ChronoLens: Measuring Language Change Across Time, Languages, and Linguistic Levels
Gagan Bhatia, Julian Schlenker, Simone Paolo Ponzetto, Steffen Eger
cs.CL, cs.AI
2026-08-04
ChronoLens uses frozen multilingual LLMs with feature-aligned crosscoders to measure change across linguistic levels in one space. Sparse features match linguistic statistics at rho=0.72, far above dense embeddings' 0.29.
Language changes over time, and it changes on different levels: word form (morphology), sentence structure (syntax), meaning (semantics), and use (pragmatics). Past computational work studied these levels with mutually incompatible representations, so it could not answer a basic question: do these levels change in lockstep, or each on its own track? Which language changes faster, and in which direction? This paper pulls all four levels into one analytical space and gives a comparable scale.
ChronoLens is a framework that treats a large model's internal features as a measuring instrument, assembled from three parts:
The data is 44.98 million documents, about 17.2 billion tokens, of parliamentary speech records across five languages (English, German, Italian, Polish, Turkish) from 1803 to 2026.
First, is the scale any good? The sparse features from the crosscoder correlate with traditional linguistic statistics at rho=0.72, far above dense embeddings' 0.29 and a vanilla sparse autoencoder's 0.28; linguistic specificity is 0.89 versus 0.74 and 0.73. These sparse features capture something linguistically meaningful.
The core finding: within a single language, morphology, syntax, semantics, and pragmatics change by roughly comparable amounts; across languages, however, when, how far, and in which direction they change differs markedly. Similar magnitudes can hide different trajectories, so cross-linguistic comparison must measure both distance and direction.
For interpretability researchers, this is a clean example of crosscoders in action: not used to "understand one model," but to align representations across many models and time periods into a shared feature set and wield it as an instrument. For computational linguistics, it is the first time the four linguistic levels can be compared on one scale, yielding the structured conclusion of "comparable change, different trajectories." The method transfers to any sparse-feature analysis of an object evolving along some variable.
The instrument itself has error. The authors concede that frozen multilingual LLMs are "imperfect measurement instruments," and early material carries residual OCR errors that contaminate features.
The four-level division is an analytical simplification that hides cross-level features, and the study covers only parliamentary corpora in five languages. Parliamentary records represent institutional discourse, not a language in its entirety.
Feature interventions show relevance, not causation. You can see that a feature correlates with semantic change, but cannot say what causes what.