Paper Review — verify claims with raw data

Concise verdict

The paper Computer modelling of innovations relative to Latin in contemporary Romance dialects encodes 145 innovations across 61 Western Romance survey points and applies historical glottometry plus clustering (Ward), MDS and t-SNE, with Random Forest feature importance (averaged 10,000 runs), to recover a robust North/South split and three regional innovation waves; the dataset and maps (Romance_innovations, HGOA, Atlas LIMSI) are provided for reproducibility and interpretation Paper Main Study

Detailed review and critique

This review analyses methods, data, results, limitations, and opportunities for follow-up work in Computer modelling of innovations relative to Latin in contemporary Romance dialects, with verbatim evidence from the paper cited inline below.

1) Data and annotation

• The authors encoded 145 binary innovations (presence 1 / absence 0) across 61 survey points sampled from the Boula de Mareüil speaking atlas; the innovations are phonetic (96), morphological (27), syntactic (8), lexical (14) and were derived from parallel fable translations plus a 120-word list Innovations and counts
• Sampling choices: 61 points favour 'average' varieties; distribution: France 21, Belgium 4, Switzerland 4, Spain 6, Portugal 1, Italy 25 — authors acknowledge Eurocentric and Western Romance focus and underrepresentation of Balkan Romance Sampling distribution and caveats
• Annotation strategy: binary encoding simplifies variation and requires human consensus on thresholds (authors met to agree); the paper documents context restrictions (e.g., word-initial, coda) to limit annotation ambiguity Annotation protocol

2) Algorithms and computational pipeline

• Analyses used: historical glottometry (wave model), agglomerative hierarchical clustering (Ward), multidimensional scaling (MDS), t-SNE, Random Forests for feature importance; Manhattan distance as pairwise distance metric for MDS/t-SNE; visualizations via Plotly and GeoJSON choropleths Algorithms and software
• Feature selection: authors used Gini importance from Random Forests and averaged importance across 10,000 runs to mitigate instability from collinearity; they avoided RFE due to correlated features Feature selection approach

3) Key results reported

• Robust North/South divide recovered by multiple methods; Oïl (northern Gallo-Romance) area is the most innovative; three prominent innovation centres: Oïl, Ibero-Romance Spain, and southern Italy (authors show maps and dendrograms) Main geographic results
• Feature importance: palatalisation of Latin CA and geminate simplification north of La Spezia-Rimini line are among top discriminators; morphosyntactic signals (non-null subjects, auxiliary selection) also discriminate but are fewer in number Feature importance
• Sardinian and Romansh behave as outliers: Sardinian is conservative and often separates; t-SNE isolates Sardinian distinctly — authors caution against literal tree-chronologies and emphasize wave-like contact history Sardinian and t-SNE

4) Strengths

• Integration of classical historical-comparative practice (distinguishing innovations vs retentions) into computational analyses addresses a known limitation of pure edit-distance methods Comparative method advantage
• Transparency and reproducibility: authors publish the full list of innovations (Romance_innovations) and link to historical glottometry resources (HGOA) and the Atlas (LIMSI) for replication and further extension Data availability

5) Limitations, possible biases, and blindspots

• Sampling bias and scope: the corpus is deliberately compact and Western-Europe centered; omission of Balkan Romance and overseas varieties reduces generality — the authors note this limitation and propose future scaling Sampling caveat
• Binary coding simplification: reducing gradient or variable innovations to 0/1 sacrifices frequency/probabilistic information (authors deliberately chose this tradeoff for parsimony) Binary coding tradeoff
• Feature weighting absent: authors did not assign perceptual or sociolinguistic weights to innovations (they discuss why and suggest that weighting requires empirical measures of salience) Feature weighting not attempted
• Algorithmic sensitivity: t-SNE is noted by authors to be locally informative but globally less faithful; Random Forest importance rankings vary with seeds despite averaging (authors acknowledge seed-dependence for the top features) Algorithmic sensitivity
• Reproducibility gap: although feature list and some maps are online, the authors do not publish the complete analysis scripts or exact random seeds; the paper gives methods and tools (Scikit-learn, SciPy, Plotly) but full code deposition would increase reproducibility further Code availability notes

6) Suggested methodological improvements and follow-ups

1. Expand sample: add more survey points, especially Balkan Romance and under-sampled Iberian, and test whether the North/South split and innovation-wave centres persist (falsifiability test authors propose)
2. Introduce weighted encodings: measure perceptual salience (speaker judgements or corpus frequency) to weight innovations before clustering and re-run feature selection
3. Release code and seeds: publish an analysis repository (Jupyter notebooks) that reproduces the Random Forest 10,000-run averaging, dendrogram pruning thresholds, and MDS/t-SNE parameters
4. Model continuous traits: for variable phonetic innovations, record proportion-of-occurrence to allow distance metrics incorporating continuous features (e.g., Bray-Curtis, Mahalanobis after PCA) rather than strict binary Manhattan
5. Compare models: include ED/acoustic-distance baselines and Bayesian contact models (e.g., sBayes-like approaches) to separate contact/inheritance/universal effects and quantify their contributions Future directions and comparisons

7) Reproducible figure: innovation counts by type and survey point (interactive)

8) Conclusion and assessment

The paper provides a rigorous, historically informed computational treatment of Romance dialect innovation patterns. The main claims (North/South divide, Oïl innovativeness, Sardinian distinctiveness, important phonetic discriminators) are well supported by the annotated dataset and the applied methods; the principal weaknesses are scope (Western Europe only), simplification of gradient phenomena to binary traits, and absence of published analysis code and exact seeds. The study is nevertheless a strong, reproducible‐oriented proof of concept that bridges comparative historical linguistics and modern data science.

If you want a fully reproducible re-run of their analyses with expanded sampling or alternative metrics (weighted features, continuous trait encodings, Bayesian contact models), you can launch a BGPT AI agent to execute that end-to-end:

Author reviews

Key sources and full-paper evidence are cited inline throughout this review. For raw data and maps see the project's online resources: Romance_innovations and HGOA (linked within the cited paper).