h t t p : / / t e d l a b . m i t . e d u / t e d l a b _ w e b s i t e / r e s e a r c h p a p e r s / G i b s o n _ 2 0 0 0 _ D L T . p d f

Ingest: The Dependency Locality Theory (Gibson 2000)

Type: types/ingest-report.md

Classification

This is a theoretical book chapter in psycholinguistics (chapter 5 of an edited volume, 2000). It states a formal cost model and tests it against rival resource theories with reported experiments. Most experiments are the author's own, and the primary reports appear elsewhere (Gibson 1998, Gibson and Ko 1998, Warren and Gibson 1999, Babyonyshev and Gibson 1999). The chapter therefore summarizes evidence rather than reporting full methods, so it is classified as a scientific paper with secondary-report evidence. Author: Edward Gibson (MIT), the originator of the dependency locality theory (DLT) and a central figure in sentence-processing research. He is an advocate of the theory he presents.

Summary

Gibson proposes that sentence comprehension draws on one fixed pool of resources with two kinds of cost. Integration cost arises when a new word is attached to an earlier head or antecedent. It is 1 energy unit (EU) for building a new discourse referent plus 1 EU for each new discourse referent (noun phrases, tensed verbs as events) that intervenes between the two linked elements. Given referents, such as the pronouns I and you, cost less to cross. Storage cost is 1 memory unit (MU) for each syntactic head still required to finish the sentence grammatically. Intuitive difficulty tracks the peak integration time in the parse, and the chapter states that the true cost function is nonlinear: several long integrations at one word are worse than one long integration. The theory explains why doubly center-embedded relative clauses are unprocessable while right-branching paraphrases with the same words are easy; why a pronoun as the innermost subject makes them easier (questionnaire, Warren and Gibson 1999), which incomplete-dependency counts cannot explain; why a complement clause containing a relative clause is easier than the reverse; why a prepositional phrase in a nested position adds difficulty; and why object-extracted relative clauses are slower than subject-extracted ones. The word-region reading-time correlations come from one self-paced reading study (32 participants, 16 items, about six regions per sentence; r = 0.79 and 0.77, combined p < 0.005). Japanese ratings and an ambiguity study (60 participants) support the storage component. Gibson leaves the referent-only distance metric, the storage-integration relation, and storage versus incomplete-dependency indexing explicitly unresolved.

Contribution assessment (our opinion)

The problem is central to psycholinguistics: what limited resource makes some unambiguous sentences hard. The chapter's main contribution is an explicit, quantitative conjecture. It separates integration from storage, it measures distance in new discourse referents rather than words, and it makes word-by-word predictions that can be tested and refuted. The pronoun-subject contrast is a well-chosen test against the incomplete-dependency theories, because those theories predict no difference. The chapter's own evidence is thin. The reading-time fits rest on a handful of regions with marginal per-sentence significance, the difficulty contrasts are mostly ratings of extreme constructions, and the rivals considered are older memory-count theories. Surprisal-style expectation accounts and antilocality findings are not addressed (rival status from background knowledge, corroborated by the Futrell, Levy and Gibson ingest). From background knowledge, DLT became one of the field's standard reference theories, so its likely significance is high as a framework. Its specific metric is a tentative simplification that later work, such as Temperley 2007, partly confirmed and partly failed.

Quotes

For simplicity , it is assumed that 1 EU is consumed for each new discourse referent in the intervening region . --- kb/sources/.snapshots/gibson-dependency-locality-theory.md @ sha256:63f9d57a3c6b5e12ea0ef5a30ac8d3009501edd243d89634bb86b4a498aec29b — section 5.3, integration cost distance assumption

Becausereferentsfor first- and second-person pronouns are already presentin the current discourse, integrating acrossthem consumesfewer resourcesthan integrating across new discourse referents, according to the discourse-based DLT structural integration cost hypothesis. --- kb/sources/.snapshots/gibson-dependency-locality-theory.md @ sha256:63f9d57a3c6b5e12ea0ef5a30ac8d3009501edd243d89634bb86b4a498aec29b — section 5.3, pronoun-subject example (6)

This RC, who John met, is therefore doubly nested. The resulting structure is so complex that it is unprocessablefor most people. --- kb/sources/.snapshots/gibson-dependency-locality-theory.md @ sha256:63f9d57a3c6b5e12ea0ef5a30ac8d3009501edd243d89634bb86b4a498aec29b — section 5.1, example (1c)

becausesentence(2) contains the samewords and expressesthe sameideasas (Ic), yet (2) is much easierto understand: (2) John met the senator [S' who attacked the reporter [S' who disliked the editor]].

--- kb/sources/.snapshots/gibson-dependency-locality-theory.md @ sha256:63f9d57a3c6b5e12ea0ef5a30ac8d3009501edd243d89634bb86b4a498aec29b — section 5.1, non-nested paraphrase (2)

It should be emphasizedthat aspectsof the DLT have been oversimplified up to this point . Importantly , the structural integration cost function is not actually linear, so that a single long structural integration will never be as complex as multiple shorter integrations taking place at the sameprocessingstate. --- kb/sources/.snapshots/gibson-dependency-locality-theory.md @ sha256:63f9d57a3c6b5e12ea0ef5a30ac8d3009501edd243d89634bb86b4a498aec29b — section 5.3, oversimplifications

In addition, the distancefunction assumedthus far increments its cost for new discourse referents alone. Processingother kinds of elementsalso probably causesmeasurableincrements in structural integration cost. --- kb/sources/.snapshots/gibson-dependency-locality-theory.md @ sha256:63f9d57a3c6b5e12ea0ef5a30ac8d3009501edd243d89634bb86b4a498aec29b — section 5.3, oversimplifications

Connections Found

For the occasion, this source is the comprehension-cost anchor of the dependency-locality batch: it supplies the mechanism by which distance costs a human reader, while its siblings supply production evidence and scope. It compares with Temperley 2007 on DLT's predictions tested in written corpora; Temperley confirms most of the derived writer rules but finds one failure (fronted adjuncts run short before long where DLT predicts long before short) and uses a different metric (words and summed distance rather than new referents and peak). It compares with Futrell, Levy and Gibson on cost as a cause of difficulty versus its imprint on word order across 53 languages, where antilocality and lossy-context surprisal are reviewed as rivals. It compares with Gibson et al. 2019 on scope: the same lead author later frames locality as one efficiency pressure among several.

On the KB side, DLT is evidence for the "parsing load" named without a mechanism in Two rewrites exposed a syntax-or-repetition tradeoff; that note's failed pass moved a temporal qualifier away from the definition it limited and detached an appositive, which lengthens dependencies. The evidence is bounded to human readers. It is likewise evidence for the human half of Human writing structures transfer to LLMs because failure modes overlap, whose per-convention transfer test still lacks the LLM half. It compares with Human-LLM differences are load-bearing for knowledge system design as a quantified human-reader cost model of exactly the kind that note says must be re-checked for agents. It compares with Local materialization should outperform distant natural-language declarations on whether distance between a dependent and what governs it raises the cost of correct use: DLT counts intervening new referents rather than raw words, which is a human precedent for that note's separation of distance from intervening competing content, though DLT works at sentence scale and is not LLM evidence. It compares with Soft-bound traditions as sources for context engineering strategies and Skill and Working Memory on the shape of a bounded processor's limit: DLT assumes one fixed pool with storage and distance-dependent integration, while Chase and Ericsson show capacity expanding with skill. Finally, it compares with Attention Is All You Need on the cost of linear distance: human integration cost grows with distance, whereas attention gives a constant path length between positions, so the mechanism does not carry over to transformer readers by construction.

Extractable Value

  1. What the source establishes about distance and cost -- for human readers, the cost of attaching a word to its head grows with the number of new discourse referents in between, and holding open predicted heads adds cost from the same pool. The evidence is ratings of center-embedded contrasts plus one small self-paced reading fit, in English and Japanese. Guidance should cite it as "a tested human comprehension model predicts higher cost for long and stacked dependencies", not as a measured cost function for ordinary technical prose. High reach for human readers; none established for LLM readers. [quick-win]
  2. Writer rules the mechanism directly supports -- the chapter's contrasts license a short, falsifiable rule set for human-read prose:
  3. do not nest a clause between a subject and its verb, and prefer a right-branching paraphrase (sentence 1c versus 2);
  4. move a heavy constituent out of the span between a head and its dependent, as Japanese fronting does (3c versus 4);
  5. attach modifiers where they do not interrupt a dependency (the nested prepositional phrase contrast);
  6. prefer subject-extracted to object-extracted relative clauses;
  7. avoid opening many predicted heads at once, such as long material before the main verb, because storage cost accumulates. Each rule names its DLT contrast, so it can be tested and dropped individually. [quick-win]
  8. Distance is counted in new referents, not words -- intervening given material (pronouns, already-introduced terms) costs less than new entities. For KB prose this suggests that an unavoidable interruption should reuse established terms rather than introduce new ones, and that a word-count or sentence-length check is a crude proxy. Any lint check must choose between this metric and Temperley's word-count sum, because they can rank sentences differently. [experiment]
  9. Mechanism for the condition-adjacency rule -- DLT gives the AGENTS.md rule "Keep necessary conditions next to their claims" a human-side rationale at sentence scale: a displaced qualifier is a long integration. It does not justify section-scale adjacency, which is a different mechanism. [quick-win]
  10. Peak cost, not average, drives difficulty -- the theory predicts that several long integrations landing on one word (as at the verbs of nested clauses) cause breakdown, so rewriting effort should target the sentence's worst attachment point rather than overall length. [experiment]
  11. Storage-minimization tie-break -- in footnote 9, Gibson proposes that when storage and integration costs conflict, the parser minimizes storage, because open predictions become later integration costs. This is a tentative ordering principle for writers choosing between two imperfect orders. [just-a-reference]

Limitations (our opinion)

The central claim is bounded in five ways. First, all evidence concerns human readers. Nothing here tests LLM comprehension, and transformer attention is not distance-limited by construction (as the Attention Is All You Need ingest records), so any benefit of these rules for agent readers is a separate conjecture that needs its own test, as Human-LLM differences are load-bearing requires. Second, the difficulty contrasts use extreme constructions (doubly nested relative clauses) that rarely occur in KB prose. The chapter does not establish effect sizes for the milder dependencies of ordinary technical writing, and its reading-time fit uses about six regions per sentence with marginal per-sentence significance. Third, the metric is admittedly simplified. Gibson says the cost function is nonlinear, that elements other than new referents probably add cost, and that the storage-integration relation and storage indexing are undecided. Writer rules derived from the exact EU counts would claim more precision than the source has. Fourth, the rivals tested are older incomplete-dependency theories. Expectation-based accounts, antilocality effects in verb-final languages, and semantic-similarity interference (which footnote 3 concedes affects nested structures) remain alternative explanations for some contrasts. An independent corpus test found one failed DLT prediction (Temperley 2007). Fifth, the chapter is a secondary summary; methods and full statistics are in the primary papers. The pdftotext capture is also scrambled in the ambiguity-resolution section and the conclusions, so this ingest relies on the clean prose of sections 5.1 to 5.3.5 and does not establish the details of the ambiguity-resolution claims (snapshot required).

Write a tentative theory note in kb/theory/ claiming that keeping each dependent near its head, measured in intervening new referents, lowers human parse cost and that reordering can achieve this without the repetition cost of splitting. Use this ingest as the mechanism, Temperley 2007 for the writer rules and the failed fronted-adjunct case, and Two rewrites exposed a syntax-or-repetition tradeoff as the KB-local failure. State the LLM-reader benefit as an untested conjecture with an explicit transfer test, before any rule enters the writing instructions.

Relevant Notes: