A corpus of Hungarian business-lunch menus from Kiskunfélegyháza, with a pipeline that turns scraped HTML into cleaned, lemmatized, categorized, and seasonally-analyzed dish data.

→ Explore the interactive visualization — a monochrome, searchable dashboard of the corpus (index.html).

1. Scrape — 0_code/hol_menuzzek.py collects daily menu HTML into 1_html/. Runs weekly via GitHub Actions.
2. HTML → text — 0_code/html_to_text.sh converts to 2_txt/.
3. Extract menus — menu_extractor.py produces daily_menus.csv.
4. Clean dishes — clean_dishes.py deduplicates and normalizes into cleaned_dishes.csv.
5. Lemmatize — process_cleaned_dishes_spacy.py (HuSpaCy) → dishes_with_lemmas.csv; create_lemma_dictionary.py → lemma_dictionary.json.
6. Categorize — categorize_dishes.py tags course, protein, and cooking method → categorized_dishes.csv.
7. Resolve synonyms — resolve_synonyms.py collapses variants using 0_code/synonym_map.json → dishes_resolved.csv.
8. Seasonal analysis — seasonal_analysis.py produces monthly counts and seasonal summaries.

make all       # run steps 2–8
make clean     # remove generated artifacts

Individual steps: make text | menus | dishes | lemmas | categorize | synonyms | seasonal.

Dependencies in requirements.txt.

Corpus: ~51,200 menu rows, ~5,050 unique dishes after resolution, Nov 2022 → Apr 2026.

1. Courses (unique dishes) — főétel 5,994 (72%) · leves 1,688 (20%) · főzelék 388 (5%) · desszert 188 (2%).
2. Proteins (unique dishes) — csirke 2,536 · sertés 2,218 · vegetáriánus 833 · marha 463 · hal 306.
3. Cooking methods (unique dishes) — rántott 1,400 · sült 1,109 · párolt 503 · töltött 491 · pörkölt 357.
4. Top dishes overall (resolved occurrences) — gyümölcsleves 2,268 · hasábburgonya 1,232 · roston csirkemell 995 · rizibizi 938 · burgonyapüré 886.
5. Top soups — gyümölcsleves 2,268 · csontleves 642 · húsleves 593 · zöldségleves 467 · lebbencsleves 269.
6. Top főzelék — gyümölcsfőzelék párizsi csirkemellel 24 · meggyfőzelék párizsi csirkével 21 · burgonyafőzelék vagdalttal 20 · sárgaborsó főzelék borsos tokánnyal 19 · meggyfőzelék párizsi csirkemellel 18.
7. Winter-peaking dishes (peak Dec–Feb) — rántott sertésszelet (75, idx 0.76) · töltött káposzta (65, 0.61) · paprikás krumpli (66, 0.53) · tarhonyás hús (64, 0.49) · paradicsomos húsgombóc (88, 0.43).
8. Spring-peaking (peak Mar–May) — sertés tokány galuskával (62, 0.58) · carbonara spagetti (60, 0.55) · magyaros burgonyaleves (63, 0.46) · rakott kelkáposzta (53, 0.41) · szilvás gombóc (51, 0.40).
9. Summer-peaking (peak Jun–Aug) — fahéjas szilvaleves (60, 0.42) · mákos tészta (51, 0.35) · zöldborsóleves (131, 0.25) · hentes tokány (56, 0.24) · lasagne (114, 0.21).
10. Autumn-peaking (peak Sep–Nov) — krumplis tészta (54, 0.72) · hamis gulyásleves (51, 0.64) · korhelyleves (50, 0.58) · rakott zöldbab (69, 0.51) · brassói aprópecsenye (80, 0.40).

Seasonal lists rank by seasonality index (peak-month concentration) among dishes with ≥50 occurrences, so they surface genuinely seasonal items rather than year-round staples. Top-soup and top-főzelék lists skip a few entries the categorizer mislabels by course (e.g. galuska tagged as a soup) — see Data notes.

Three tiers based on natural breaks in the frequency distribution. Counts are total occurrences across the corpus (Nov 2022 → Apr 2026).

20 dishes, ~12,750 occurrences combined. The menu backbone — appear constantly across restaurants and seasons. Mains (7,695) lead soups (4,452); chicken (5,071) edges pork (3,629) and vegetarian (3,189); frying dominates (rántott 5,900).

53 dishes, ~4,663 occurrences. Familiar dishes that rotate in and out of weekly menus.

Profile: main courses (2,351) and soups (2,044) are roughly balanced — soups carry more weight here than in Tier 1. Vegetarian dishes (826) — főzelék, rakott, pasta — take the largest protein share, ahead of beef (494), pork (393) and chicken (213). The dominant cooking methods shift away from frying: rakott (366) leads, followed by pörkölt (300), sült (226), and töltött (129). Characteristic dish types: stews (gulyás-, csirke-, sertéspörkölt), layered casseroles (rakott karfiol, székelykáposzta), thicker soups (tarhonya-, frankfurti, brokkoli krém), and a handful of comfort-food desserts (máglyarakás).

248 dishes, ~4,798 occurrences. Variety items that keep menus from feeling repetitive.

Profile: mains (2,563) and soups (1,749) again split most of the weight, but főzelék (287) appears as a real category — almost absent in Tier 1. Protein distribution flattens: vegetarian 537, chicken 509, pork 452, beef 313 — no single protein dominates. Methods diversify further: pörkölt (510) leads, joined by rakott (148) and párizsi-style (148) — techniques rare in higher tiers. Characteristic dish types: regional and old-style stews (hentes/sertés/borsos/csikós tokány, hamis gulyás, korhelyleves), heritage dishes (erdélyi rakott káposzta, krumplis tészta, szilvás gombóc), and lighter brothy soups (lencse-, csirkeaprólék, húsgombóc). This is where the menus pick up character and local variation.

4,726 long-tail dishes appearing 1–9 times each (~9,277 occurrences). Daily specials, one-offs, and seasonal experiments — not part of the hierarchy.

• Coverage. The scraper (GitHub Actions, weekly) collects HTML into 1_html/; the analysis pipeline (steps 2–8) is run by hand. This refresh reprocessed the full corpus through 2026-04-05 — the previous README froze at June 2025 because the pipeline hadn't been re-run, even though scraping continued.
• Menu boilerplate. The newer site layout interleaves price lines, section headers (Levesek:, Főételek:), greetings, and season's-greetings into the menu blocks. menu_extractor.py captured these as dish rows, and previously they survived into the results (~40 raw strings, ~1,600 occurrences — Kínálatunkat itt találja alone appeared 606×). clean_dishes.py's is_noise_line() now filters them (price tokens, trailing-colon labels, greeting/announcement and section-header phrases), so they no longer reach the dish data.
• Casing. Some canonical names retain their original capitalization where synonym/lemma resolution didn't map them to a lowercase form (curated synonyms in synonym_map.json are lowercase; auto-grouped and unmatched dishes keep their cleaned, sentence-cased form). Display names above are lowercased for consistency; this introduces no count collisions (verified) and is cosmetic — it does not affect any count.