Gutenberg+: A More Temporally Faithful Corpus for Diachronic NLP
Proceedings of the Workshop on Structured Linguistic Data and Evaluation (SLiDE)
Abstract
We introduce Gutenberg+, a temporally more faithful version of the Project Gutenberg (PG) corpus, one of the most widely used resources for diachronic text analysis. Despite its popularity, the PG corpus contains a major yet overlooked flaw: around 15% of its entries are collections (e.g., anthologies of books, letters, or poems) rather than atomic works, which distorts temporal analyses since such collections may span multiple decades. We present an automatic method to detect and split these collections into their constituent works, producing a finer-grained and temporally consistent corpus. We further re-annotate publication years using LLM-based retrieval-augmented generative methods, demonstrating the potential of LLMs to enhance structured linguistic resources. To illustrate the utility of Gutenberg+, we conduct a small-scale diachronic case study on negation, showing that our refined corpus captures more nuanced cross-linguistic variation than the original PG data. Finally, we release the corpus in UIMA format with full metadata and linguistic annotations, providing a standardized resource for future research on diachronic language change.