Chronoberg: A 250-Year Diachronic Corpus with Temporal Affective Lexicons for Language Model Analysis
Niharika Hegde, Subarnaduti Paul, Lars-Joel Frey, Manuel Brack, Kristian Kersting, Martin Mundt, Patrick Schramowski
cs.CL, cs.AI
2025-09-26
Chronoberg is 2.7B tokens of English books from 1750-2000, with era VAD scores for 335,804 words. Swapping in a period synonym flips 50-75% of hate labels in 3 of 4 LLMs.
Most LLM training text sits close to the present. Common Crawl and Wikipedia show usage drift over months or a few years, not how a word's charge moves across two centuries. Change is partly a shift in what a word points at. It is also pejoration, a slide toward a worse evaluation, and amelioration, a slide toward a better one. A hate classifier that memorized today's lexicon will reread an old book with today's offenses. The paper's opening example is a line about a woman sent to strew the flowers, which a modern reader loads with connotations the sentence did not carry.
Each existing resource misses a piece. Google Books and EEBO n-grams cover long spans and throw away the sentence. COHA, COCA, and CCOHA are carefully annotated and not freely usable at scale. TiC-LM and TemporalWiki target short-horizon factual drift. The NRC VAD lexicon scores roughly 45,000 contemporary English words on valence, arousal, and dominance, and those scores are a snapshot of now. Chronoberg is an open, year-aligned English book corpus whose affect scores move with time.
The source is Project Gutenberg, about 73,500 books. Catalog dates are often digitization dates. Publication year is inferred mainly from OpenLibrary, and author lifespan is used to drop posthumous editions. On 100 hand-checked books from 1611 to 1912, OpenLibrary hits the exact year for 49. Among books where a year comes back, it lands within five years 76% of the time, with a mean absolute error of 3.05 years and a standard deviation of 5.20. Bins narrower than 15 years are a bad idea. After undated, anonymous, non-English, and translated works are removed, 25,061 books remain, spanning 1750 to 2000: 2.7 billion tokens and 91 million sentences. Counts rise through the 1920s. Copyright then cuts the supply, because most books published after 1929 are still protected. The five 50-year windows get thinner toward the present, and their edges are not literary periods.
Each window gets its own Word2Vec model, trained for 10 epochs with a window of 5 and 300 dimensions. CADE aligns those spaces to a compass model of the full corpus, so nearest neighbors can be compared across time. A word is scored by averaging NRC VAD values of about 20 neighbors. Finding 20 that the modern lexicon actually lists usually means scanning the top 100 nearest vectors. Ten neighbors bias the score. Five hundred wash it toward neutral. The lexicon covers 335,804 words, each on three axes from -1 to 1.
The anchor is a modern human rating. Historical text only chooses the neighbors. Asylum moves from +0.27 to -0.65 as its company shifts from sanctuary to the mental hospital. Febrile moves from -0.58 to +0.33 as metaphor crowds out the clinical sense of danger. A modern thermometer held against an old map of neighbors tracks who stands nearby. It does not record how a reader in 1750 felt. Hand-rating a third of a million words backward across 250 years is not available, so the numbers stay trajectories.
Sentence scores skip a flat average, which neutral function words pull toward zero. Only adjectives, verbs, and adverbs count. Each clause is averaged, and the sentence takes the minimum, so a polite second clause cannot wash out an insult. In Figure 2 that pattern scores -0.005 if every word is averaged and -0.53 if the minimum clause is kept. The corpus then leans negative: in each window about 28% of sentences fall below -0.05 and about 50% sit above +0.05.
A word counts as shifting when the first-to-last valence gap exceeds 0.1 and the positive or negative class changes. High confidence needs two successive steps past the noise floor and a net change above 0.2. Medium is a single step, or a smaller net move. Low is drift that never clears the per-step floor. Those tiers hold 44,049, 36,301, and 2,518 words, about a quarter of the lexicon.
Of 119 words with a shift already documented in the linguistics literature, 12 fall outside the measurement window. The 40 in-scope evaluative items break down as 30 pejorations, 9 ameliorations, and 1 contronym.
| Tier | Words | Direction matches the literature |
| High | 14 | 12 (86%) |
| Medium | 21 | 11 (52%) |
| Low | 5 | 0 |
| All | 40 | 23 |
Eight of nine high-tier pejorations match, including awful and bastard. Eight of nine ameliorations match. Of the 17 misses, 11 run the wrong way because a dominant book sense pulls neighborhood valence upward. Bully and egregious pejorated mainly in speech. Five words, including fairy, queen, and bent, are too rare as insults in edited books to move the score. Sanction cancels itself, one sense against the other.
Non-evaluative changes, such as metaphor, narrowing, and generalization, should leave valence flat. Only 17 of 67 do. Among 24 words SemEval-2020 called stable, 9 stay flat here and 15 move. SemEval compares two time points. Chronoberg spans 250 years, so some of that motion can be real long-horizon drift. High-tier evaluative change is the part that holds. Medium tier, at 52%, is close to a coin flip.
Valence is a diagnostic of temporal mismatch. The paper is explicit that it does not stand in for hate. On HateCheck, RoBERTa reaches precision 0.963 and recall 0.973, with many false alarms. Perspective API reaches precision 0.993 and recall 0.389, and a threshold of 0.7 is there to strip false positives. RoBERTa casts a wide net, Perspective filters it, and the OpenAI Moderation API is an extra baseline. Table 3 is a set of examples, not an error rate. A line saying black should never be worn at a wedding is flagged as hate at valence +0.10. In an 1850s sentence, faggot still means a bundle of sticks. Current models treat the modern slur, while the temporal VAD score stays positive.
Sentences where RoBERTa and the VAD score already disagree were then given to Mistral-8x7B, Qwen2.5-72B, Deepseek-R1-70B, and GPT-OSS-120B. Replacing the target with [MASK], or with a non-toxic neighbor from that era's Word2Vec space, cuts the disagreement. The figure marks drops from about 9% and 13% up to 74%. The text summarizes three of the four models as a 50 to 75% drop, with many hate labels flipping to non-hate. The VAD score does not move. The classifier is following the surface string.
For anyone moderating public-domain books or archives, a period synonym is a direct probe. If labels flip in bulk on the same sentence, the system is still tied to the modern string. Sentences, affect scores, and code are public. Against n-gram archives, the gain is the sentence. Against COHA-style corpora, the gain is 2.7 billion tokens that do not require a license and can be grouped by year. Continual pretraining, lifelong learning, and unlearning are named as later uses. None of those runs is in the paper.
Neighbor averaging is not a new embedding method. What remains is the resource, and a controlled negative result: on three models, half to three quarters of the conflict sentences flip after one lexical swap. The scores are a side signal. Using them as hate labels gives them the wrong job.
The limitations section states four bounds. Date error is a mean absolute error of 3.05 years and a standard deviation of 5.20, so bins should be at least 15 years wide. The VAD numbers are trajectories, and the tiers are indicative. Edited books omit the spoken and subcultural registers where pejoration usually starts, which is why fairy and queen fail. Hate is subjective across eras, a binary label is too coarse, and the lexicon belongs next to a moderator, not in its place. A separate use constraint: do not train a modern classifier on these hate tags, or historical prejudice gets written into the new model. This is print English by prolific authors from 1750 to 2000, and it does not travel past that.
Several holes stay open. The 50 to 75% figure uses only sentences where RoBERTa already disagreed with VAD, so the corpus-wide false-positive rate is unmeasured. Taking the minimum clause pushes text negative, and disagreement with that score is then read as the model overreacting. Circularity, modern ratings painted onto old neighbors, is checked only indirectly by 12 of 14 high-tier words. That sample is small, and medium tier is near chance. Contextual diachronic embeddings are not in the comparison, so the expectation that a Transformer would leave the VAD reading unchanged is untested. Fifty-year bins average away turns inside the bin. Copyright thins the window after 1920, and Table 3 never reports agreement by decade.