LACE: Locality-Aware Complexity Estimation

ACL ARR 2026 January Submission100 Authors

Keywords: syntactic complexity; Dependency Locality Theory; theory-grounded evaluation; psycholinguistic metrics; structural linguistic difficulty; model robustness; diagnostic analysis; instruction tuning effects

Abstract:

Syntactic complexity contributes to language processing difficulty, yet most evaluation metrics rely on shallow proxies or conflate syn-
tactic and lexical difficulty, leaving the structural contribution unmeasured. We address this gap by introducing two locality-aware syntac-
tic complexity metrics inspired by Dependency Locality Theory (DLT): LACE-CORE, which quantifies the memory load and integration dif-
ficulty of syntactic dependencies, and LACE-FULL, which additionally captures the cost of introducing new discourse referents. We bench-
mark these metrics across three dimensions: (1) agreement with human judgments of text simplification, (2) overlap with other complexity
metrics, and (3) prediction of difficulty in downstream QA tasks. Our results show that LACE-FULL aligns more closely with human judgments of simplified text, while LACE-CORE provides the most length-independent signal of structural complexity. These findings establish LACE as a theory-grounded benchmark for syntactic complexity, with potential applications to text simplification, readability, and accessibility.

Paper Type: Long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: robustness; probing; model analysis; linguistic theories; computational psycholinguistics; reasoning; evaluation; benchmarking;

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data analysis

Languages Studied: English

Submission Number: 100