CHRONOBERG: Temporal NLP and Geodesy Approaches

• CHRONOBERG is a dual-nature framework: one part is a temporally structured corpus for diachronic language analysis, while the other uses clock-based measurements in general relativity to recover Earth’s gravity.
• The NLP component curates 25,061 English books (1750–2000), employing time-binned Word2Vec embeddings and Compass Aligned Distributional Embeddings to track semantic and affective shifts.
• The geodesy component leverages precise clock comparisons and exact redshift relations to infer multipolar gravitational fields, integrating both geometric and Doppler factors.

Searching arXiv for CHRONOBERG to ground the article in the cited literature. CHRONOBERG denotes two distinct constructs in contemporary arXiv literature. In natural language processing, it is a temporally structured corpus of English book texts spanning 1750–2000, designed to capture language evolution, diachronic semantics, and temporal awareness in foundation models (Hegde et al., 26 Sep 2025). In relativistic geodesy, it denotes a chronometric gravimetry concept in which clock comparisons on ground and in space are used to recover the Earth’s stationary gravity field through exact general-relativistic redshift relations (Philipp et al., 2023). The shared name therefore refers not to a single unified framework, but to two separate programs centered on temporal structure: one linguistic and data-centric, the other physical and measurement-centric.

1. Terminological scope

The name CHRONOBERG is used for two unrelated technical objects.

Usage	Domain	Core object
CHRONOBERG	NLP and foundation models	A temporally structured corpus of English book texts with temporal annotations and historically calibrated VAD lexicons
CHRONOBERG	Relativistic geodesy	A chronometric geodesy concept for gravity field recovery from clock comparisons

A frequent source of confusion is the assumption that CHRONOBERG names a single methodology. The literature instead supports a disambiguated reading. The corpus-oriented CHRONOBERG addresses diachronic variation, semantic drift, moderation drift, and temporal generalization in LLMs. The geodetic CHRONOBERG addresses gravitational redshift, relativistic level surfaces, multipole recovery, and frequency transfer in stationary spacetimes. A plausible implication is that the acronym’s recurrence reflects a shared emphasis on time-resolved inference rather than a shared scientific lineage.

2. Corpus construction, temporal structure, and curation

As a linguistic resource, CHRONOBERG is a temporally structured, large-scale corpus of English book texts curated from Project Gutenberg and spanning 250 years of Late Modern English, from 1750 to 2000 (Hegde et al., 26 Sep 2025). It was introduced to address a limitation of common web-scale corpora used for LLMs: their lack of long-horizon temporal structure, which can prevent models from contextualizing diachronic evolution in semantics and social norms.

The released corpus contains 25,061 English-language books after filtering, corresponding to 2.7B tokens and 91M sentences. The starting pool was 73,500 Project Gutenberg items. Translations were removed (8.4%) to avoid temporal misalignment from translation timing, English-only filtering retained 80.3%, and non-textual materials accounted for 1.7% of exclusions. Books without inferred publication year were excluded (40.7%). The processed variant removes non-alphanumeric characters, and both raw and sentence-annotated versions are provided. Sentence segmentation and sentence-level annotation are organized per publication year.

Temporal organization is central to the dataset design. The corpus spans five primary 50-year intervals: 1750–1799, 1800–1849, 1850–1899, 1900–1949, and 1950–2000. Year-by-year grouping is also provided, but the authors recommend coarse bins no smaller than 15 years for evaluation in order to exceed metadata uncertainty and mitigate publication-date inference variance. Coverage is uneven across decades: it grows to the 1920s and then drops sharply because of copyright restrictions post-1929.

Metadata recovery is a defining technical contribution. Because original publication dates in Project Gutenberg are often absent or incorrect, CHRONOBERG implements metadata recovery and inference using external bibliographic sources, specifically OpenLibrary and Wikipedia, together with consistency checks against author lifespan. Posthumous publications are excluded whenever inconsistent. Manual validation over 100 sampled books from 1611–1912 found OpenLibrary to be the most reliable predictor, with mean absolute error $\pm 3.05$ years, standard deviation 5.20 years, $\mathrm{acc@5} = 76\%$, and $\mathrm{acc@7} = 79\%$. Wikipedia had lower coverage and higher error, while majority vote across predictors produced slightly better recall but higher MAE, 4.05 years, and poor scalability.

Topical coverage is broad but not uniform. Frequent subjects include Fiction (23.9%), History (12.0%), Juvenile fiction (9.4%), Social Life and Customs (5.8%), and 19th century (4.9%), alongside England, United States, Great Britain, Description and Travel, and Conduct of Life. Bookshelves include American Bestsellers (1895–1923), Science Fiction, Children’s Literature/Series, World War I, US Civil War, Historical Fiction, Humor, and Native America. The distribution implies strong pre-1920 representation and a substantial fiction skew.

3. Temporal annotations, VAD induction, and lexical semantic change

CHRONOBERG operationalizes diachronic affective semantics through time-sensitive Valence–Arousal–Dominance analysis (Hegde et al., 26 Sep 2025). The core method is to train separate Word2Vec embeddings for each 50-year slice, using 10 epochs, window size 5, and 300 dimensions, and then align the embedding spaces with Compass Aligned Distributional Embeddings (CADE) so that vectors can be compared directly across time.

The induced lexicons extend the contemporary NRC VAD resource to 335,804 words across five epochs. For each word in each slice, the method retrieves semantically similar neighbors in the aligned time-specific embedding space and averages the human-annotated NRC VAD scores of those neighbors. The paper gives the construction as

$$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$

Here, $\mathbf{e}_w \in \mathbb{R}^d$ is the embedding of target word $w$, $s(e_w,e_u)$ is the cosine similarity, $\mathcal{V}_{\text{VAD}}$ is the set of words with NRC VAD annotations, and $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$. Empirically, averaging Top-20 neighbors is favored; Top-100 nearest neighbors are retrieved to ensure 20 neighbors with NRC VAD coverage. Top-5 is reported as too biased, whereas Top-500 over-smooths toward neutrality.

Sentence-level annotations are designed to foreground emotionally salient material. The procedure averages over adjectives, verbs, and adverbs, represented as JJ/VB/RB, and then computes clause-level means. The final sentence score uses the minimum clause-level valence in order to reflect the strongest affective polarity in complex sentences. The paper states this as

$$\widehat{\mathbf{A}}_{\text{VAD}}(\text{sent}) = \min_{C_i \in \text{Clauses}} \left( \frac{ \sum_{t \in C_i,\ \text{pos}(t)\in\{\text{JJ,VB,RB}\}} \mathbf{A}_{\text{VAD}}(t) }{ N_{C_i,\{\text{JJ,VB,RB}\}} } \right).$$

Quantitatively, the lexicon assigns VAD scores in $\mathrm{acc@5} = 76\%$0. Sentence-level distributions are reported as roughly 28% negative, defined by valence $\mathrm{acc@5} = 76\%$1, and 50% positive, defined by valence $\mathrm{acc@5} = 76\%$2, per epoch with $\mathrm{acc@5} = 76\%$3; the remainder is near-neutral. Diachronic shifts are substantial: 7,885 words shifted from positive to negative, while 8,787 shifted from negative to positive. The 7,000 most variable words contribute to contextual changes in more than 90,000 sentences.

The paper provides concrete examples of long-range affective drift across epochs. Over the 1750s to 1950s, asylum changes from 0.27 to $\mathrm{acc@5} = 76\%$4, germs from 0.15 to $\mathrm{acc@5} = 76\%$5, homeless from 0.11 to $\mathrm{acc@5} = 76\%$6, and weird from 0.30 to $\mathrm{acc@5} = 76\%$7. In the opposite direction, febrile changes from $\mathrm{acc@5} = 76\%$8 to 0.33, infatuation from $\mathrm{acc@5} = 76\%$9 to 0.40, and destiny from $\mathrm{acc@7} = 79\%$0 to 0.44. The paper interprets these changes as reflecting semantic re-specialization, including medicalization, urbanization, and colloquial shifts.

In comparative perspective, CHRONOBERG is positioned against EEBO, Google Ngrams, COCA, COHA, CCOHA, TemporalWiki, and TiC-LM. The stated distinction is that EEBO and Google Ngrams are large-scale but lack sentence context and modern ML compatibility, COCA/COHA/CCOHA are valuable for American English but smaller in scale and temporal depth, and TemporalWiki/TiC-LM emphasize contemporary factual updates rather than centuries-long semantic analysis.

4. Temporal robustness, moderation drift, and continual learning

The corpus was used to test whether modern LLM-based tools can contextualize discriminatory language and sentiment historically, and whether sequentially trained LLMs can encode diachronic semantic shifts without catastrophic forgetting (Hegde et al., 26 Sep 2025). The results are consistently negative.

For discriminatory-language detection, nine tools were evaluated on HateCheck. Many are described as near chance. RoBERTa achieved the highest recall, Perspective API the highest precision, and OpenAI moderation was added for comparison. The paper then used a two-stage pipeline in which RoBERTa flags candidates and Perspective API filters them to reduce false positives.

Qualitative examples expose historical-context failures. The sentence “Black should never be worn at a wedding” has neutral valence, approximately 0.10, but is flagged as hateful. The expression “tory hussy” is misclassified because of modern connotations of hussy. The dataset also corroborates known diachronic shifts in words such as faggot and gay. Quantitatively, for sentences flagged as hateful by RoBERTa, the temporal VAD lexicons contradict hatefulness 59% of the time in early eras, 1750–1850s, reducing to approximately 50% in later eras. Disagreement between OpenAI moderation and RoBERTa is approximately 85%, which the paper presents as evidence of high volatility among modern tools when historical context is involved.

The sequential-training experiments use Pythia 1.4B with the gpt-neo-1.3B tokenizer on NVIDIA A100-80GB GPUs. Three training regimes are compared: Sequential Training (ST) over consecutive 50-year intervals; Single-Interval Baselines; and continual-learning baselines comprising Elastic Weight Consolidation (EWC) and Low-Rank Adaptation (LoRA). The reported configuration is 30 epochs, batch size 64, micro-batch 4, gradient accumulation 8, Adam with weight decay 0.1, $\mathrm{acc@7} = 79\%$1, $\mathrm{acc@7} = 79\%$2, learning rate $\mathrm{acc@7} = 79\%$3, cosine decay scheduler, 100 warmup steps, and sequence length 2048. LoRA uses $\mathrm{acc@7} = 79\%$4 and $\mathrm{acc@7} = 79\%$5.

Evaluation is by perplexity across eras on two test sets per interval: valence-stable words and valence-shifting words. Sequential Training exhibits forgetting on earlier intervals, mild for valence-stable material but severe for valence-shifting content. Forward generalization worsens as temporal jumps increase; from 1750 to 1950, perplexity increases by approximately 27%. On valence-shifting words, deterioration is described as substantially exacerbated.

EWC reduces catastrophic forgetting and keeps previously seen intervals closer to diagonal performance than ST, but it still struggles on valence-shifting words because it preserves past information rather than anticipating future semantic drift. An appendix summary reports average perplexity increase over time of 12% for EWC versus 34% for ST, and forward generalization of 29% for EWC versus 33% for ST. LoRA presents an intermediate trade-off, retaining knowledge nearly as well as EWC while providing more plasticity than EWC and less forgetting than ST. For valence-stable words, the initial interval sees only approximately a 15% rise at the end of training; for valence-shifting content, cross-temporal perplexity increases more sharply, for example a $\mathrm{acc@7} = 79\%$6 rise. Its appendix summary gives 15% average perplexity increase and 27% forward generalization.

The overall conclusion is explicit: present sequential and continual-learning approaches struggle to encode diachronic semantic drift, especially for terms with shifting affect. The paper therefore recommends temporally aware pipelines that integrate coarse temporal binning, aligned embeddings, volatility detection, era-aware objectives, and evaluation protocols sensitive to historical change.

5. General-relativistic CHRONOBERG: clocks, redshift, and stationary spacetimes

In relativistic geodesy, CHRONOBERG names a clock-based gravity field recovery concept formulated in Einstein’s general relativity on a stationary spacetime $\mathrm{acc@7} = 79\%$7 with metric signature $\mathrm{acc@7} = 79\%$8 and timelike Killing vector field $\mathrm{acc@7} = 79\%$9 (Philipp et al., 2023). The basic objects are standard clocks moving on timelike worldlines $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$0, with normalized four-velocity $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$1 satisfying $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$2, and light signals modeled as null geodesics $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$3 with tangent $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$4 satisfying $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$5.

The measured frequency of a light ray by an observer is defined as $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$6, with the equivalent sign convention $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$7 or $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$8. For emission at $$\mathcal{N}_k(w) = \operatorname{Top}k_{u\in \mathcal{V}_{\text{VAD}\setminus\{w\}}} s(e_w,e_u), \qquad \widehat{\mathbf{A}}_{\text{VAD}}(w) = \frac{1}{|\mathcal{N}_k(w)|} \sum_{u\in \mathcal{N}_k(w)} \mathbf{A}_{\text{VAD}}(u).$$9 and reception at $\mathbf{e}_w \in \mathbb{R}^d$0, the exact frequency transfer is

$\mathbf{e}_w \in \mathbb{R}^d$1

In a stationary spacetime with adapted coordinates $\mathbf{e}_w \in \mathbb{R}^d$2, the metric is written

$\mathbf{e}_w \in \mathbb{R}^d$3

with $\mathbf{e}_w \in \mathbb{R}^d$4, where $\mathbf{e}_w \in \mathbb{R}^d$5 is the redshift potential. For static observers aligned with the Killing field, the conserved $\mathbf{e}_w \in \mathbb{R}^d$6 yields the standard stationary result

$\mathbf{e}_w \in \mathbb{R}^d$7

with $\mathbf{e}_w \in \mathbb{R}^d$8. For two stationary observers, $\mathbf{e}_w \in \mathbb{R}^d$9 is constant along the link.

Proper time along a timelike worldline obeys

$w$0

where $w$1. The decomposition separates gravito-electric contributions through $w$2, kinematic terms through $w$3, and frame-dragging contributions through the cross-terms $w$4, or $w$5 in axisymmetry.

The paper also gives the weak-field, slow-motion expansion

$w$6

or equivalently

$w$7

This makes explicit the separation between Newtonian potential difference, transverse Doppler, longitudinal Doppler, and higher-order relativistic effects such as frame dragging and propagation corrections.

6. Multipoles, observation geometry, and gravity field recovery

The geodetic CHRONOBERG concept uses redshift measurements between clocks on ground and in space to determine the Earth’s gravitational potential and, more specifically, its multipolar structure (Philipp et al., 2023). In the Newtonian limit, the geopotential is represented by spherical-harmonic multipoles, while in the relativistic treatment the paper employs Thorne’s ACMC multipolar expansion for the stationary vacuum metric. The expansion is given in terms of mass multipoles $w$8 and spin multipoles $w$9, with the dominant Earth gravito-magnetic term being the spin dipole $s(e_w,e_u)$0. Frame dragging appears through $s(e_w,e_u)$1, and near Earth the spin-dipole contribution is

$s(e_w,e_u)$2

The observation geometry is expressed via the emitter–observer map and the time-transfer function $s(e_w,e_u)$3, whose derivatives provide the light-ray direction: $s(e_w,e_u)$4 and $s(e_w,e_u)$5. With these definitions, the exact redshift factorizes into a geometric part and a Doppler part,

$s(e_w,e_u)$6

where

$s(e_w,e_u)$7

This factorization is central to inversion: geometry carries $s(e_w,e_u)$8 and $s(e_w,e_u)$9, while the second factor carries longitudinal Doppler terms.

Several measurement configurations are treated explicitly. For two stationary ground clocks, $\mathcal{V}_{\text{VAD}}$0, so redshift directly constrains the gravito-electric potential and thus the mass multipoles. For a stationary ground clock and a moving satellite, the paper gives

$\mathcal{V}_{\text{VAD}}$1

and for a geodesic satellite, on the geoid timescale $\mathcal{V}_{\text{VAD}}$2,

$\mathcal{V}_{\text{VAD}}$3

For two geodesic clocks in space, the redshift includes the ratio of conserved energies and both geometric and Doppler factors. Two-way mirror tracking is also analyzed; in Schwarzschild radial free fall the expansion becomes $\mathcal{V}_{\text{VAD}}$4, which combines Doppler and GR corrections.

The inversion strategy is presented as a least-squares problem. One constructs an observation vector $\mathcal{V}_{\text{VAD}}$5 from measured redshifts and known kinematic quantities, builds a model vector $\mathcal{V}_{\text{VAD}}$6 from the multipolar $\mathcal{V}_{\text{VAD}}$7 and $\mathcal{V}_{\text{VAD}}$8, and solves $\mathcal{V}_{\text{VAD}}$9 over a global dataset of ground–ground, ground–space, and space–space links. The framework is described as compatible with existing gravity mission methods and with the Newtonian energy approach in the appropriate weak-field limit.

The paper also assesses feasibility. State-of-the-art optical clocks are reported at fractional frequency stability $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$0, corresponding to approximately 1 cm in geopotential height, with $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$1. Fiber links achieve stability at least an order of magnitude better than free-space links; free-space Earth–space comparisons currently realize stabilities around $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$2, often $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$3. ACES aims at $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$4 tests of the redshift, and Galileo eccentric orbits have reached $\mathbf{A}_{\text{VAD}}(u) \in \mathbb{R}^3$5. The intended role of CHRONOBERG is therefore as an additional data channel for advanced data fusion alongside satellite tracking, accelerometry, and gradiometry.

7. Limitations, ethical constraints, and future directions

The two CHRONOBERG programs are both explicitly bounded by methodological caveats, though the nature of those caveats differs sharply by field.

For the corpus, the paper includes a disclaimer that historical language samples may be offensive (Hegde et al., 26 Sep 2025). It further states that VAD lexicons are not moderation tools and should not be treated as definitive labels of harmfulness. Harmful language is described as subjective and contextual, and contemporary classifiers are shown to misinterpret historical content. Technical limitations include publication-year inference uncertainty, with MAE on the order of 3–5 years; genre and era bias, especially heavy pre-1920 representation and strong fiction presence; dependence on synchronic NRC VAD ratings for historically calibrated induction; and practical preprocessing constraints, including removal of non-alphanumeric characters in the processed split. The stated future directions are decade-level analyses tied to historical eras and literary movements, broader continual and lifelong learning, machine unlearning protocols for historically contingent slurs or outdated facts, and extension to additional languages and longer time ranges as copyright windows open.

For chronometric geodesy, the main limitations arise from modeling assumptions rather than annotation uncertainty (Philipp et al., 2023). The framework is exact but assumes stationarity and vacuum in the sense of slowly varying mean fields; tides, atmosphere, hydrology, and non-gravitational accelerations must be added a posteriori in adiabatic or perturbative form. The emitter–observer problem must be solved numerically, the time-transfer function must encode propagation effects such as Shapiro delay and bending, and unique null geodesics are assumed. A plausible implication is that practical deployment depends as much on orbit determination, link modeling, and multi-sensor fusion as on clock performance alone.

Taken together, the two literatures use CHRONOBERG to formalize temporal dependence in domains where static treatments are inadequate. In one case, temporal structure is a property of linguistic data and semantic drift; in the other, it is a property of relativistic signal propagation and gravitational inference. The name thus marks two independent efforts to make temporal context an explicit computational and observational variable.