pattern open. The paper does not measure processing cost. It reviews the cost evidence and theories: reading-time slowdowns when a subject–verb dependency is lengthened (Grodner and Gibson 2005), antilocality effects in verb-final languages, working-memory decay and interference, parser search time, statistical complexity, and information locality from lossy-context surprisal. It explicitly takes no position on which motivation is correct. It also reviews derived ordering predictions: short-before-long constituents after a head in head-initial languages such as English, long-before-short in head-final ones, and stronger closeness pressure for word pairs that predict each other more.

Contribution assessment (our opinion)

The problem is important. Whether a single processing-cost principle explains word-order universals is a central question for functional linguistics, and the efficiency hypothesis needs tests that can fail. The paper's main advance is methodological and evidential, not a new conjecture: DLM dates to Behaghel, and corpus comparisons with random baselines predate this work, including the authors' own 2015 study. What it adds is a stacked set of independently motivated null baselines across 53 languages, an explicit narrowest interpretation that names unrun rival constraints, and a model-based way to separate grammar from usage in corpus data. That split answers a standard objection that corpora reveal only usage. The breadth makes "DLM beyond these baselines" well corroborated. The grammar-versus-usage attribution is weaker because it inherits the linearization model's local-subtree, projective decomposition, and the authors acknowledge this. The head-finality finding is a useful documented anomaly rather than an explanation. Our guess is that the paper's durable value is as a reference test design and a consolidation of DLM evidence, while the causal question (DLM itself versus rival constraints) stays open. Our originality judgment rests on the paper's own literature review and general background, not on a literature search.

Quotes

For now, we focus on showing large-scale evidence for DLM as an empirical principle driving quantitative patterns of word order. We do not take a position on the precise functional motivation for DLM: we believe such an argument would be best served by detailed experimental studies, rather than the relatively coarse-grained approach we take here. --- kb/sources/.snapshots/futrell-levy-gibson-dependency-locality.md @ sha256:1c2cef7ae6abaa7a2d5f8d80db1814a62740db0e2f1d674b15a9abdbae8ce8d1 — §2.2, after its review of parser-search, memory decay and interference, lossy-context surprisal and statistical-complexity motivations

There is a strong tendency in predominantly head-initial languages such as English to order constituents after a head from short to long (Behaghel 1909, Wasow 2002). --- kb/sources/.snapshots/futrell-levy-gibson-dependency-locality.md @ sha256:1c2cef7ae6abaa7a2d5f8d80db1814a62740db0e2f1d674b15a9abdbae8ce8d1 — §2.3, short-before-long preferences

While DLM predicts short-before-long orders in head-initial contexts, it predicts a long-before-short preference in head-final contexts, as shown in 8. Correspondingly, a long-before-short preference has been demonstrated experimentally in Japanese, an overwhelmingly head-final language (Yamashita & Chang 2001) --- kb/sources/.snapshots/futrell-levy-gibson-dependency-locality.md @ sha256:1c2cef7ae6abaa7a2d5f8d80db1814a62740db0e2f1d674b15a9abdbae8ce8d1 — §2.3, long-before-short preferences

Connections Found

For the occasion, the source's role is a bounded human-side basis for word-ordering guidance, not evidence about reading cost for any reader. It shows that human grammar and usage keep dependents near their heads; it only reviews the experiments that tie distance to comprehension cost. The KB has no theory note behind the AGENTS.md rule "Keep necessary conditions next to their claims", which is the closest existing writing guidance. The nearest evidence is Two rewrites exposed a syntax-or-repetition tradeoff. That note's first rewrite failed partly by moving a temporal qualifier away from the definition it limited and detaching an appositive. The paper supplies a human mechanism for why such detachment is costly. It does not show the same cost for LLM readers, and the two-rewrites failure concerns meaning errors during rewriting, not reading cost. For Human writing structures transfer to LLMs because failure modes overlap, the paper supplies the human-side rationale for one convention, keeping linked elements adjacent. That note's per-convention test can then ask whether LLMs share the failure; the paper does not answer it.

The paper also compares with the Attention Is All You Need ingest on the cost of linear distance between dependent elements. Human language reduces that cost by reordering the signal. Sequence models first reduced it with input reversal, then removed the architectural condition with attention. For a dual audience, this sharpens the open question: does a locality rule help attention-based readers, or does its warrant for them come only from preventing attachment errors? No KB experiment separates those warrants. Methodologically, the paper is evidence for An experiment identifies only the contrast it actually runs. Its §4.3 limits the conclusion to the baselines it ran and names unrun constraints (grammar complexity bounds, noise robustness, uniform information density) as open rivals. Its §5 separates grammar from usage by adding a second comparison condition rather than reinterpreting the first.

Extractable Value

  1. What the source establishes about distance and cost -- the paper's own evidence is distributional: human production keeps head–dependent distance shorter than independently motivated baselines, in grammar and in usage. The link to processing cost is reviewed, not tested. Reading-time slowdown from a lengthened subject–verb dependency (Grodner and Gibson 2005) is the cited experimental anchor. Antilocality effects in verb-final languages are a reviewed counterexample, and the authors decline to choose among memory, search-time and information-theoretic motivations. Guidance citing this paper should therefore say "human languages and writers minimize dependency distance, and controlled experiments link long dependencies to human reading slowdown", not "short dependencies are measured to be cheaper to read". High reach for human readers; none established for LLM readers. [quick-win]
  2. Operational rules the source supports for English prose -- four rules follow from the reviewed theory and fit the corpus evidence, at different strengths. (a) Keep a modifier, qualifier or argument next to the word it attaches to; this is Behaghel's "highest law" operationalized as DLM, and it is the rule the corpus results most directly support. (b) Keep the subject close to its verb; do not insert long relative clauses or parentheticals between them. The reviewed reading-time experiment varies exactly this distance. (c) After a head, order sister constituents short before long, so a long object or clause goes last. This is a reviewed derivation for head-initial languages, with English corpus support. It reverses for head-final languages, so it is English-specific guidance, not a universal. (d) When two dependents compete for the nearest slot, the more strongly linked one should be closer; this is the information-locality extension, reviewed and supported only by the authors' cited corpus work. All four are tendencies, not rules every sentence obeys. Footnote 5 notes that grammar can override small length differences. Distance is counted in words; the reviewed alternative measures (syllables, discourse referents) correlate highly. [experiment]
  3. Competing ordering pressures a writing rule must allow -- the paper names non-locality biases that also shape order: given before new, animate or definite early, and easy-first production. A DLM rule that ignores these could push writers to place new long material early against given-before-new. For KB prose, the practical conflict is a long new claim versus a short back-reference; the source gives no resolution, only the observation that the biases interact. [just-a-reference]
  4. Scale boundary for the AGENTS.md condition–claim rule -- the paper measures word-level syntactic dependencies within sentences. "Keep necessary conditions next to their claims" also applies across sentences and paragraphs, where the paper says nothing. Within a sentence, the paper supports keeping a condition clause attached to its head; beyond the sentence, the rule needs another warrant, such as attachment-error prevention during rewriting from the two-rewrites evidence note. [deep-dive]
  5. Stacked null baselines with a stated narrowest interpretation -- a reusable template for claiming an effect: build baselines from independently motivated constraints, claim only the residual beyond the strongest one, and name unrun rival constraints. It corroborates the contrast note; the KB already holds the principle. [just-a-reference]

Limitations (our opinion)

The central corpus result is well supported but bounded by what was not run. The baselines omit several independently motivated constraints that the authors name, so "DLM is the cause" remains one explanation among untested rivals. The grammar/usage split depends on linearization models that represent only order among a head's immediate dependents and only projective orders. Any ordering regularity beyond that decomposition is counted as "usage", so the usage effect could partly be unmodeled grammar. The evaluation of those models' adequacy (acceptability and same-meaning judgments) was done only for English, as reported from earlier work. The functional mechanisms (memory decay, interference, parser search) are reviewed, not tested here; the paper explicitly takes no position on which motivation is correct. Corpus regularity in production is not evidence that a reader benefits from a given reordering: speakers could minimize distance for production ease alone. The paper reviews both production and comprehension accounts but tests neither. UD annotation choices could bias toward short dependencies; the replication across alternative parses reduces but does not remove that risk. Corpus genres vary, and comparisons fix sentence length rather than control genre. The derived ordering rules (short-before-long, information locality) are reviewed from other work, not tested in this paper. Most importantly for this KB, every result concerns human language. Nothing in the paper licenses the claim that LLM readers bear the same distance cost, and the transformer architecture is a plausible reason they might not, as in the Attention Is All You Need ingest. Writing guidance citing this source for agent-facing prose must keep that transfer gap beside the claim.

Write a theory note under kb/theory/ stating the bounded word-ordering rules (attach modifiers and conditions to their heads, keep subject near verb, put long constituents last in English) with this paper as human-side evidence, the reader-cost link marked as reviewed rather than tested, and LLM-reader benefit marked as untested; adopt it into the writing instructions only through that note.

Relevant Notes: