# ChronoLens：跨时间、语言与语言层面测量语言变迁

- 来源：HuggingFace Daily Papers（社区热门论文）
- 发布时间：2026-08-04 08:00
- AIHOT 分数：48
- AIHOT 链接：https://aihot.virxact.com/items/cmsg5fy9y06cerolge7tiu2ts
- 原文链接：https://arxiv.org/abs/2608.03507

## AI 摘要

纽伦堡工业大学与曼海姆大学推出 ChronoLens 框架，结合冻结多语言模型、特征对齐的 crosscoders 与事后语言干预，在统一分析空间中比较五种语言、四个语言层面的历史变迁。

## 正文

Gagan Bhatia

Julian Schlenker

Simone Paolo Ponzetto

Steffen Eger

University of Technology Nuremberg

University of Mannheim

(gagan.bhatia, steffen.eger)@utn.de

Abstract

Historical language change affects morphology, syntax, semantics, and pragmatics, yet computational studies typically examine these levels with incompatible representations and therefore cannot determine whether they evolve together across languages. We address this problem by asking how the magnitude and direction of change vary across linguistic levels, languages, and historical periods within a single analytical space. We introduce ChronoLens, a framework that combines frozen multilingual language models, feature-aligned crosscoders, and post-hoc linguistic interventions, and apply it to 44.98 million documents and approximately 17.2 billion tokens from five parliamentary traditions spanning 1803–2026. The resulting sparse representations agree substantially more strongly with linguistic statistics than dense embeddings or a pooled sparse autoencoder (ρ=0.72 versus 0.29 and 0.28), and reveal that morphology, syntax, semantics, and pragmatics generally change by comparable amounts within a language, while languages differ markedly in when, how far, and in which direction they change. These findings show that historical language change is a structured, multidimensional process: similar magnitudes can conceal different trajectories, and meaningful cross-linguistic comparison requires measuring both distance and direction.

ChronoLens: Measuring Language Change Across Time, Languages, and Linguistic Levels

Gagan Bhatia1, Julian Schlenker2, Simone Paolo Ponzetto2, Steffen Eger1 1University of Technology Nuremberg 2University of Mannheim (gagan.bhatia, steffen.eger)@utn.de

1 Introduction

Language change over time affects morphological form, syntactic structure, meaning, and pragmatic function, but computational studies rarely examine these levels within the same analytical framework (Degaetano-Ortlieb and Teich, 2018; Bizzoni et al., 2019; Reinig et al., 2024). Most computational research on historical language change has focused on lexical semantics, using static embeddings, contextual representations, usage similarity, or optimal transport to compare word meanings across periods (Hamilton et al., 2016b; Giulianelli et al., 2020; Periti and Tahmasebi, 2024; Periti and Montanelli, 2024; Kishino et al., 2025). Morphological comparisons commonly use annotated form distributions, while computational studies of pragmatics often model speech acts or other communicative functions (Berdicevskis et al., 2018; Baayen, 2009; Reinig et al., 2024; Subramanian et al., 2019). Because these research areas use different representations, datasets, and units of measurement, their trajectories cannot be compared directly (Degaetano-Ortlieb and Teich, 2018; Bizzoni et al., 2019; Periti and Montanelli, 2024). This incompatibility prevents us from answering a basic question about language change: do morphology, syntax, semantics, and pragmatics change together, or do they follow distinct historical trajectories? Studies relating lexical and grammatical development suggest that changes at different levels can interact, but these analyses have largely focused on one language, register, or pair of linguistic measurements (Degaetano-Ortlieb and Teich, 2018; Bizzoni et al., 2019; Chen et al., 2026). A change observed in one language may reflect a broader development shared by several linguistic communities, or it may result from that language’s grammar, political history, or corpus composition (Hamilton et al., 2016a; Niu et al., 2023; Krielke et al., 2024). Determining which explanation is more plausible requires a representation in which linguistic levels, languages, and historical periods are directly comparable.

Multilingual language models provide part of this representation because they encode several languages within a common dense space (Schuster et al., 2019; Martins et al., 2024). Dense model dimensions, however, do not correspond to stable linguistic features because linguistic information is distributed across dimensions and individual dimensions may participate in several unrelated computations (Bricken et al., 2023; Huben et al., 2024; Templeton et al., 2024). Sparse autoencoders address this problem by decomposing dense activations into a sparse set of more selective features (Huben et al., 2024; Gao et al., 2025; Lieberum et al., 2024). Yet sparse autoencoder features are not canonical: separate training runs or datasets need not recover the same feature inventory (Leask et al., 2025; Karvonen et al., 2025). Consequently, feature j in a dictionary trained on one language or period has no guaranteed correspondence to feature j in another dictionary (Leask et al., 2025; Deng et al., 2025). Recent approaches connect either languages and features, corpora and time, or model checkpoints and linguistic capabilities. To our knowledge, no prior work places multiple human languages, historical periods, and linguistic levels in a shared, feature-aligned representation framework (Deng et al., 2025; Jing et al., 2026; Bayazit et al., 2026; Fedorova et al., 2026).

We introduce ChronoLens, a framework for comparing historical change across five languages and four linguistic levels within a common analytical space. We study English, German, Italian, Polish, and Turkish using a unified corpus of 44.98 million parliamentary documents and approximately 17.2 billion tokens collected from 22 open sources spanning 200+ years (Coole et al., 2020; Blätte and Blessing, 2018; Ogrodniczuk and Nitoń, 2020; Cova, 2025; Güngör, 2018; Erjavec et al., 2023, 2024). Parliamentary proceedings provide dated records produced under recurring institutional roles and communicative conventions, which has made them a common resource for comparative political and historical language research (Erjavec et al., 2023, 2024; de Jong et al., 2024; Skubic and Fišer, 2024). Their long, precisely dated coverage and relatively stable institutional context facilitate cross-temporal and cross-lingual comparison while limiting variation due to changing genres. Previous parliamentary studies have examined ideological change, political framing, solidarity, and speech acts, but generally within one language or with one predefined linguistic outcome (Walter et al., 2021; Kostikova et al., 2024; Ghafouri et al., 2025; Reinig et al., 2024).

Our contributions are as follows: (i) we introduce ChronoLens, a unified framework for comparing historical language change across five languages, multiple historical periods, and four linguistic levels: morphology, syntax, semantics, and pragmatics; (ii) we construct a multilingual diachronic corpus comprising 44.98 million parliamentary and political documents and approximately 17.2 billion tokens from 22 open sources, spanning the period from 1803 to 2026; (iii) we develop a feature-aligned methodology that combines frozen multilingual language models with crosscoders and post-hoc probe interventions, enabling sparse features to be compared directly across languages and periods without using linguistic labels during feature learning; and (iv) we provide an empirical analysis showing that the resulting features are more strongly aligned with observed linguistic statistics than dense embeddings or sparse autoencoders, and that historical change is coordinated across linguistic levels but differs substantially across languages in its timing, magnitude, and direction.

2 Related Work

Computational approaches to language change. Most computational work on diachronic change focuses on lexical semantics (Hamilton et al., 2016b; Eger and Mehler, 2016; Periti and Tahmasebi, 2024; Kishino et al., 2025), while syntactic change is commonly measured through dependency distance and structural complexity (Liu et al., 2022; Krielke et al., 2025; Chen et al., 2026). Work on morphology has examined productivity, morphosyntactic complexity, and the relation between morphological structure and meaning (Baayen, 2009; Berdicevskis et al., 2018; Cotterell and Schütze, 2018; Nagata et al., 2026). Moreover, shared processing pressures such as dependency-length minimization provide a reason to expect partial convergence across languages (Futrell et al., 2015; Gibson et al., 2019; Futrell et al., 2020; Hahn and Xu, 2022; Niu et al., 2023; Xu and Futrell, 2024). However, prior studies generally apply different representations to different phenomena. We instead compare morphology, syntax, semantics, and pragmatics within one shared representation space.

Parliamentary and political discourse. Parliamentary corpora have supported diachronic and comparative research on ideology, framing, migration, solidarity, and speech acts (Erjavec et al., 2023, 2024; Walter et al., 2021; Ghafouri et al., 2025; Kostikova et al., 2024; Reinig et al., 2024). This work typically targets one language, concept, or prediction task. In contrast, we use comparable parliamentary material to investigate whether several linguistic levels follow shared or language-specific historical trajectories.

Sparse and feature-aligned representations. Sparse autoencoders recover interpretable, language-selective, and culturally selective features from language-model activations (Huben et al., 2024; Deng et al., 2025; Andrylie et al., 2025; Zou et al., 2026; Jing et al., 2026). However, independently trained dictionaries need not contain aligned features (Leask et al., 2025). Crosscoders address this problem by learning a shared feature index across models or checkpoints (Lindsey et al., 2024; Jiralerspong and Bricken, 2026; Minder et al., 2026; Bayazit et al., 2026). We adapt them to languages and historical periods, assign linguistic interpretations post hoc, and distinguish the magnitude and direction of change.

媒体内容 · 前往原文查看

Language Code Docs (M) Tokens (B) Coverage

English en 17.78 6.8 1803–2026

Italian it 5.15 3.6 1848–2022

German de 4.45 3.6 1867–2026

Polish pl 15.61 2.5 1919–2025

Turkish tr 1.99 0.7 1950–2023

Total 44.98 17.2 1803–2026

Table 1: Corpus overview by language. The full corpus contains approximately 17.2B tokens.

3 Dataset

Sources and coverage. We introduce a multilingual diachronic corpus of parliamentary speech and related political text. The corpus draws on 22 open official and research sources across five languages, combining long parliamentary records with smaller complementary sources such as party manifestos. The main parliamentary sources include UK Hansard and TheyWorkForYou for English (Coole et al., 2020; Odell and Simon, 2019), the Polish Parliamentary Corpus and ParlaMint for Polish (Ogrodniczuk and Nitoń, 2020; Erjavec et al., 2024), ItaParlCorpus, IPSA, and ParlaMint for Italian (Cova, 2025; Frasnelli and Palmero Aprosio, 2024; Erjavec et al., 2024), GermaParl, German parliamentary proceedings, Reichstag material, DeuParl, ParlaMint, and official Bundestag records for German (Blätte and Blessing, 2018; Walter et al., 2021; Erjavec et al., 2024), and TBMM and ParlaMint data for Turkish (Güngör, 2018; Erjavec et al., 2024). Party manifestos are drawn from the Manifesto Project (Volkens et al., 2017). All sources are mapped to a unified record schema with shared fields for language, date, document type, source, and text. This schema is necessary for cross-lingual comparison, since a metric can only be compared across languages when the underlying records carry the same temporal and document-level metadata. The final corpus contains 44.98M documents and approximately 17.2B tokens, spanning 1803 to 2026. Table 1 summarizes the corpus by language. Appendix A lists the individual sources.

Quality control and density. Because historical OCR error can look like language change, we apply source-specific quality control before sampling. For OCR-derived Reichstag material from 1867 to 1942, we use ABBYY FineReader character confidence and remove pages below a confidence threshold or with more than 15% low-confidence glyphs, which removes about 8% of the oldest pages. Born-digital sources do not have OCR confidence scores, so we screen them with a character-n-gram gibberish detector, which removes about 0.5% of born-digital text. We treat pre-1949 data cautiously because OCR noise is concentrated there. Since diachronic analysis also requires continuous coverage, not only large total size, Figure 5 reports both the temporal span of each language and per-decade document density after OCR filtering.

4 ChronoLens

Figure 1: Overview of crosscoder training and post-hoc linguistic attribution. (a) Matched cross-lingual or diachronic sentence tuples are constructed by sampling stratum. (b) Condition-specific encoders and decoders learn shared sparse features through reconstruction alone. (c) The trained model yields aligned but linguistically unlabeled features and decoder directions. (d) Held-out feature ablations assign each feature to the linguistic level with the largest selective prediction drop.

We introduce ChronoLens, a framework for studying how multilingual representations change over time. The framework connects three dimensions of variation: language, historical period, and linguistic level. Figure 1 summarizes the ChronoLens pipeline. It measures how far each language moves through the learned feature space, whether pairs of languages move in similar directions, and whether the four linguistic levels exhibit aligned trajectories within a language. Each sentence is encoded once with a frozen multilingual language model, and the resulting representation is used throughout the pipeline. We train crosscoders under two comparison conditions: a cross-lingual setting that contrasts languages and a period-level setting that contrasts historical periods within one language. The crosscoders learn sparse features without linguistic supervision; only after training do we assign these features to morphology, syntax, semantics, or pragmatics. This separation prevents the linguistic labels from shaping the learned feature inventory.

4.1 Representation Learning and Linguistic Attribution

Input tuples. The primary analysis uses parliamentary sentences sampled independently of predefined target words. Cross-lingual comparisons cover 1950–2020, the period available for all five languages, whereas within-language analyses use the full historical record available for each language. We construct input tuples by matching sentences on token length. The matched sentences are neither translations nor paraphrases. Length matching controls for systematic differences across languages and periods that could otherwise allow the crosscoder to distinguish conditions from sentence length rather than linguistic content. In the cross-lingual setting, each tuple contains one matched sentence from each language. In the period-level setting, each tuple contains matched sentences from different historical periods of the same language (Fig. 1(a)).

Contextual representations. Using pretrained multilingual LLMs, we encode each sentence in an input tuple as a contextualized representation (Fig. 1(a)).111Refer to Appendix C.2 for more details. Our analysis uses four multilingual backbones: Qwen3-8B (Qwen Team, 2025), Llama-3.1-8B (Llama Team, 2024), Mistral-Nemo-2407 (Jiang et al., 2023; MistralAI, 2025), and EuroLLM-9B-2512 (Martins et al., 2024). We select these models for their broad language coverage and their variation in model family, tokenizer, and pretraining data. Agreement across them is therefore less likely to result from one model’s training procedure. EuroLLM provides an additional contrast because it was developed specifically for European languages.

Crosscoder training. Using the contextualized representations from each multilingual backbone, we train a crosscoder that learns a shared feature index together with a separate decoder for each language or period (Lindsey et al., 2024; Minder et al., 2026) (Fig. 1(b)). We train each crosscoder on 20k input tuples per decade and language, without linguistic labels, and reserve 10% for validation. Sampling may reuse sentences when a stratum contains too few distinct instances; we therefore report both the number of tuples and the number of distinct sentences. Appendix C.1 provides the complete sampling procedure. For an input tuple {𝐱~c}c=1C, where c indexes a condition, either a language in the cross-lingual setting or a historical period in the period-level setting, the crosscoder computes

𝐟 =BatchTopK⁡(ReLU⁡(∑c=1C𝐱~c​𝐖encc+𝐛enc)), (1)

𝐱^c =𝐟𝐖decc+𝐛decc. (2)

The encoder projects each condition-specific representation into a shared feature space and sums the projected representations across the tuple. After adding the shared bias, ReLU removes negative activations and BatchTopK retains only the strongest positive activations. The resulting vector 𝐟 is thus one sparse representation whose feature indices are shared across all conditions (Fig. 1(c)). Each condition-specific decoder maps this vector back into the representation space of condition c. The shared feature vector aligns feature identity across conditions, whereas the separate decoders allow the same feature to contribute differently to each language or period. We set the dictionary size, i.e., the number of learned sparse features, to twice the backbone hidden dimension and use a BatchTopK target active fraction of 0.10. Appendix C.3 reports the training objective, sparsity settings, and reconstruction checks.

Linguistic level assignment. We assign crosscoder features to linguistic levels using a suite of 23 complementary sentence-level tasks. The five morphology tasks characterize the main predicate through tense, mood, voice, inflectional load, and deverbal nominalization density, drawing on prior work on multilingual morphological complexity, compositional morphology, and grammaticalization (Berdicevskis et al., 2018; Cotterell and Schütze, 2018; Nagata et al., 2026). The six syntax tasks measure clause embedding, length-adjusted tree depth, dependency distance, head direction, nominal modification, and coordination, all established properties of dependency structure and diachronic syntactic variation (Liu et al., 2022; Krielke et al., 2025; Chen et al., 2026). To avoid equating sentence meaning with policy topic, the six semantic tasks characterize predicate and argument types, abstractness, negation, quantification and policy-frame task (Otmakhova and Frermann, 2025). The six pragmatics tasks cover deixis, stance, modality, communicative act, evidentiality, and politeness, following work that treats political language use as more than sentence form alone (Reinig et al., 2024; Subramanian et al., 2019). Labels are derived primarily from Universal Dependencies parses and multilingual lexical resources (Nivre et al., 2020; Qi et al., 2020); continuous measures are discretized into low, medium, and high classes. Semantic and pragmatic labels use deterministic weak-supervision rules that combine parse-derived morphosyntactic and dependency cues with curated multilingual lexicons for predicate and argument types, quantification, modality, stance, evidentiality, and forms of address.

After crosscoder training, we fit a linear probe for every language-task pair and compare it with an otherwise identical permuted-label control (Hewitt and Liang, 2019). Because probing is performed only after feature learning, these labels interpret the learned representation without influencing the feature inventory. For each held-out sentence, we ablate one crosscoder feature at a time and recompute the correct-label probability under every task probe (Fig. 1(d)). We aggregate the resulting probability decreases over the tasks belonging to each linguistic level. A feature is assigned to the level with the largest aggregate decrease only when that effect satisfies the attribution threshold and clearly exceeds its effects on the other levels; otherwise, the feature remains unassigned. Because these levels are not mutually exclusive, the assignment denotes a feature’s dominant selective effect rather than an exclusive linguistic interpretation and prevents double counting across level-specific trajectories. This procedure provides a post-hoc linguistic attribution of the shared feature dictionary. Appendix C.4 reports the thresholds and random-direction controls.

4.2 Evaluation and Diachronic Measurements

Baselines and measurement validation. To assess the quality of the learned crosscoder features, we compare them with two controlled baselines. The embedding Hamilton et al. (2016b); Hagen (2025) baseline uses the frozen sentence representations directly; for feature-level evaluation, we use their first 512 principal components. The pooled SAE Andrylie et al. (2025); Karvonen et al. (2025) learns a shared sparse dictionary from representations pooled across all five languages but uses one decoder for each language. In contrast, the crosscoder jointly encodes condition-specific inputs into a shared feature vector and reconstructs them with condition-specific decoders.

We evaluate four properties commonly used to assess sparse representations (Gao et al., 2025; Balagansky et al., 2025; Karvonen et al., 2025; Kantamneni et al., 2025): (1) Reconstruction is the fraction of held-out activation variance left unexplained by the reconstruction, with lower Fraction of Variance Unexplained (FVU) indicating greater fidelity; it is undefined for uncompressed embeddings. (2) Trajectory stability is the mean cosine between the full-data displacement vector and vectors obtained by resampling sentences within each language–decade cell, so higher values indicate that the estimated direction does not depend strongly on the sampled sentences. (3) Linguistic agreement is the mean Spearman correlation between decade-to-decade representational displacement and direct changes in independently measured linguistic indicators for morphology, syntax, semantics, and pragmatics (Periti and Tahmasebi, 2024; Chen et al., 2026; Otmakhova and Frermann, 2025; Reinig et al., 2024). (4) Linguistic specificity is the proportion of a feature’s total level-aggregated ablation effect concentrated on its most affected linguistic level; 0.25 corresponds to equal effects across the four levels, whereas larger values indicate more level-selective features. We evaluate these effects against permuted-label probe controls following Hewitt and Liang (2019).

Measuring historical change. We compute a separate trajectory for each language and linguistic level. Let Dℓ,t be the set of sentences in language ℓ and decade t, 𝐟s the sparse feature vector of sentence s, and Sλ the features assigned to linguistic level λ. The representation of language ℓ at level λ in decade t is the mean activation of those features:

𝐮λ​(ℓ,t)=1|Dℓ,t|​∑s∈Dℓ,t𝐟s​[Sλ]. (3)

Here, 𝐟s​[Sλ] denotes the entries of 𝐟s belonging to level λ. We compute all quantities separately for each multilingual backbone and omit the model index for readability. For each language, we measure change relative to its first available decade t0ℓ:

Magnitudeλ⁡(ℓ,t)=‖𝐮λ​(ℓ,t)−𝐮λ​(ℓ,t0ℓ)‖2‖𝐮λ​(ℓ,t0ℓ)‖2. (4)

The numerator is the Euclidean distance from the language’s initial representation, while the denominator normalizes for differences in feature scale across linguistic levels. A value of 0 denotes no change from the initial decade, and larger values denote greater displacement. To compare directions, let t0ℓ,ℓ′ and t1ℓ,ℓ′ be the first and last decades available for both languages ℓ and ℓ′. Their displacement vectors are

𝚫λℓ=𝐮λ​(ℓ,t1ℓ,ℓ′)−𝐮λ​(ℓ,t0ℓ,ℓ′), (5)

with 𝚫λℓ′ defined analogously. We measure directional alignment using cosine similarity:

Directionλ⁡(ℓ,ℓ′)=𝚫λℓ⋅𝚫λℓ′‖𝚫λℓ‖2​‖𝚫λℓ′‖2. (6)

Values near 1 indicate parallel change, values near −1 indicate change in opposite directions, and values near 0 indicate unrelated directions. Magnitude therefore measures how far each language moves, whereas direction measures whether two languages move similarly.

5 Results

All results use the same matched sentence samples, the same four backbones, and the same crosscoder configuration described in §4; every reported value is the mean over the four measuring models, with their spread reported alongside. Magnitudes follow Eq. 4 and directions follow Eq. 6. We first test whether the crosscoder provides a more linguistically valid representation than dense embeddings and a pooled sparse autoencoder (§5.1). We then examine the magnitude and temporal profile of change in each language (§5.2). Finally, we test whether similar magnitudes imply similar directions, both across languages and across linguistic levels (§5.3).

5.1 Crosscoders recover more linguistically grounded features

媒体内容 · 前往原文查看

Criterion Embeddings SAE Crosscoder

Held-out reconstruction, FVU ↓ n/a 0.07 0.04

Trajectory stability ↑ 0.90 0.91 0.92

Linguistic agreement, ρ↑ 0.29 0.28 0.72

Linguistic specificity, ↑ 0.74 0.73 0.89

Table 2: Controlled comparison of dense embeddings, a SAE, and the crosscoder. FVU is not applicable to embeddings because they do not reconstruct a compressed representation.

Table 2 shows that the crosscoder’s main advantage is linguistic rather than geometric. Agreement with direct changes in tense, subordination, policy-topic, and speech-act distributions increases to ρ=0.72, compared with 0.29 for embeddings and 0.28 for the pooled SAE. Linguistic specificity similarly increases to 0.89, from 0.74 and 0.73. In contrast, the pooled SAE does not improve over the original embeddings on either measure. A shared sparse dictionary is therefore not sufficient by itself. The improvement appears when shared feature identities are combined with condition-specific decoders. The comparison is less differentiated on reconstruction and stability. The crosscoder reduces FVU from 0.07 to 0.04, but all three representations produce highly stable directions, with bootstrap cosines between 0.90 and 0.92. Thus, the crosscoder does not obtain its linguistic advantage by producing substantially smoother trajectories. It preserves the stable signal already present in the representations while organizing it into features that better correspond to linguistic variation. Table 3 examines four representative features from the English analysis, showing how the learned inventory distinguishes historical periods, linguistic levels, and recurring lexical or grammatical patterns.

媒体内容 · 前往原文查看

Feat. Level Peak Top-activating excerpt

2373 Morph. P1 He thought it unwise to leave them …permanently alienated.

5236 Prag. P2 asked the Minister whether it is the intention of the Government …?

7333 Sem. P4 those whose asylum cases are outstanding are detained …

1082 Prag. P4 Does he not agree that immigration creates problems?

Table 3: Representative period-specific features from the English crosscoder. Peak reports the dominant period: P1 = 1803–1899, P2 = 1900–1945, P3 = 1946–1979, P4 = 1980–2004, and P5 = 2005–2026. Linguistic levels are assigned by held-out feature ablation rather than inspection of the examples. Shading shows activation intensity within each excerpt.

Table 3 shows that period specificity is not equivalent to topic specificity. Feature 5236 captures the P2 written-question construction asked the Minister whether; its strongest activations span different policy content, indicating a pragmatic parliamentary frame rather than a migration subtopic. Feature 1082 captures a later pragmatic form, Does he not agree that, associated with adversarial oral questioning in P4.

5.2 Historical change is comparable in magnitude across levels but differs across languages

Figure 2: Magnitude of historical change over each language’s available record. At decade t, the trajectory reports Magnitudeλ⁡(ℓ,t) relative to the language’s first available decade, averaged over linguistic levels. Shaded intervals show variation across the four measuring models. Because the first available decade differs by language, the figure compares complete within-language trajectories rather than a common historical interval.

Figure 3: Magnitude of historical change between 1950 and 2020 by language and linguistic level. All languages are evaluated over the same historical interval. Darker cells indicate greater magnitude of historical change from the 1950 representation.

Figure 2 reports the magnitude of historical change relative to each language’s first available decade. The languages differ in both their final magnitude and their temporal profile. German has the largest final magnitude indicating largest change, reaching 0.89 after a marked increase from approximately 1980 onward. English increases more gradually, reaches its highest magnitude in the 1990s, and ends at 0.62. Polish reaches approximately 0.53 around 1980 but decreases to 0.45 by the final decade. Italian and Turkish both end at 0.53, although their trajectories cover different periods and develop differently over time. Italian and Turkish have the same final value but different temporal profiles, while the final values for English and Polish are lower than their earlier maxima. Comparisons across languages must therefore consider both the magnitude at a given decade and the trajectory through which that magnitude develops.

Figure 3 compares the magnitude of change over the common 1950–2020 interval. Averaged across linguistic levels, Turkish has the highest magnitude at 0.61, followed by German at 0.43. English and Italian both average 0.35, while Polish averages 0.34. This ranking differs from the full-record comparison in Figure 2: German has the highest magnitude over its complete record, whereas Turkish has the highest magnitude within the common 1950–2020 period. Magnitude comparisons therefore depend on the historical interval used. Variation across languages is larger than variation across linguistic levels. The Turkish syntax and semantics cells also show the largest between-model spreads. This variation may partly reflect model-specific tokenization of Turkish: its agglutinative morphology can produce different subword segmentations across tokenizers, and such differences can affect morphology-sensitive evaluations (Arnett and Bergen, 2025; Başar and Bisazza, 2026). An analysis of individual linguistic measures provides a more direct interpretation of the aggregate magnitudes (Appendix D). We examine 23 measures grouped under morphology, syntax, semantics, and pragmatics. Across five languages, 37 of the 90 fitted trends remain significant. Personal deixis increases in all five languages, significantly in German, Italian, Polish, and Turkish. Passive voice decreases in English, German, and Turkish but increases in Polish. Long dependencies decrease in English, German, and Italian but increase in Polish. ross-language agreement is strongest for pragmatics, with a mean pairwise cosine similarity of 0.92, while syntax has no common overall direction, with a mean of 0.00. These results show that similar aggregate magnitudes can result from different changes in the underlying linguistic measures.

5.3 Magnitude and direction reveal distinct patterns

Figure 4: Magnitude and direction of historical change over each language’s available record. The length of each arrow represents the net magnitude of historical change, computed from the first to the last available decade and averaged over linguistic levels. The relative angles summarize the direction metric, that is, the cosine similarity between the corresponding displacement vectors. Languages with smaller angular separation have more similar directions of change.

Figure 4 shows that magnitude and direction capture different properties of historical change. Italian and Turkish have the same net magnitude of change, 0.53, but different directions. German and Turkish show the opposite pattern: their net magnitudes differ, 0.89 and 0.53, but their directions are similar. The magnitude of change therefore does not determine the direction of change. The directional configuration is not explained by genealogical relatedness alone. English and Italian have similar directions, as do German and Turkish, whereas English and German do not form the closest pair despite both being Germanic languages. This pattern is consistent with the possibility that parliamentary language responds to shared cultural, political, and institutional developments. Prior work has distinguished culturally associated semantic change from language-internal drift and has shown that cultural differences can be recovered from patterns of language use (Hamilton et al., 2016a; Garimella et al., 2016). It also complements studies that identify language-specific periods of semantic and syntactic change (Hamilton et al., 2016b; Periti and Tahmasebi, 2024; Degaetano-Ortlieb and Teich, 2018; Krielke et al., 2025; Chen et al., 2026), as well as work on historical variation in political framing and parliamentary speech acts (Otmakhova and Frermann, 2025; Reinig et al., 2024). Our results extend these findings by showing that (i) morphology, syntax, semantics, and pragmatics generally have comparable magnitudes of change within a language, and (ii) similar magnitudes do not imply similar directions. Single-level analyses and magnitude alone therefore capture only part of the observed diachronic structure.

6 Conclusion

We introduced ChronoLens, a framework for measuring historical language change jointly across languages, periods, and linguistic levels. The crosscoder produces representations that agree more strongly with direct linguistic stats while preserving stable historical trajectories. Across five parliamentary traditions, we find that morphology, syntax, semantics, and pragmatics generally change by comparable amounts within a language, but that languages differ substantially in the magnitude, and direction of this change. Future work can test whether these patterns generalize beyond parliamentary discourse and to a broader range of languages, linguistic measurements, and historical corpora.

Limitations

Our division into morphology, syntax, semantics, and pragmatics is also an analytical simplification. These levels are not mutually exclusive, and individual phenomena or learned features may span several of them. The hard assignment used in the trajectory analysis identifies a feature’s dominant probe effect and prevents double counting, but it can obscure genuinely cross-level features. The resulting trajectories should therefore be interpreted as changes along four operational dimensions rather than as a complete decomposition of linguistic change. The analysis is restricted to parliamentary discourse in five languages with unequal historical coverage; residual OCR errors, particularly in the earliest material, may still resemble linguistic change despite our filtering. Moreover, frozen multilingual language models remain imperfect measurement instruments whose tokenization, pretraining data, and language coverage may affect the recovered trajectories. Finally, feature interventions establish relevance to probe predictions, but they do not identify the political, cultural, or institutional causes of the observed changes.

Broader Impact

ChronoLens provides a common framework for comparing historical change across languages and linguistic levels, which may support research in computational linguistics, political science, history, and the digital humanities. At the same time, parliamentary records represent institutional discourse produced by political actors rather than the language use of entire populations. Cross-linguistic similarities should therefore not be interpreted as essential properties of national communities or as direct evidence that languages are becoming uniformly more alike. The framework is best used to generate hypotheses that are subsequently evaluated against original texts, dated historical events, and social or institutional evidence. Extending the analysis to additional genres, regions, and less-resourced languages will be important for preventing conclusions about historical language change from being dominated by well-documented European parliamentary traditions.

Ethical Considerations

Our analysis uses publicly available parliamentary and political texts from official and research sources and reports aggregate language–period patterns rather than predictions about individual speakers. Nevertheless, parliamentary records may contain identifiable speakers and discussions of sensitive political or social issues. Any release of derived data should preserve source attribution, licensing conditions, and applicable restrictions rather than redistributing source material indiscriminately. The automatic parsers, multilingual lexicons, probes, and pretrained language models used in the pipeline may also encode cultural and language-specific biases; their outputs should be treated as operational measurements, not objective labels or diagnoses of linguistic communities. The resulting representations should not be used for individual political profiling, targeted persuasion, or ranking languages and populations. Reproducibility materials should document data provenance, filtering, sampling, model versions, and known measurement limitations. Finally, we only used LLMs for code generation.

References

L. M. Andrylie, I. Rahmanisa, M. K. Ihsani, A. F. Wicaksono, H. A. Wibowo, and A. F. Aji (2025) Sparse autoencoders can capture language-specific concepts across diverse languages. arXiv preprint arXiv:2507.11230. Cited by: §2, §4.2.

C. Arnett and B. Bergen (2025) Why do language models perform worse for morphologically complex languages?. In Proceedings of the 31st International Conference on Computational Linguistics, O. Rambow, L. Wanner, M. Apidianaki, H. Al-Khalifa, B. D. Eugenio, and S. Schockaert (Eds.), Abu Dhabi, UAE, pp. 6607–6623. External Links: Link Cited by: §5.2.

R. H. Baayen (2009) Corpus linguistics in morphology: morphological productivity. In Corpus Linguistics: An International Handbook, A. Lüdeling and M. Kytö (Eds.), pp. 900–919. Cited by: §1, §2.

N. Balagansky, Y. Aksenov, D. Laptev, V. Kurochkin, G. Gerasimov, N. Koriagin, and D. Gavrilov (2025) Train one sparse autoencoder across multiple sparsity budgets to preserve interpretability and accuracy. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V. Peng (Eds.), Suzhou, China, pp. 10171–10179. External Links: Link, Document, ISBN 979-8-89176-332-6 Cited by: §4.2.

E. Başar and A. Bisazza (2026) A morphology-aware evaluation of Turkish syntax in large language models. In Proceedings of the Second Workshop Natural Language Processing for Turkic Languages (SIGTURK 2026), K. Oflazer, A. Köksal, and O. Varol (Eds.), Rabat, Morocco, pp. 95–102. External Links: Link, Document, ISBN 979-8-89176-370-8 Cited by: §5.2.

D. Bayazit, A. Mueller, and A. Bosselut (2026) Crosscoding through time: tracking emergence & consolidation of linguistic representations throughout LLM pretraining. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, pp. 1353–1377. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: §1, §2.

A. Berdicevskis, Ç. Çöltekin, K. Ehret, K. von Prince, D. Ross, B. Thompson, C. Yan, V. Demberg, G. Lupyan, T. Rama, and C. Bentz (2018) Using Universal Dependencies in cross-linguistic complexity research. In Proceedings of the Second Workshop on Universal Dependencies (UDW 2018), Brussels, Belgium, pp. 8–17. External Links: Link, Document Cited by: §1, §2, §4.1.

Y. Bizzoni, S. Degaetano-Ortlieb, K. Menzel, P. Krielke, and E. Teich (2019) Grammar and meaning: analysing the topology of diachronic word embeddings. In Proceedings of the 1st International Workshop on Computational Approaches to Historical Language Change, N. Tahmasebi, L. Borin, A. Jatowt, and Y. Xu (Eds.), Florence, Italy, pp. 175–185. External Links: Link, Document Cited by: §1.

A. Blätte and A. Blessing (2018) The GermaParl corpus of parliamentary protocols. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), N. Calzolari, K. Choukri, C. Cieri, T. Declerck, S. Goggi, K. Hasida, H. Isahara, B. Maegaard, J. Mariani, H. Mazo, A. Moreno, J. Odijk, S. Piperidis, and T. Tokunaga (Eds.), Miyazaki, Japan. External Links: Link Cited by: §1, §3.

T. Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, et al. (2023) Towards monosemanticity: decomposing language models with dictionary learning. Transformer Circuits Thread. External Links: Link Cited by: §1.

Y. Chen, W. Zhao, A. Breitbarth, M. Stoeckel, A. Mehler, D. Schlechtweg, and S. Eger (2026) Syntactic language change in english and german: metrics, parsers, and convergences. PLOS One 21 (4), pp. e0346096. External Links: Link, Document Cited by: §1, §2, §4.1, §4.2, §5.3.

M. Coole, P. Rayson, and J. Mariani (2020) Unfinished business: construction and maintenance of a semantically tagged historical parliamentary corpus, UK Hansard from 1803 to the present day. In Proceedings of the Second ParlaCLARIN Workshop, D. Fišer, M. Eskevich, and F. de Jong (Eds.), Marseille, France, pp. 23–27 (eng). External Links: Link, ISBN 979-10-95546-47-4 Cited by: §1, §3.

R. Cotterell and H. Schütze (2018) Joint semantic synthesis and morphological analysis of the derived word. Transactions of the Association for Computational Linguistics 6, pp. 33–48. External Links: Link, Document Cited by: §2, §4.1.

J. Cova (2025) A new database for italian parliamentary speeches: introducing the itaparlcorpus dataset. Italian Political Science Review/Rivista Italiana di Scienza Politica 55 (1), pp. 77–86. External Links: Document Cited by: §1, §3.

A. de Jong, T. Kuzman, M. Larooij, and M. Marx (2024) ParlaMint ngram viewer: multilingual comparative diachronic search across 26 parliaments. In Proceedings of the IV Workshop on Creating, Analysing, and Increasing Accessibility of Parliamentary Corpora (ParlaCLARIN) @ LREC-COLING 2024, Torino, Italia, pp. 110–115. External Links: Link Cited by: §1.

S. Degaetano-Ortlieb and E. Teich (2018) Using relative entropy for detection and analysis of periods of diachronic linguistic change. In Proceedings of the Second Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, B. Alex, S. Degaetano-Ortlieb, A. Feldman, A. Kazantseva, N. Reiter, and S. Szpakowicz (Eds.), Santa Fe, New Mexico, pp. 22–33. External Links: Link Cited by: §1, §5.3.

B. Deng, Y. Wan, B. Yang, Y. Zhang, and F. Feng (2025) Unveiling language-specific features in large language models via sparse autoencoders. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, pp. 4563–4608. External Links: Link, Document, ISBN 979-8-89176-251-0 Cited by: §1, §2.

S. Eger and A. Mehler (2016) On the linearity of semantic change: investigating meaning variation via dynamic graph models. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), K. Erk and N. A. Smith (Eds.), Berlin, Germany, pp. 52–58. External Links: Link, Document Cited by: §2.

T. Erjavec, M. Kopp, N. Ljubešić, T. Kuzman, P. Rayson, P. Osenova, M. Ogrodniczuk, u. undefinedöltekin, D. Koržinek, K. Meden, J. Skubic, P. Rupnik, T. Agnoloni, J. Aires, S. Barkarson, R. Bartolini, N. Bel, M. Calzada Pérez, R. Darģis, S. Diwersy, M. Gavriilidou, R. van Heusden, M. Iruskieta, N. Kahusk, A. Kryvenko, N. Ligeti-Nagy, C. Magariños, M. Mölder, C. Navarretta, K. Simov, L. M. Tungland, J. Tuominen, J. Vidler, A. I. Vladu, T. Wissik, V. Yrjänäinen, and D. Fišer (2024) ParlaMint ii: advancing comparable parliamentary corpora across europe. Language Resources and Evaluation 59 (3), pp. 2071–2102. External Links: ISSN 1574-0218, Link, Document Cited by: §1, §2, §3.

T. Erjavec, M. Ogrodniczuk, P. Osenova, N. Ljubešić, K. Simov, A. Pančur, M. Rudolf, M. Kopp, et al. (2023) The ParlaMint corpora of parliamentary proceedings. Language Resources and Evaluation 57 (1), pp. 415–448. External Links: Document, Link Cited by: §1, §2.

M. Fedorova, A. Kutuzov, and K. Umarova (2026) DHPLT: large-scale multilingual diachronic corpora and word representations for semantic change modelling. In The Proceedings for the 6th International Workshop on Computational Approaches to Language Change (LChange’26), N. Tahmasebi, P. Cassotti, S. Montariol, A. Kutuzov, N. Huebscher, E. Spaziani, and N. Baes (Eds.), Rabat, Morocco, pp. 87–96. External Links: Link, Document, ISBN 979-8-89176-362-3 Cited by: §1.

V. Frasnelli and A. Palmero Aprosio (2024) There’s something new about the Italian parliament: the IPSA corpus. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), N. Calzolari, M. Kan, V. Hoste, A. Lenci, S. Sakti, and N. Xue (Eds.), Torino, Italia, pp. 16037–16046. External Links: Link Cited by: §3.

R. Futrell, R. P. Levy, and E. Gibson (2020) Dependency locality as an explanatory principle for word order. Language 96 (2), pp. 371–412. External Links: Document Cited by: §2.

R. Futrell, K. Mahowald, and E. Gibson (2015) Large-scale evidence of dependency length minimization in 37 languages. Proceedings of the National Academy of Sciences 112 (33), pp. 10336–10341. External Links: Document Cited by: §2.

L. Gao, T. D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu (2025) Scaling and evaluating sparse autoencoders. In The Thirteenth International Conference on Learning Representations (ICLR), External Links: Link Cited by: §1, §4.2.

A. Garimella, R. Mihalcea, and J. Pennebaker (2016) Identifying cross-cultural differences in word usage. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, Y. Matsumoto and R. Prasad (Eds.), Osaka, Japan, pp. 674–683. External Links: Link Cited by: §5.3.

V. Ghafouri, R. McNeil, T. Yankov, M. Sumption, L. Rocher, S. A. Hale, and A. Mahdi (2025) Framing migration: a computational analysis of UK parliamentary discourse. External Links: 2509.14197, Link, Document Cited by: §1, §2.

E. Gibson, R. Futrell, S. T. Piantadosi, I. Dautriche, K. Mahowald, L. Bergen, and R. Levy (2019) How efficiency shapes human language. Trends in Cognitive Sciences 23 (5), pp. 389–407. External Links: Document Cited by: §2.

M. Giulianelli, M. Del Tredici, and R. Fernández (2020) Analysing lexical semantic change with contextualised word representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Online, pp. 3960–3973. External Links: Link, Document Cited by: §1.

O. Güngör (2018) A corpus of grand national assembly of turkish parliament’s transcripts. External Links: Link Cited by: §1, §3.

T. Hagen (2025) Lexical semantic change annotation with large language models. In Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025), Albuquerque, New Mexico, pp. 172–178. External Links: Link, Document Cited by: §4.2.

M. Hahn and Y. Xu (2022) Crosslinguistic word order variation reflects evolutionary pressures of dependency and information locality. Proceedings of the National Academy of Sciences 119 (24), pp. e2122604119. External Links: Document Cited by: §2.

W. L. Hamilton, J. Leskovec, and D. Jurafsky (2016a) Cultural shift or linguistic drift? comparing two computational measures of semantic change. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, J. Su, K. Duh, and X. Carreras (Eds.), Austin, Texas, pp. 2116–2121. External Links: Link, Document Cited by: §1, §5.3.

W. L. Hamilton, J. Leskovec, and D. Jurafsky (2016b) Diachronic word embeddings reveal statistical laws of semantic change. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1489–1501. External Links: Document, Link Cited by: §1, §2, §4.2, §5.3.

J. Hewitt and P. Liang (2019) Designing and interpreting probes with control tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, pp. 2733–2743. External Links: Link, Document Cited by: §C.5, Table 9, §4.1, §4.2.

R. Huben, H. Cunningham, L. R. Smith, A. Ewart, and L. Sharkey (2024) Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations (ICLR), External Links: Link Cited by: §1, §2.

A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de Las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, and W. E. Sayed (2023) Mistral 7b. CoRR abs/2310.06825. External Links: Document, Link Cited by: Table 7, §4.1.

Y. Jing, W. Qiu, Y. Peng, and Z. Sui (2026) HistLens: mapping idea change across concepts and corpora. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, pp. 14326–14351. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: §1, §2.

T. Jiralerspong and T. Bricken (2026) Cross-architecture model diffing with crosscoders: unsupervised discovery of differences between llms. External Links: 2602.11729, Link Cited by: §C.3, §2.

S. Kantamneni, J. Engels, S. Rajamanoharan, M. Tegmark, and N. Nanda (2025) Are sparse autoencoders useful? A case study in sparse probing. In Proceedings of the 42nd International Conference on Machine Learning (ICML), External Links: Link Cited by: §4.2.

A. Karvonen, C. Rager, J. Lin, C. Tigges, J. Bloom, D. Chanin, Y. Lau, E. Farrell, C. McDougall, K. Ayonrinde, M. Wearden, A. Conmy, S. Marks, and N. Nanda (2025) SAEBench: a comprehensive benchmark for sparse autoencoders in language model interpretability. External Links: 2503.09532, Link Cited by: §1, §4.2, §4.2.

R. Kishino, H. Yamagiwa, R. Nagata, S. Yokoi, and H. Shimodaira (2025) Quantifying lexical semantic shift via unbalanced optimal transport. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vienna, Austria, pp. 15913–15933. External Links: Link, Document Cited by: §1, §2.

A. Kostikova, B. Paassen, D. Beese, O. Pütz, G. Wiedemann, and S. Eger (2024) Fine-grained detection of solidarity for women and migrants in 155 years of German parliamentary debates. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Miami, Florida, USA, pp. 5884–5907. External Links: Link, Document Cited by: §1, §2.

M. Krielke, D. Alves, and L. Talamo (2025) Tracing syntactic complexity: exploring the evolution of average dependency length across three centuries of scientific English. In Proceedings of the Eighth International Conference on Dependency Linguistics (Depling, SyntaxFest 2025), Ljubljana, Slovenia, pp. 13–23. External Links: Link Cited by: §2, §4.1, §5.3.

M. Krielke, S. Degaetano-Ortlieb, and E. Teich (2024) Cross-linguistic dependency length minimization in scientific language. Languages in Contrast 24 (1). External Links: Link, Document Cited by: §1.

P. Leask, B. Bussmann, M. T. Pearce, J. I. Bloom, C. Tigges, N. A. Moubayed, L. Sharkey, and N. Nanda (2025) Sparse autoencoders do not find canonical units of analysis. In The Thirteenth International Conference on Learning Representations (ICLR), Cited by: §1, §2.

T. Lieberum, S. Rajamanoharan, A. Conmy, L. Smith, N. Sonnerat, V. Varma, J. Kramár, A. Dragan, R. Shah, and N. Nanda (2024) Gemma scope: open sparse autoencoders everywhere all at once on Gemma 2. External Links: 2408.05147, Link Cited by: §1.

J. Lindsey, A. Templeton, J. Marcus, T. Conerly, J. Batson, and C. Olah (2024) Sparse crosscoders for cross-layer features and model diffing. Transformer Circuits Thread. Note: Research update External Links: Link Cited by: §2, §4.1.

X. Liu, H. Zhu, and L. Lei (2022) Dependency distance minimization: a diachronic exploration of the effects of sentence length and dependency types. Humanities and Social Sciences Communications 9 (1). External Links: Link, Document Cited by: §2, §4.1.

Llama Team (2024) The Llama 3 herd of models. CoRR abs/2407.21783. External Links: Document, Link Cited by: Table 7, §4.1.

P. H. Martins, P. Fernandes, J. Alves, N. M. Guerreiro, R. Rei, D. M. Alves, J. Pombal, M. A. Farajian, M. Faysse, M. Klimaszewski, P. Colombo, B. Haddow, J. G. C. de Souza, A. Birch, and A. F. T. Martins (2024) EuroLLM: multilingual language models for europe. CoRR abs/2409.16235. External Links: Document, Link Cited by: Table 7, §1, §4.1.

J. Minder, C. Dumas, C. Juang, B. Chugtai, and N. Nanda (2026) Overcoming sparsity artifacts in crosscoders to interpret chat-tuning. External Links: 2504.02922, Link Cited by: §C.3, §C.3, §C.3, §2, §4.1.

MistralAI (2025) External Links: Link Cited by: Table 7, §4.1.

R. Nagata, D. Mochihashi, M. Ido, Y. Kubota, N. Otani, Y. Kawasaki, and H. Takamura (2026) Cross-lingual and word-independent methods for quantifying degree of grammaticalization. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), V. Demberg, K. Inui, and L. Marquez (Eds.), Rabat, Morocco, pp. 4775–4787. External Links: Link, Document, ISBN 979-8-89176-380-7 Cited by: §2, §4.1.

R. Niu, Y. Wang, and H. Liu (2023) The cross-linguistic variations in dependency distance minimization and its potential explanations. In Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation, Hong Kong, China, pp. 559–569. External Links: Link Cited by: §1, §2.

J. Nivre, M. de Marneffe, F. Ginter, J. Hajič, C. D. Manning, S. Pyysalo, S. Schuster, F. M. Tyers, and D. Zeman (2020) Universal dependencies v2: an evergrowing multilingual treebank collection. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), pp. 4034–4043. External Links: Link Cited by: §4.1.

E. Odell and B. Simon (2019) Evanodell/hansard: hansard 0.8.0. Zenodo. External Links: Document, Link Cited by: §3.

M. Ogrodniczuk and B. Nitoń (2020) New developments in the Polish parliamentary corpus. In Proceedings of the Second ParlaCLARIN Workshop, D. Fišer, M. Eskevich, and F. de Jong (Eds.), Marseille, France, pp. 1–4 (eng). External Links: Link, ISBN 979-10-95546-47-4 Cited by: §1, §3.

Y. Otmakhova and L. Frermann (2025) Narrative media framing in political discourse. In Findings of the Association for Computational Linguistics: ACL 2025, Vienna, Austria, pp. 9167–9196. External Links: Link, Document Cited by: §4.1, §4.2, §5.3.

F. Periti and S. Montanelli (2024) Lexical semantic change through large language models: a survey. ACM Computing Surveys 56 (11), pp. 282:1–282:38. Cited by: §1.

F. Periti and N. Tahmasebi (2024) A systematic comparison of contextualized word embeddings for lexical semantic change. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Mexico City, Mexico, pp. 4262–4282. External Links: Link, Document Cited by: §1, §2, §4.2, §5.3.

P. Qi, Y. Zhang, Y. Zhang, J. Bolton, and C. D. Manning (2020) Stanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pp. 101–108. External Links: Document, Link Cited by: §4.1.

Qwen Team (2025) Qwen3 technical report. CoRR abs/2505.09388. External Links: Document, Link Cited by: Table 7, §4.1.

I. Reinig, I. Rehbein, and S. P. Ponzetto (2024) How to do politics with words: investigating speech acts in parliamentary debates. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), Torino, Italia, pp. 8287–8300. External Links: Link Cited by: §1, §1, §2, §4.1, §4.2, §5.3.

T. Schuster, O. Ram, R. Barzilay, and A. Globerson (2019) Cross-lingual alignment of contextual word embeddings, with applications to zero-shot dependency parsing. External Links: 1902.09492, Link Cited by: §1.

J. Skubic and D. Fišer (2024) Parliamentary discourse research in political science: literature review. In Proceedings of the IV Workshop on Creating, Analysing, and Increasing Accessibility of Parliamentary Corpora (ParlaCLARIN) @ LREC-COLING 2024, Torino, Italia, pp. 1–11. External Links: Link Cited by: §1.

S. Subramanian, T. Cohn, and T. Baldwin (2019) Target based speech act classification in political campaign text. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM 2019), Minneapolis, Minnesota, pp. 273–282. External Links: Link, Document Cited by: §1, §4.1.

A. Templeton, T. Conerly, J. Marcus, J. Lindsey, T. Bricken, et al. (2024) Scaling monosemanticity: extracting interpretable features from Claude 3 Sonnet. Transformer Circuits Thread. External Links: Link Cited by: §1.

A. Volkens, P. Lehmann, T. Matthieß, N. Merz, S. Regel, B. Weßels, and Wissenschaftszentrum Berlin Für Sozialforschung (WZB) (2017) Manifesto project dataset. Manifesto Project (en). External Links: Document, Link Cited by: §3.

T. Walter, C. Kirschner, S. Eger, G. Glavaš, A. Lauscher, and S. P. Ponzetto (2021) Diachronic analysis of german parliamentary proceedings: ideological shifts through the lens of political biases. External Links: 2108.06295, Link Cited by: §1, §2, §3.

W. Xu and R. Futrell (2024) Syntactic dependency length shaped by strategic memory allocation. In Proceedings of the 6th Workshop on Research in Computational Linguistic Typology and Multilingual NLP, St. Julian’s, Malta, pp. 1–9. External Links: Link, Document Cited by: §2.

C. Zou, D. Jiao, and L. Hu (2026) Deciphering cultural representations in large language models via sparse autoencoders. In Findings of the Association for Computational Linguistics: ACL 2026, M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, pp. 5656–5677. External Links: Link, Document, ISBN 979-8-89176-395-1 Cited by: §2.

Appendix A Dataset Sources

Figure 5: Tokens per decade (Calculated using Llama-3 tokenizer, log scale) for each language, 1800s–2020s; colour key at top left.

媒体内容 · 前往原文查看

Lang. Source Country Documents Coverage Description

en hansard_historic GB 10,379,346 1803–2004 UK Commons and Lords, Historic Hansard.

en theyworkforyou GB 7,149,348 1945–2026 UK Commons and Westminster Hall debates, TheyWorkForYou.

en parlamint_gb GB 673,121 2015–2022 UK component of ParlaMint 5.0.

en manifesto_uk GB 84 1964–2024 UK party manifestos from the Manifesto Project.

pl ppc_pl PL 13,898,425 1919–2025 Polish Parliamentary Corpus, Sejm and Senate.

pl parlamint_pl PL 229,012 2015–2022 Polish component of ParlaMint 5.0.

pl sejm_pl PL 148 2015–2023 Sejm proceedings and interpellations.

pl manifesto_pl PL 43 1991–2019 Polish party manifestos from the Manifesto Project.

it itaparl IT 5,639,906 1948–2022 Camera dei Deputati speech turns, ItaParl.

it parlamint_it IT 174,184 2013–2022 Italian component of ParlaMint 5.0.

it italian_parliament IT 34,857 1848–2022 Camera and Senato, Kingdom of Italy through Republic, full-text OCR.

it manifesto_it IT 112 1963–2018 Italian party manifestos from the Manifesto Project.

de parlamint_de_beta DE 2,684,792 1949–2025 ParlaMint-DE beta, Bundestag debates.

de reichstag_bsb DE 2,645,175 1867–1939 Reichstagsprotokolle, full-text OCR.

de germaparl DE 1,042,888 1949–2021 GermaParlTEI Bundestag debates.

de bundestag_official DE 620,975 2025–2026 Bundestag DIP API, official recent records.

de parlamint_at AT 232,980 1996–2022 Austrian parliament, German-language ParlaMint component.

de ddb_newspapers DE 100,000 1945–2024 German Digital Library historical newspapers.

de manifesto_de DE 104 1949–2025 German party manifestos from the Manifesto Project.

tr tbmm_speeches_v1 TR 1,207,674 1950–2023 Turkish Grand National Assembly linked corpus.

tr parlamint_tr TR 682,387 2011–2022 Turkish component of ParlaMint 5.0.

tr manifesto_tr TR 28 1954–2018 Turkish party manifestos from the Manifesto Project.

Table 4: All sources in the corpus. Counts are post-deduplication counts in the unified record schema. The main analyses use parliamentary speech; manifestos and German newspapers are retained for corpus breadth.

Appendix B Dataset examples

Every one of the 22 sources in Table 4 is normalised onto the single 16-field record schema described in §3. Table 5 shows one complete record and Table 6 shows one record per language, drawn from each language’s principal parliamentary source. Together they make concrete what the cross-lingual comparison rests on: the lang, year and doc_type fields carry the same meaning in every collection, so a metric computed over English speech is computed over the same kind of object as the same metric over Italian or Turkish speech.

媒体内容 · 前往原文查看

Field Value

id germaparl_a98144794f94707d

source germaparl

country DE

lang de

doc_type parliamentary_speech

date 1994-05-26

year 1994

title Bundestag 1994-05-26 WP12 Nr230

speaker Gerd Wartenberg (Berlin)

speaker_role NA

party SPD

house Bundestag

granularity speech

url (null)

meta {"leg": "12", "sess": "230"}

text Herr Staatssekretär Lintner, meinen Sie denn — bei aller Kritik an diesem Gesetz —, daß es sinnvoll ist, diese alten Hüte wieder hervorzuholen und diesen Streit fortzusetzen: Sind wir ein Einwanderungsland oder kein Einwanderungsland? Glauben Sie, daß uns das weiterführt? Es sind in die Bundesrepublik Deutschland 6 Millionen Menschen zugewandert. Können wir nicht einmal von dieser Basis ausgehen, diese Menschen hier vernünftig zu integrieren […] (Beifall bei der SPD und dem BÜNDNIS 90/DIE GRÜNEN)

Table 5: One complete record from the unified corpus (source germaparl, German Bundestag, 26 May 1994), abridged only in text.meta holds the source-specific fields that do not fit the shared schema and is stored as a JSON string so the Parquet schema stays byte-identical across sources; here it records the legislative period and sitting number. Fields absent at source are null (url) or carry the source’s own missing-value token (speaker_role = NA).

媒体内容 · 前往原文查看

Lang. Source Year Speaker / Party text (excerpt)

en hansard_historic 1803 The Speaker / — acquainted the House that the House h3d, in obedience to his Majesty’s command, attended in the House of Peers, to hear his Majesty’s most gracious Speech from the Throne […]

de germaparl 1994 G. Wartenberg / SPD Sind wir ein Einwanderungsland oder kein Einwanderungsland? […] Es sind in die Bundesrepublik Deutschland 6 Millionen Menschen zugewandert. […]

it itaparl 1950 O. L. Scalfaro / DC SCALFARO. Voi non credete nella Costituzione; ne usate come tappa per la conquista violenta del potere. (Interruzione del deputato Grilli). […] Dove vie liberta nel giudizio, si ha affermazione di giustizia […]

pl ppc_pl 1919 F. Radziwiłł / KPK Wysoka Izbo! W chwili gdy przypada na mnie obowiązek zagajenia tego Wysokiego Zgromadzenia, oprócz serdecznej radości czuję wzruszenie […]

tr tbmm_speeches_v1 2018 M. A. Kaya / AKP Sayın Başkanım, tabii ki konuşmacının özellikle ortaya koymuş olduğu gerçek dışı beyanlarına girmeden önce şundan bahsetmek isterim […]

Table 6: One record per language, each taken from that language’s principal parliamentary source, showing the same 16-field schema realised across all five.

Data schema.

Three properties of the release are worth stating explicitly, because they constrain how the records may be used. First, granularity distinguishes a speech (one uninterrupted turn) from a segment (a chunk of a longer document, suffixed #segk in the id) and from a session (a whole sitting); metrics that assume a single speaker must be restricted to granularity = speech. Second, date is a string and may be partial—1919 for an OCR’d annual volume, 1964-10 for a manifesto, a full YYYY-MM-DD for a modern sitting— so the integer year field, not date, is the safe key for the decade binning used throughout. Third, id is inherited from the originating source and is a document identifier, not a primary key: in the session-structured sources (germaparl, bundestag_official, theyworkforyou) every speech of one sitting shares its sitting’s identifier, so records must be addressed by row rather than looked up by id.

Appendix C ChronoLens methodological details

C.1 Concept matching and sentence sampling

Concept inventory.

The analysis covers 13 concepts. Migration and gender are the target concepts. Defence, democracy, economy, environment, Europe, religion, security, taxation, and technology represent general political discourse. Road and water are frequency-matched controls. We keep the three groups separate because a pattern found for all political concepts may reflect a change in parliamentary discourse generally, while a pattern also found for road and water is more likely to reflect corpus or model variation.

Each concept is represented by a language-specific set of lowercased stems. A stem is matched at the beginning of a word, with longer alternatives tested first. Prefix matching is used because exact word matching would miss common inflected forms, particularly in German, Polish, and Turkish. It may also include words that share a stem but not the intended sense. We therefore report the full lexicons with the released data and inspect ambiguous terms during error analysis.

A sentence is assigned to at most one concept. When several concepts occur in the same sentence, the first matched concept determines the assignment. This rule prevents the same sentence from contributing to several concept trajectories, although it means that concept counts are a partition of the harvested sentences rather than an estimate of their total corpus frequency.

Concept masking.

For every matched sentence, we retain the original text and create a masked version in which the concept expression and its inflectional ending are replaced with <concept>. The policy-frame labels and matching strata are calculated from the masked text. During activation extraction, we also exclude the concept span from mean pooling. These two operations prevent the concept term itself from determining either the matching stratum or the model representation.

Matched strata.

We divide sentences into strata based on token length and dominant policy frame. Length is divided into four bands: fewer than 12 tokens, 12 to 24 tokens, 25 to 44 tokens, and at least 45 tokens. The policy frame is selected from a fixed multilingual inventory that covers domains such as economy, migration, security, rights, welfare, employment, environment, law, education, religion, and foreign policy.

Cross-lingual and temporal tuples are sampled only within a shared stratum. The sentences in a tuple therefore have similar lengths and broad policy content. This matching does not make them translations or paraphrases, but it removes two simple sources of variation that the crosscoder could otherwise use to distinguish conditions.

Temporal balancing.

We retain at most 400 sentences for each concept, language, and decade. The cap prevents recent decades from dominating the analysis because they contain more digitized material. Reservoir sampling with a fixed seed is used so that the selected sentences remain reproducible. A decade is used as a crosscoder condition only when it contains at least 250 distinct sentences after matching. The final data provide approximately 293 distinct sentences per decade condition, 855 per period-level condition, and 3,187 per cross-lingual condition. These values refer to distinct sentence representations, not the larger number of tuples obtained by resampling them.

C.2 Backbones, representations, and layer selection

We represent each sentence by mean-pooling the content-token residual stream from one block of a frozen language model. Extraction uses a single forward pass under no_grad, and vectors are stored in fp16. When a target concept is present, its tokens are excluded from pooling so that changes in its surface form are not mistaken for contextual change; topic-agnostic sentences are pooled in full. We use four multilingual backbones, Qwen3-8B, Llama-3.1-8B, Mistral-Nemo-2407, and EuroLLM-9B-2512, chosen to vary model family, tokenizer, and pretraining data.

For each backbone, we evaluate layers at approximately 25%, 50%, and 75% of model depth and select the layer with the highest mean probe selectivity across four linguistic levels and five languages. Selectivity is defined as accheld-out​(real)−accheld-out​(control), where the control probe uses permuted labels; sentence-length and OCR-noise probes are reported but excluded from selection. Table 7 gives the selected layers: L9 for Qwen3-8B, L16 for Llama-3.1-8B, L20 for Mistral-Nemo-2407, and L10 for EuroLLM-9B-2512. Differences between the best and second-best depths are small (0.002–0.026), indicating that the linguistic properties are similarly decodable across the middle layers. We therefore fix one selected layer per backbone before computing any historical results.

媒体内容 · 前往原文查看

Backbone Blocks dmodel 25% 50% 75%

Qwen3-8B (Qwen Team, 2025) 36 4096 L9: 0.254† L18: 0.235 L27: 0.248

Llama-3.1-8B (Llama Team, 2024) 32 4096 L8: 0.278 L16: 0.280† L24: 0.273

Mistral-Nemo-2407 (Jiang et al., 2023; MistralAI, 2025) 40 5120 L10: 0.277 L20: 0.287† L30: 0.278

EuroLLM-9B-2512 (Martins et al., 2024) 42 4096 L10: 0.295† L21: 0.266 L32: 0.269

Table 7: Backbone architectures and layer-selection results. The final three columns report mean probe selectivity at approximately 25%, 50%, and 75% of model depth, averaged over the four linguistic levels and five languages. The selected layer for each backbone is shown in bold and marked with †.

C.3 Crosscoder training and calibration

Input tuples.

A crosscoder receives matched tuples with one representation from every condition. In the cross-lingual setting, a tuple contains one representation from each of the five languages. In the period-level setting, it contains one representation from each available historical period. In the decade-level setting, it contains one representation from every decade that meets the minimum data requirement.

Each crosscoder is trained on 20,000 sampled tuples. We reserve 10% for validation, leaving 18,000 training tuples and 2,000 validation tuples. Sampling may reuse a sentence when the shared strata do not contain enough distinct examples. We therefore record both the number of sampled tuples and the number of distinct sentences in every condition.

Condition standardization.

Activation magnitudes can differ between languages and historical periods. Without normalization, a condition with larger vector norms could contribute more to the reconstruction loss. We standardize each condition separately:

𝐱~c=𝐱c−𝝁cσc,σc=𝔼​‖𝐱c−𝝁c‖22d, (7)

where c denotes a condition and d is the model hidden dimension. This normalization preserves directional information while reducing scale differences between conditions.

Crosscoder objective.

Following prior work on crosscoders (Jiralerspong and Bricken, 2026; Minder et al., 2026), one sparse feature vector is inferred jointly from all conditions:

𝐟 =ReLU⁡(∑c=1C𝐱~c​𝐖encc+𝐛enc), (8)

𝐱^c =𝐟𝐖decc+𝐛decc. (9)

The feature index is shared across conditions, while each condition has its own encoder and decoder weights. This structure allows feature j to be compared directly across languages or periods.

The dictionary size is set relative to the hidden dimension of the backbone rather than fixed across models. We use an expansion factor of two, so a model with hidden dimension d receives 2​d features. This keeps the degree of overcompleteness comparable across backbones.

Sparsity.

We use BatchTopK rather than an L1 penalty. BatchTopK retains the strongest feature activations in each batch but does not directly penalize their magnitude. This matters because the shared and condition-specific classification depends on decoder magnitudes. An L1 penalty can shrink a feature unevenly across conditions and make a shared feature appear specific (Minder et al., 2026).

The target active fraction is 0.10. An auxiliary reconstruction term revives features that have remained inactive for 200 optimization steps. During inference, a threshold estimated from the surviving training activations removes small activations. Thresholds are calibrated separately for each condition because a threshold estimated from the joint tuple representation does not transfer directly to a single condition.

Shared and condition-specific features.

For feature j, let

nc,j=‖𝐖decc​[j]‖2 (10)

be its decoder norm in condition c. A feature is considered shared when the ratio between its largest and smallest decoder norms is at most four. It is considered condition-specific when its largest norm is at least four times its second-largest norm. Features that meet neither rule are marked as mixed. Features with negligible total decoder norm are marked inactive.

The main threshold is four, but we also calculate the split at ratios of two, three, six, and eight. This sensitivity analysis shows whether the conclusion depends on one boundary.

Latent-scaling check.

Decoder norms can overstate specificity when training suppresses a shared direction in one condition or divides one shared mechanism across several features. For every initially specific feature, we test whether its decoder direction also explains reconstruction or residual structure in the other conditions. If it does, the feature is relabeled as shared. This check follows the latent-scaling analysis of Minder et al. (2026).

Reconstruction value.

We evaluate reconstruction on the held-out tuples using fraction of variance unexplained:

FVU=∑i‖𝐱~i−𝐱^i‖22∑i‖𝐱~i−𝐱~¯‖22. (11)

An FVU of zero indicates perfect reconstruction, while an FVU of one is no better than predicting the condition mean. We mark a checkpoint as unusable when held-out FVU exceeds 0.50 or when any condition falls below the minimum number of distinct sentences. Feature-level conclusions are not drawn from unusable checkpoints.

Condition-shuffled null.

A decomposition may produce apparently specific features even when its conditions contain no systematic difference. We therefore train a null crosscoder with the same data, strata, condition count, dictionary size, and optimization procedure. In the null tuples, every position is sampled from the pooled condition data, making condition identities exchangeable.

We compare the real and null crosscoders using their number of specific features and their decoder-norm ratios. The real crosscoder is considered to resolve its conditions only when it contains at least twice as many specific features as the null and also meets the minimum feature-coverage requirement. When it does not pass this test, analyses that depend on the shared-specific split are marked unsupported rather than interpreted as evidence that the conditions are identical.

C.4 Feature attribution and interventions

Candidate features.

For each crosscoder, we score the 1,000 features with the highest activation mass. Restricting the intervention analysis to these features keeps the computation manageable and avoids testing features that are rarely active. The main trajectory analysis is not restricted to features that pass every diagnostic because the resulting feature set is too small to define stable trajectory spaces. Instead, the diagnostics are reported as evidence about the reliability of individual features.

Diagnostic gate.

The six diagnostics are temporal selectivity, linguistic specificity, necessity, sufficiency, split-half robustness, and cross-lingual recurrence where applicable. Their conjunction is stored as a diagnostic indicator. It is not used to remove features from the main analysis. This choice avoids constructing trajectories from a very small and uneven set of surviving features, while still showing how often a feature receives support beyond correlation.

Linguistic attribution by ablation.

Let fs,j be the activation of feature j on sentence s, and let 𝜹c,j be its decoder direction mapped back to the original activation scale. We remove the feature only from sentences on which it is active:

𝐱s−j=𝐱s−fs,j​𝜹c,j. (12)

For each linguistic level λ, we measure the average decrease in the probability assigned to the correct label:

Dλ​(j)=1|Aj|​∑s∈Aj[pλ​(ys∣𝐱s)−pλ​(ys∣𝐱s−j)], (13)

where Aj is the set of sentences on which the feature is active. The feature is assigned to the level with the largest decrease. Measuring only active sentences prevents a localized feature effect from being diluted by sentences on which the feature contributes nothing.

We use two criteria to distinguish a selective level attribution from a small or ambiguous probe effect. First, the largest aggregate probability decrease must exceed an absolute threshold of 0.01, which excludes effects that are negligible on the probability scale. Second, it must be at least 1.2 times the second-largest decrease, which requires a minimum separation between the two most affected levels. These criteria implement a deliberately conservative assignment rule: features with weak effects or comparable effects across several levels remain unassigned rather than being forced into a single category. Features active on fewer than 20 held-out sentences also remain unassigned because their intervention effects cannot be estimated reliably. For cross-lingual shared features, the final level is determined by the majority assignment across languages.

Temporal selectivity.

A feature must first show some relation to historical time. We compare its activation on early and late sentences using a held-out Mann-Whitney statistic converted to a direction-independent area under the curve. The feature passes the temporal-selectivity diagnostic when its AUC is at least 0.55. This criterion establishes association with time but does not by itself establish a functional role.

Necessity.

Necessity asks whether removing a feature weakens the later-period prediction of the probe associated with its assigned level. We compute the mean decrease in the relevant probability on later sentences where the feature is active. The effect must exceed the 95th percentile obtained from random directions with the same norm.

Sufficiency.

Sufficiency asks whether adding the feature to earlier representations moves the probe toward the later-period prediction. We insert the decoder direction using the feature’s mean activation on later firing sentences. As with necessity, the effect must exceed the matched random-direction null. The random comparison is used because a large edit to an activation may change a probe even when the edited direction has no special linguistic role.

Split-half robustness.

We divide sufficiently large matching strata into two random halves and calculate the feature’s mean activation in each half. A feature passes the robustness diagnostic when the two activation profiles have a Spearman correlation of at least 0.4. This test checks whether the feature depends on a small number of sentences within a stratum.

Cross-lingual recurrence.

For cross-lingual shared features, we test whether inserting the feature has the same directional effect in a majority of the other languages. This test uses the semantics probe because its policy-frame classes have the same interpretation across the five languages. Tense and subordinate-clause labels are not fully equivalent across languages, so we do not use them for this cross-lingual sign test.

Threshold sensitivity.

We repeat the attribution and trajectory analyses using absolute thresholds in {0.005,0.01,0.02} and relative-separation thresholds in {1.1,1.2,1.5}. Across these settings the composition of the assigned inventory is essentially fixed (morphology 11.3–12.3%, syntax 16.3–17.5%, semantics 69.7–71.9%, pragmatics 0.4–0.6%), the same 30 of 56 (concept, level) cells clear the minimum-feature criterion, and the per-cell endpoint convergence indices stay tightly coupled to the main setting (Pearson r≥0.86, Spearman ρ≥0.84, sign agreement 87–100%). Pragmatics, however, never reaches that count at any setting (0.4–0.6% of assigned features), so the filtered trajectories cover morphology, syntax and semantics only. The pre-registered target × decade convergence contrast keeps its sign and its significance at every setting (β^ = −0.29 to −0.12, p≤0.020), target-concept morphology (−0.18 to −0.02) and semantics (−0.10 to −0.07) diverge in all 9 runs (slope per century) while control and political cells converge in 54 of 54, and the pre-registered verdict is C throughout, the only sign that moves being target syntax (positive in 6 of 9), although stricter thresholds reduce the number of assigned features from 3,581 (at 0.005, 1.1) to 1,297 (at 0.02, 1.5). We therefore use 0.01 and 1.2 in the main analysis as an intermediate setting that excludes weak and ambiguous effects while retaining sufficient feature coverage for stable trajectory estimation.

媒体内容 · 前往原文查看

τabs τrel N cells r sign β^ p

0.005 1.1 3,581 30 0.90 97% −0.12 0.020

0.005 1.2 3,507 30 0.96 97% −0.14 0.009

0.005 1.5 3,275 30 0.97 93% −0.15 0.006

0.01 1.1 2,429 30 0.93 100% −0.18 <10−3

0.01 1.2 2,373 30 1.00 100% −0.18 <𝟏𝟎−𝟑

0.01 1.5 2,198 30 0.99 100% −0.18 <10−3

0.02 1.1 1,436 30 0.90 87% −0.29 <10−4

0.02 1.2 1,405 30 0.91 87% −0.25 <10−3

0.02 1.5 1,297 30 0.86 87% −0.24 <10−3

Table 8: Sensitivity of level attribution and of the trajectory analysis to criterion 2’s absolute floor τabs and relative separation τrel. N: assigned features (all concepts); cells: (concept, level) cells clearing the minimum feature count; r and sign: Pearson correlation and sign agreement of the per-cell endpoint convergence index against the main setting (bold); β^: target × decade convergence contrast (negative = targets diverge relative to controls).

C.5 Probe diagnostics

The four probes differ in their number of classes and in the distribution of their automatically derived labels. Raw accuracy is therefore not directly comparable across tasks: in particular, a probe can achieve high accuracy when one class dominates the held-out data. We report three diagnostics separately for each language–task pair. Majority is the proportion of held-out examples assigned to the most frequent class; accuracy is the probe’s held-out accuracy; and selectivity is the difference between held-out accuracy and the accuracy of the same probe trained on permuted labels (Hewitt and Liang, 2019).

媒体内容 · 前往原文查看

Language Task Classes Majority Accuracy Selectivity

English Tense 3 0.51 0.93±0.01 0.50±0.00

English Subordination 2 0.79 0.93±0.01 0.19±0.00

English Policy topic 13 0.08 0.75±0.04 0.67±0.04

English Speech act 3 0.84 0.97±0.01 0.21±0.01

German Tense 3 0.65 0.87±0.01 0.34±0.01

German Subordination 2 0.62 0.91±0.00 0.36±0.01

German Policy topic 13 0.08 0.71±0.03 0.64±0.03

German Speech act 3 0.90 0.95±0.00 0.17±0.01

Italian Tense 3 0.71 0.92±0.01 0.33±0.01

Italian Subordination 2 0.72 0.89±0.06 0.23±0.10

Italian Policy topic 13 0.10 0.57±0.29 0.49±0.30

Italian Speech act 3 0.96 0.97±0.01 0.55±0.03

Polish Tense 3 0.46 0.90±0.02 0.50±0.02

Polish Subordination 2 0.73 0.91±0.01 0.28±0.01

Polish Policy topic 13 0.08 0.78±0.03 0.70±0.03

Polish Speech act 3 0.92 0.96±0.00 0.14±0.01

Turkish Tense 13 0.58 0.91±0.01 0.66±0.00

Turkish Subordination 2 0.66 0.81±0.01 0.26±0.00

Turkish Policy topic 13 0.08 0.73±0.03 0.65±0.03

Turkish Speech act 3 0.93 0.96±0.01 0.59±0.01

Table 9: Probe diagnostics by language and task. Majority is the proportion of held-out examples belonging to the most frequent class. Accuracy is held-out probe accuracy, and selectivity is held-out accuracy minus accuracy under permuted labels (Hewitt and Liang, 2019).

Table 9 shows why accuracy and selectivity must be considered jointly. Tense selectivity ranges from 0.33 to 0.50 in the four languages for which the task is available. Policy-topic selectivity is also comparatively high (0.49–0.70), although the Italian result varies substantially across backbones. Subordination selectivity ranges from 0.19 to 0.36. Speech-act accuracy is uniformly high (0.95–0.97), but the majority class accounts for 0.84–0.96 of the held-out examples; accordingly, speech-act selectivity is only 0.05–0.21. We therefore treat these probes as task-based operationalizations rather than exhaustive measurements of morphology, syntax, semantics, and pragmatics, and interpret findings based on lower-selectivity tasks with additional caution.

These diagnostics measure the predictability and class balance of the automatically derived labels. They do not directly establish the correctness of those labels across languages or historical periods. A language- and period-stratified manual label audit remains necessary to quantify label quality independently of probe performance.

C.6 Additional trajectory and statistical details

Trajectory coordinates.

For concept κ, language ℓ, decade t, and linguistic level λ, we average the activation vector over the shared features assigned to that level:

𝐮λ​(κ,ℓ,t)=1|Cκ,ℓ,t|​∑s∈Cκ,ℓ,t𝐟s​[𝒮λshared]. (14)

A cell must contain at least 25 sentences. Cells with fewer than eight available features are retained but marked as low coverage.

Alignment.

The Convergence Index measures whether two languages become closer, but it does not indicate whether their changes point in the same direction. We therefore calculate the cosine between their net displacement vectors:

parend=cos⁡(𝐮ℓ,tlast−𝐮ℓ,tfirst,𝐮ℓ′,tlast−𝐮ℓ′,tfirst). (15)

We also calculate stepwise alignment as the mean cosine between corresponding decade-to-decade changes. A pair is labeled parallel when stepwise alignment is above 0.20 and anti-parallel when it is below −0.20. Values between these thresholds remain unclassified.

Mechanism sharing.

For every historical bin, we calculate the proportion of feature activation mass assigned to cross-lingual shared features. We estimate its change over time with a linear slope. A positive slope indicates increasing use of shared features, while a flat or negative slope indicates that representational convergence is not accompanied by greater mechanism sharing.

Mixed-effects specification.

The main model is

CI∼decade×concept​_​group×level, (16)

with variance components for concept and language pair. Decade is centered and scaled so that coefficients describe change per century. We fit the model separately for every backbone. A pooled model includes backbone as an additional variance component.

We report the direction of each effect, its variation across backbones, and agreement on the resulting convergence category. Individual features are not aligned across backbones because their feature spaces are learned independently. Cross-model comparison is performed only at the level of the estimated findings.

Appendix D Observable linguistic changes

The main analysis measures change in learned representations. To interpret these aggregate changes, we also examine 18 observable linguistic measures derived from the same sentence-level annotations used for feature attribution. The measures are grouped into morphology, syntax, semantics, and pragmatics. Examples include passive voice for morphology, subordination for syntax, human subjects for semantics, and personal deixis for pragmatics.

We use the common 1950–2020 interval and balance the samples by decade, sentence length, and topic. For each measure and language, we fit a linear trend and report the estimated change over the full interval in percentage points. A positive value means that the measure became more frequent between 1950 and 2020. A negative value means that it became less frequent.

We test whether each fitted trend differs from a flat trajectory and apply Benjamini–Hochberg correction across the 90 language–measure tests. Bold values in Table 10 have q<0.05. Several measures are based on terciles defined separately using each language’s historical distribution. Their absolute frequencies should therefore not be compared across languages. The direction and size of their changes over time can still be compared.

Figure 6: Observable linguistic changes between 1950 and 2020. The left panel reports the fitted change in 18 measures for five languages. Positive values indicate increases and negative values indicate decreases; bold values have Benjamini–Hochberg corrected q<0.05. The upper middle panel reports cosine similarity between each pair of languages, with 95% bootstrap intervals. Positive cosine values indicate that the two languages tend to show increases and decreases in the same measures. Negative values indicate that measures that increase in one language tend to decrease in the other. Values near zero indicate no consistent shared direction. The upper right panel reports these similarities separately for morphology, syntax, semantics, and pragmatics. The lower panels show selected measure frequencies by decade. Tercile-based measures are defined separately for each language, so only their changes over time should be compared across languages.

媒体内容 · 前往原文查看

Level Measure English German Italian Polish Turkish

Morphology Past reference −14.9 −6.0 −3.1 +0.5 +6.3

Passive voice −6.9 −10.6 −6.9 +9.8 −8.9

High inflectional load −0.4 −0.3 +10.3 +11.1 +12.7

High nominal density −1.3 −2.2 −0.1 +7.2 +11.0

Syntax Any subordination +1.9 −4.3 +13.0 +16.7 +10.1

Deep parse tree +2.1 −1.0 +0.3 +2.9 −9.0

Long dependencies −4.7 −7.6 −9.6 +12.0 +11.6

Left-branching −4.9 −4.4 −5.2 +22.6 +16.8

Heavy NP modification −4.8 −3.4 −8.2 −23.0 +4.3

Much coordination +9.3 +14.9 −2.2 −20.7 −1.5

Semantics Stative predicate +1.8 −0.5 −1.9 −0.2 −1.6

Human subject +2.8 +2.7 −1.4 +12.9 +10.9

Abstract reference −0.6 −1.4 +7.4 −15.5 −2.1

Abstract object +4.1 +0.6 +5.7 −7.5 +2.3

Quantified reference −1.5 +0.9 −10.6 +0.1 −2.8

Pragmatics Personal deixis +7.3 +7.9 +23.5 +33.0 +20.0

Stance-marked +3.0 +2.4 +10.3 +6.7 +8.4

Reported voice −1.9 −3.7 −3.4 +11.8 −1.4

Table 10: Signed fitted change in 18 observable linguistic measures between 1950 and 2020, measured in percentage points. Bold values have Benjamini–Hochberg corrected q<0.05. Positive values indicate that a measure became more frequent, while negative values indicate that it became less frequent.

Of the 90 fitted trends, 37 remain significant after correction. Personal deixis shows the clearest shared pattern: it increases in all five languages, with significant increases in German, Italian, Polish, and Turkish. Several other measures do not share one direction across languages. Long dependencies decrease significantly in English, German, and Italian but increase significantly in Polish. Left-branching decreases significantly in English and German but increases significantly in Polish and Turkish. Passive voice decreases significantly in English, German, and Turkish but increases significantly in Polish. Coordination increases significantly in English and German but decreases significantly in Polish. Thus, languages can have similar overall amounts of change while differing in which linguistic properties increase or decrease.

D.1 Cross-language agreement

To measure whether two languages change in similar ways, we represent each language by a vector containing its fitted changes for the measures within one linguistic level. We then compute cosine similarity between the vectors of each language pair. A positive value means that the two languages tend to increase and decrease in the same measures. A negative value means that they tend to change in different directions. A value near zero means that there is no consistent relation between their patterns of change.

媒体内容 · 前往原文查看

Level # measures Mean cosine Minimum Maximum Pairs with CI excluding 0 Leave-one-measure-out mean Sign agreement Timing r

Morphology 4 +0.21 −0.62 (DE–PL) +0.81 (EN–DE) 7/10 [+0.08,+0.27] 0.45 +0.18

Syntax 6 0.00 −0.63 (DE–PL) +0.87 (EN–DE) 9/10 [−0.07,+0.04] 0.47 +0.04

Semantics 5 +0.29 −0.57 (IT–PL) +0.82 (DE–TR) 3/10 [+0.12,+0.35] 0.56 +0.07

Pragmatics 3 +0.92 +0.73 (DE–PL) +1.00 (IT–TR) 10/10 [+0.53,+0.99] 0.87 +0.43

Table 11: Cross-language agreement by linguistic level. Cosine similarity is computed between the observable change vectors of each language pair. “Pairs with CI excluding 0” reports how many of the ten pairwise 95% bootstrap intervals exclude zero. “Leave-one-measure-out mean” gives the range of the mean cosine after removing one measure at a time. Sign agreement is the mean proportion of measures with the same trend direction. Timing correlation is the mean correlation between the corresponding decade-level trajectories.

Table 11 shows that cross-language agreement differs by linguistic level. Pragmatics has the strongest agreement, with a mean cosine similarity of 0.92. All ten language pairs have positive bootstrap intervals that exclude zero. This result is driven mainly by the widespread increase in personal deixis, although the leave-one-measure-out mean remains positive, ranging from 0.53 to 0.99.

Morphology and semantics show weaker positive agreement, with mean cosine similarities of 0.21 and 0.29. Syntax has a mean of 0.00. This does not mean that the syntactic measures remain unchanged. Instead, different language pairs show different directions: four significant pairwise similarities are positive and five are negative. There is therefore no single syntactic pattern shared by all five languages.

媒体内容 · 前往原文查看

Language pair Morphology Syntax Semantics Pragmatics All measures

English–German +0.81 +0.87 +0.48 +0.98 +0.81

Polish–Turkish +0.41 +0.55 +0.61 +0.90 +0.63

English–Italian +0.42 +0.47 +0.35 +0.99 +0.51

Italian–Turkish +0.67 −0.21 +0.09 +1.00 +0.43

German–Turkish +0.12 −0.52 +0.82 +0.94 +0.01

German–Italian +0.56 +0.10 −0.42 +0.95 +0.36

English–Turkish −0.16 −0.54 +0.66 +0.99 −0.01

Italian–Polish +0.22 +0.27 −0.57 +0.87 +0.38

English–Polish −0.33 −0.38 +0.12 +0.82 −0.03

German–Polish −0.62 −0.63 +0.73 +0.73 −0.23

Table 12: Cosine similarity between observable change vectors for every language pair and linguistic level. Bold values have 95% bootstrap intervals that exclude zero. The final column computes cosine similarity over all 18 measures.

Table 12 shows that the direction of change depends on both the language pair and the linguistic level. For example, English and German have positive similarities for morphology, syntax, and pragmatics, while German and Polish have negative similarities for morphology and syntax but positive similarities for semantics and pragmatics. A positive overall value therefore does not imply agreement at every linguistic level.

The pairwise results are not explained by language-family membership alone. The mean cosines for pairs within the Indo-European group and pairs involving Turkish are, respectively, 0.18 and 0.26 for morphology, 0.12 and −0.18 for syntax, 0.12 and 0.55 for semantics, and 0.89 and 0.95 for pragmatics. Some pairs involving Turkish have high semantic or pragmatic similarity, while some pairs of Indo-European languages have negative similarity for morphology or syntax. Because the analysis contains only five languages, these comparisons are descriptive. They do not establish a general relation between genealogical relatedness and the direction of historical change.

D.2 Long-window results

The common 1950–2020 interval supports direct comparison across languages but does not use the full historical record available for English, German, Italian, or Polish. Table 13 therefore reports the largest significant trends over each language’s complete available interval. The values are fitted percentage-point changes per century. They should not be compared as total changes because the historical spans differ.

媒体内容 · 前往原文查看

Language Span Largest significant changes per century

English 1800–2020 Past reference −34.5; personal deixis +21.5; human subject +15.0; passive voice −12.1.

German 1860–2020 Personal deixis +22.1; human subject +20.7; passive voice −11.3; heavy NP modification −8.3.

Italian 1840–2020 Coordination +11.7; left-branching −9.3; heavy NP modification +8.8; human subject −8.4.

Polish 1910–2020 High nominal density +11.6; deep parse tree +9.2.

Turkish 1950–2020 Personal deixis +28.8; left-branching +24.3; high inflectional load +18.3; high nominal density +15.8.

Table 13: Largest significant observable changes over each language’s complete historical record. Values are fitted percentage-point changes per century, and all listed trends have q<0.05. The unequal intervals make this analysis supplementary to the balanced 1950–2020 comparison.