Introduction
While massive datasets for historical prose exist, poetry has long been a "small data" field. This represents a significant gap in linguistics, as poetic language—dense, novel, and succinct—serves as an ideal laboratory for studying semantic change. This research addresses the scarcity of large-scale diachronic poetry corpora and tests whether established laws of semantic change hold true in highly stylized verse.
The Deutsches Lyrik Korpus (DLK)
The study introduces the Deutsches Lyrik Korpus (DLK) v1, the largest poetry corpus of its kind. Comprising approximately 75,000 poems and over 11 million tokens, it spans from 1575 (Baroque) to 1936 (Modernism). The corpus includes 269 authors and was compiled from sources like Textgrid and the German Text Archive (DTA).
Methodology: Diachronic Embeddings and Trope Detection
The researchers utilized a joint learning approach for diachronic word embeddings. By representing each word's vector as a main embedding plus a time-specific offset, the model shares information across centuries while capturing era-specific nuances.
To identify shifts in style and linguistic innovation, the team performed a self-similarity analysis, measuring how much a word's meaning resembles itself across 50-year windows. Additionally, they employed Principal Component Analysis (PCA) on the cosine similarity trajectories of word pairs to detect emerging and declining poetic tropes.
Results: Literary Seismography
The analysis revealed that the "law of linear semantic change" applies to poetry: the further apart two time periods are, the more word meanings diverge linearly.
A key finding was the "Pairwise Self-Similarity" plot, which acts as a seismograph for literary history. A sharp dip in similarity occurs around 1750, aligning with the onset of the Romantic period—a time of significant linguistic upheaval and semantic diversification in German poetry.
The Evolution of Poetic Tropes
The study successfully mapped the life cycles of specific metaphors. While some associations like "love" (Liebe) and "fidelity" (Treue) remained stable across four centuries, others showed distinct trajectories:
- Rising Tropes: The Romantic era saw a spike in associations between "love" and concepts like "magic" (Zauber) or "freshness" (frische).
- Falling Tropes: Associations like "cheap love" (using billigkeit, which historically meant "equity" but shifted toward "low price") vanished as word meanings changed in broader society.
Quantitative validation showed that the first two PCA components explained 86% of the variance in word-pair trajectories, suggesting that tropes follow collective historical trends rather than random shifts.
Conclusion
This research demonstrates that computational methods can effectively reconstruct literary history and track the evolution of complex tropes. Despite challenges with data sparsity in the "long tail" of rare vocabulary, the DLK provides a robust foundation for future diachronic linguistic studies in poetry.
Summary updated