{{TITLE}} · corpus-lens dashboard

Corpus map · nonlinear embedding of the corpus

TF-IDF vectors (top-30 distinctive keywords per paper, union vocabulary, L2 normalised) projected to 2D. Each dot is a paper; nearby dots use similar vocabulary. PaCMAP preserves both local and global structure; UMAP biases toward local structure; PCA is the linear baseline.

method

colour by

size by

show labels

Click any dot to open the per-paper view. The clusters that emerge here are empirical – they don't know about programme assignments. When a programme's colour clusters tightly, the editorial decomposition matches the vocabulary structure; where colours interleave, the vocabulary cuts the corpus differently than the programmes do.

Programme × rhetorical dimension heatmap

Mean density (per 1k tokens) of each rhetorical marker within each programme. Cell colour intensity is the value; cell text shows the number. Hover for the per-cell sample size; click a programme name to filter the per-paper view.

Distinctive vocabulary fingerprint per programme

Top 8 most-distinctive (df ≤ 3) keywords aggregated within each programme. These are the words the editorial decomposition agrees should anchor each programme; they are what would fingerprint any new paper into that programme.

Method

For each paper, we extract plain text from paper.html (dropping <script>, <style>, <pre>, <code>, the auto-injected enhancement panels, and the References section). Then:

TF-IDF keywords – top-20 distinctive terms per paper using log((N+1)/(df+1)) + 1 for IDF.
Distinctive keywords – subset of the above with df ≤ 3 (used by ≤ 3 papers total).
Top bigrams – two-word phrases ranked by raw frequency, stopwords removed.
Rhetorical fingerprint – per-1000-word density of: hedging (may, might, could, presumably, plausibly, suggests, …), assertion (proves, establishes, demonstrates, definitively, …), empirical (experiment, dataset, observe, baseline, …), theoretical (theorem, lemma, proof, axiom, corollary, …), critical (refutes, contradicts, fails, mistaken, …), and quantitative (numbers with units).
Paper similarity – cosine of TF-IDF top-30 keyword vectors. Top-5 most-similar papers per paper.
Inverted keyword index – for the top 2000 corpus terms, which papers use each one and how much.

Source: tools/corpus_analyze.py. Stdlib-only; no sklearn or external models. Run on demand; output in assets/corpus_analysis.json · corpus_keyword_index.json · corpus_similarity.json.