📄 Methods, Data, and Conceptual Change: Reflections from Two Quantitative Diachronic Case Studies

Summary

The paper asks how the structure of a historical corpus limits which kinds of meaning change a quantitative method can detect. The authors do not propose a new technique. They set two existing approaches side by side.

• Quad-based concept modelling. This method comes from the Linguistic DNA project. It runs over EEBO-TCP, a huge, genre-mixed and orthographically unstable collection of Early Modern English print. After spelling normalisation and lemmatisation, it finds sets of four strongly associated lemmas inside a wide text window. These sets are then grouped into "constellations" around a focal word.
• SynFlow. This dependency-based method is applied to a small slice of the Royal Society Corpus 6.0.4, covering late-eighteenth-century chemistry writing, which the authors re-parsed. It tracks which words fill the syntactic slots around a concept.

Each method reveals a different kind of change:

• The window approach shows divergence at the level of discourse, such as liberty versus freedom.
• The slot approach shows relational shifts, such as air gaining chemically specific modifiers.

The authors conclude that what counts as observed change is produced jointly by method and dataset. Matching the method to the data, and documenting preprocessing openly, are therefore essential.

Key points

• EEBO-TCP holds roughly 765M words (c.1470s–1690s) of religious, political, pamphlet and literary texts. Quads are four co-occurring lemmas found within a moving window, usually 100 words.
• The RSC subset covers 1750–1799 and three topics (Biochemistry, Chemistry 1, Chemistry 2, taken from Menzel et al., 2021). It contains 913 texts, 351,849 tokens and a vocabulary of 73,628. It was drawn from the ~78.6M-token open release.
• The RSC subset was re-parsed with Stanza. OCR noise (e.g. "head – ach"), historical spellings (e.g. "inconveniency") and leftover preprocessing errors caused wrong lemmas and wrong or unrecognised POS tags (tagged X).
• The near-synonyms liberty and freedom often co-occur, yet their quad constellations point to distinct discourse environments.
• In SynFlow, the reconceptualisation of air appears mostly in the adjective slot, which moves from generic modifiers to chemically specific ones.

Interesting ideas

• I1. The type of context used (wide-window association vs. dependency relations) should follow how reliably that structure can be recovered from a given dataset, not analyst taste.
• I2. A concept can be profiled by the fillers of its syntactic slots, such as its adjective modifiers, and change shows up as those fillers shift.
• I3. Words that look interchangeable lexically can be separated by their association networks rather than by their frequency.
• I4. Spelling normalisation plus lemmatisation is a precondition for counting associations at all, not just cleanup.
• I5. Even a "latest, optimised" corpus release needs hands-on inspection of parser output, and every transformation should be documented for reuse.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tree): Estimate parse reliability per genre or source (e.g. the X-tag rate), and fall back to non-syntactic evidence where the trees are untrustworthy.
• I2 (world): Compile per-noun slot-filler profiles from parsed text into the ontology, giving an explainable picture of how a corpus characterises an entity type.
• I3 (world): Store near-synonyms as distinct concepts in the knowledge graph, with the contexts that distinguish them attached.
• I4 (tokens): Add a variant-spelling and OCR-split normalisation map ahead of the lemmatiser so that older fables and tales reach the same lemmas as modern text.
• I5 (learn): Routinely audit silver data for unknown-tag and bad-lemma rates as a negative-control check, and log each preprocessing step.