Large language models typically train on social media and web crawls, which offer scale but little long-term temporal depth. The new Chronoberg corpus directly addresses this gap by assembling a 250-year diachronic dataset, allowing researchers to study how language and sentiment evolve across centuries rather than just years.
The corpus pairs its historical texts with temporal affective lexicons—word-sentiment mappings that shift over time. This lets analysts trace not only word usage but also how emotional connotations change, which is critical for interpreting older texts with modern models.
Because the source is a single arXiv preprint, the details are limited to the abstract. The authors argue that existing corpora lack the temporal structure needed for certain analyses, and Chronoberg is positioned as a remedy. No comparisons to other diachronic resources are given in the available text, so its relative advantages beyond the stated design remain to be seen.