Human language is constrained by a fundamental cognitive bottleneck: our working memory. When we process a sentence, we must hold words in our mind until their syntactic partners appear. If the distance between a "head" (like a verb) and its "dependent" (like its object) is too large, the mental effort required to maintain that connection increases, leading to processing errors or slower comprehension.
To mitigate this, natural languages across the globe appear to be optimized for Dependency Length Minimization (DLM). This principle suggests that word order is not arbitrary but is shaped to keep related words as close together as possible, reducing the "dependency distance" the brain must navigate.
The Memory Bottleneck and Information Locality
The drive for efficiency stems from the sequential nature of speech and text. As Richard Futrell argues in Linguistic structure from a bottleneck on sequential information processing, the brain faces a predictive burden; it is significantly more efficient to process units that are concatenated rather than interleaved. This leads to a design principle known as information locality, where elements with high mutual information—those that help predict one another—are placed in close proximity to minimize the cost of holding predictors in memory (Richard Futrell).
Evidence for this optimization is found in the very structure of our grammars. Gildea et al. demonstrated in Optimizing Grammars for Minimum Dependency Length that English word order is significantly closer to a theoretical minimum for dependency length than to a random arrangement. This suggests that efficiency is a primary driver of how languages evolve over time. Indeed, Juzek et al. found that scientific English has become measurably more efficient over the last 300 years, shifting toward nominal structures that shorten dependency arcs (Exploring diachronic syntactic shifts with dependency length: the case of scientific English).
Asymmetry and the Short-Before-Long Preference
One of the most visible manifestations of DLM is the ordering of constituents in coordination (e.g., "cats and dogs"). Humans exhibit a consistent preference for placing shorter, simpler linguistic units before longer, more complex ones.
In A Corpus-based Syntactic Analysis of Two-termed Unlike Coordination, Kallini et al. show that when joining different syntactic categories, simpler phrases like Noun Phrases almost always precede complex subordinate clauses. This "short-before-long" pattern serves to minimize the distance to the external governor of the phrase. However, this preference is sensitive to the position of that governor. Przepiórkowski et al. note in Conjunct Lengths in English, Dependency Length Minimization, and Dependency Structure of Coordination that while the preference is strong when the governor precedes the coordination, it vanishes when the governor follows it, suggesting that word order is a dynamic balancing act rather than a rigid rule.
This optimization happens even at the most granular levels. Gulordava et al. found that DLM influences word order even within three- or four-word noun phrases, where secondary modifiers shift positions to minimize the total stretch of dependencies (Dependency length minimisation effects in short spans: a large-scale analysis of adjective placement in complex noun phrases).
Computational Efficiency and Parsing
The same principles that make language easier for humans to speak also make it easier for machines to parse. Dependency parsers—algorithms that map the relationships in a sentence—perform significantly better when dependency lengths are minimized.
Eisner et al. exploited this in Parsing with Soft and Hard Constraints on Dependency Length by introducing "vine grammars" that penalize long-distance connections, reducing parsing time from cubic to linear complexity. Similarly, Nilsson et al. showed that transforming wide "star" structures into sequential "chains" reduces the search space for parsers by shortening the average arc length (Graph Transformations in Data-Driven Dependency Parsing).
Representational choices also matter. Manning et al. investigated how reindexing functional heads (like prepositions) can create more compact structures that are easier for transition-based parsers to identify (Does Universal Dependencies need a parsing representation? An investigation of English). For head-final languages like Japanese, Kanayama et al. found that switching to right-headed models significantly improved accuracy by aligning with the language's natural branching direction and reducing dependency length (Coordinate Structures in Universal Dependencies for Head-final Languages).
The Limits of Efficiency
While DLM is a powerful force, it is not the only one. Language processing is a "preference trade-off" where syntactic constraints compete with semantic knowledge (Lenhart K. Schubert). For instance, Agarwal et al. found that in Hindi, recent discourse context and "lexical boosts" can override the preference for shorter dependencies, leading speakers to choose more complex word orders if they have been recently primed (Dual Mechanism Priming Effects in Hindi Word Order).
Ultimately, dependency length minimization serves as a universal pressure that keeps language functional. By keeping related words close, natural language ensures that the information flow remains within the bounds of human cognitive capacity, a pattern that persists from ancient scientific texts to modern large language models (Tur et al.).