| """ |
| arzuhal_retrieval.py — Vatandas sorusu -> kanun maddesi retrieval benchmark'i (Pusula modulleri uzerine). |
| |
| Kosullar (ayni qid, farkli sorgu metni): |
| clean : gundelik soru (vatandasin kendi cumlesi) — ANA kosul |
| resmi : ayni sorunun hukuki terimlerle hali (madde gorulmeden uretildi); yoksa gundelik kullanilir (degismemis sayilir) |
| ascii_full, ascii_partial, upper, typo, colloquial, mobile : Pusula bozulmalari, gundelik soruya uygulanir |
| Etiket kumeleri: ana = E+K (etiket kontrolunden gecen), kati = yalniz E. |
| """ |
| from types import SimpleNamespace |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| from pusula_eval import METRICS, paired_bootstrap |
|
|
|
|
| def doc_text(row): |
| head = f"{row['law_title']} madde {row['article']}" |
| title = (row.get("title") or "").strip() |
| return f"{head} — {title}. {row['text']}" if title else f"{head}. {row['text']}" |
|
|
|
|
| def build_corpus(arts): |
| """Mulga olmayan tum hedef kanun maddeleri (gecici/ek dahil). Mulga maddeler gercek bir aramada da cikmaz.""" |
| a = arts[~arts["repealed"]] |
| return {r["doc_id"]: doc_text(r) for _, r in a.iterrows()} |
|
|
|
|
| def attach_labels(questions, review): |
| """review: qid, dogru_mu (E/K/H). Eslesmeyen soru olmamali (eksik degerlendirme sessizce gecmesin).""" |
| lab = dict(zip(review["qid"], review["dogru_mu"].str.strip().str.upper())) |
| missing = sorted(set(questions["qid"]) - set(lab)) |
| assert not missing, f"Değerlendirmesi olmayan soru: {missing[:10]} ({len(missing)})" |
| bad = {v for v in lab.values()} - {"E", "K", "H"} |
| assert not bad, f"Geçersiz etiket: {bad}" |
| q = questions.copy() |
| q["etiket"] = q["qid"].map(lab) |
| return q |
|
|
|
|
| def build_conditions(q, perturbations, perturb_fn, seed=0): |
| """-> queries {kosul: {qid: metin}}, changed {kosul: set(qid)}, qrels. 'clean' = gundelik.""" |
| clean = dict(zip(q["qid"], q["gundelik"])) |
| resmi = {qid: (r if isinstance(r, str) and r.strip() else clean[qid]) for qid, r in zip(q["qid"], q["resmi"])} |
| queries = {"clean": clean, "resmi": resmi} |
| for p in perturbations: |
| if p != "clean": |
| queries[p] = {qid: perturb_fn(t, qid, p, seed=seed) for qid, t in clean.items()} |
| changed = {c: {qid for qid, t in qs.items() if t != clean[qid]} for c, qs in queries.items() if c != "clean"} |
| qrels = {qid: {doc: 1} for qid, doc in zip(q["qid"], q["doc_id"])} |
| return queries, changed, qrels |
|
|
|
|
| def as_sets(corpus, qrels, name="arzuhal"): |
| """pusula_dense.benchmark_fingerprint'in bekledigi bicim.""" |
| return {name: SimpleNamespace(corpus=corpus, qrels=qrels)} |
|
|
|
|
| def law_hit_at_1(ranked, rels): |
| """Ilk sonuc dogru KANUNDAN mi (madde yanlis olsa bile)? Model en azindan hukuk alanini bulabiliyor mu.""" |
| if not ranked: |
| return 0.0 |
| gold_law = next(iter(rels)).split(":")[0] |
| return float(ranked[0].split(":")[0] == gold_law) |
|
|
|
|
| ARZ_METRICS = {**METRICS, "kanun@1": law_hit_at_1} |
|
|
|
|
| def mean_over(pq_metric, qids): |
| return float(np.mean([pq_metric[q] for q in qids])) if qids else float("nan") |
|
|
|
|
| def retention(pq, cond, changed, qids, metric): |
| """Kosulun DEGISTIRDIGI sorgularda: bozuk/temiz ortalama orani (Pusula ile ayni tanim).""" |
| sel = [q for q in qids if q in changed[cond]] |
| if not sel: |
| return float("nan"), 0 |
| c = mean_over(pq["clean"][metric], sel) |
| return (mean_over(pq[cond][metric], sel) / c if c > 0 else float("nan")), len(sel) |
|
|
|
|
| def leaderboard(results, conditions, changed, subsets, metric="ndcg@10", rank_by="clean", n_boot=2000, seed=0): |
| """results: {short: {"per_query": {kosul: {metrik: {qid: deger}}}, "spec": {...}}} |
| subsets: {"ana": [qid...], "kati": [qid...]}. Siralama ana kumede rank_by kosuluna gore. |
| Komsu modeller eslestirilmis bootstrap ile; %95 GA sifiri kapsiyorsa ayni grup.""" |
| main = subsets["ana"] |
| rows = [] |
| for short, r in results.items(): |
| pq = r["per_query"] |
| row = {"model": short, "params_m": r["spec"].get("params_m")} |
| for c in ["clean"] + [c for c in conditions if c != "clean"]: |
| row[c] = mean_over(pq[c][metric], main) |
| for c in conditions: |
| if c != "clean": |
| row[f"koruma_{c}"], _ = retention(pq, c, changed, main, metric) |
| for name, qids in subsets.items(): |
| if name != "ana": |
| row[f"clean_{name}"] = mean_over(pq["clean"][metric], qids) |
| row["kanun@1"] = mean_over(pq["clean"]["kanun@1"], main) |
| row["recall@10"] = mean_over(pq["clean"]["recall@10"], main) |
| rows.append(row) |
| lb = pd.DataFrame(rows).sort_values(rank_by, ascending=False).reset_index(drop=True) |
| group, groups, sig = 1, [], [] |
| for i in range(len(lb)): |
| if i > 0: |
| a = [results[lb.model[i - 1]]["per_query"][rank_by][metric][q] for q in main] |
| b = [results[lb.model[i]]["per_query"][rank_by][metric][q] for q in main] |
| _, ci = paired_bootstrap(b, a, n_boot=n_boot, seed=seed) |
| s = ci[0] > 0 |
| sig.append(s) |
| if s: |
| group += 1 |
| groups.append(group) |
| lb.insert(1, "grup", groups) |
| lb["rank"] = range(1, len(lb) + 1) |
| if "clean_kati" in lb: |
| lb["rank_kati"] = lb["clean_kati"].rank(ascending=False, method="min").astype(int) |
| return lb |
|
|
|
|
| def per_law(results, q, metric="ndcg@10", qids=None, cond="clean"): |
| """Model x kanun tablosu (ana kume). Kanun adi varsa ad, yoksa numara.""" |
| by = "law_title" if "law_title" in q else "law" |
| qids = set(qids or q["qid"]) |
| sub = q[q["qid"].isin(qids)] |
| out = {} |
| for short, r in results.items(): |
| m = r["per_query"][cond][metric] |
| out[short] = {law: float(np.mean([m[x] for x in g["qid"]])) for law, g in sub.groupby(by)} |
| return pd.DataFrame(out).T |
|
|
