📄 Machines in the media: semantic change in the lexicon of mechanization in 19th-century British newspapers

Summary

The paper asks whether computational tools can follow the meaning shifts that 19th-century industrialisation caused in English words about machines and transport, and whether these tools agree with traditional philological accounts. The authors combined two digitised British newspaper collections (about 4.6 billion tokens, 1801–1920) and divided them into decades. They trained a separate Word2Vec model for each decade and aligned the resulting spaces with Orthogonal Procrustes. For each word, they measured the cosine similarity between every decade's vector and the 1910s vector. They then ran the PELT changepoint algorithm on these similarity curves. Detected shifts were checked against nearest neighbours, Görlach's (1999) history of 19th-century English and OED sense dates. Most detections agreed with the expert literature. For coach, the models picked up the railway-carriage sense in the same decade as its first OED attestation, and the data suggest this use was not only American. The study shows that large OCR'd newspaper archives can support semantic-change research in digital humanities when validation is done carefully.

Key points

• Data: Heritage Made Digital (12 titles, about 2.3B tokens) plus Living with Machines newspapers (107 titles, about 2.3B tokens). Text was lowercased, punctuation and OCR line breaks were removed, and NLTK stop words were dropped.
• Model selection: a grid search on the 1821–1830 slice scored candidate settings by mean cosine similarity of five OED/Historical Thesaurus synonym pairs. The winning configuration was skip-gram, 5 epochs, 200 dimensions, window 3 and min count 1.
• Changepoint detection: PELT from the ruptures library, run with jump 1 and penalties of 0.5 and 0.25.
  • At penalty 0.5, eight words were flagged: coach, gear, traffic and train (1830), stamp (1840), fellow and railway (1860), and matches (1880).
  • At penalty 0.25, wheel was also flagged (1810, 1880), and traffic gained a second changepoint in 1860.
• Mismatches with the literature:
  • Fellow and wheel, which Görlach calls stable, were still flagged.
  • Bulb got no changepoint. The authors attribute this to its tangled polysemy: it already had senses in anatomy, botany and electricity before 1800.
• Negative control: the words from the synonym pairs, which the OED treats as stable, produced no changepoints at penalty 0.5, apart from identical.

Interesting ideas

• I1. Testing the detector on words known to be stable (here, the synonym-pair vocabulary) shows how often it raises false alarms, and that rate is what makes its positive findings believable.
• I2. Instead of picking hyperparameters with an abstract loss, they used a few hand-checked synonym pairs with known relations as a cheap intrinsic benchmark.
• I3. Every decade is compared with one fixed reference period (the latest) rather than only with its neighbour, which turns change into a single curve that a changepoint algorithm can segment.
• I4. A numeric detection is only accepted as meaningful after inspecting nearest neighbours and matching them to dated dictionary senses, so every flagged change comes with a human-readable reason.
• I5. The authors note that heavily polysemous words with many pre-existing senses blur a single-vector signal, which is why bulb showed no clear changepoint.

For Mova

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

• I1 (learn): Any automatic signal Mova uses, such as witness-parser disagreement or a repair-rule trigger, could be run on a held-out set where no change or error is expected, so its false-positive rate is reported alongside its hits.
• I2 (learn): A small curated list of pairs known to share or lack a VerbNet class or WordNet relation could serve as a fast sanity check when choosing settings for lexicon-induction or silver-data steps.
• I3 (world): Tracking an entity's state against a fixed anchor point in the story, not only step to step, could help locate the exact event where a holding or attribute shifted.
• I4 (world): When Mova's compiled lexicon assigns or changes a word sense, linking that decision to concrete WordNet/FrameNet evidence would keep the step explainable and auditable.
• I5 (frames): Sense-sensitive role assignment should flag highly polysemous predicates for frame-level disambiguation instead of trusting a single averaged representation.