File size: 4,851 Bytes
f8c0823 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | """Auditoría y versión limpia (v1.1) de MachismoDetect y SinOdio-LATAM, con el mismo método que Racismo/Clasismo."""
import glob, json, os, re, unicodedata
import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score
from sklearn.model_selection import GroupShuffleSplit
SEED = 42
ABBREV = {"xq": "porque", "pq": "porque", "xk": "porque", "pk": "porque", "q": "que", "ke": "que", "k": "que", "pa": "para",
"tb": "tambien", "tmb": "tambien", "toy": "estoy", "vdd": "verdad", "td": "todo", "nd": "nada", "grax": "gracias"}
LAUGHS = {"jaja", "jajaja", "jeje", "lol", "jajaj", "xd", "v"}
def skeleton(t):
t = unicodedata.normalize("NFKD", str(t).lower())
t = "".join(c for c in t if not unicodedata.combining(c))
t = re.sub(r"[^a-zñ0-9 ]+", " ", t); t = re.sub(r"(.)\1{1,}", r"\1", t)
return " ".join(w for w in (ABBREV.get(w, w) for w in t.split()) if w not in LAUGHS)
def splits_of(prefix):
return {os.path.basename(f).split("__")[1].replace(".parquet", ""): pd.read_parquet(f) for f in sorted(glob.glob(f"data/{prefix}__*.parquet"))}
def to_long(name, parts):
out = []
for sp, d in parts.items():
if name == "machismodetect":
meta = d[["nivel", "subcategoria", "pais", "contexto"]]
out.append(pd.concat([meta.assign(texto=d.texto_machista, etiqueta=1), meta.assign(texto=d.texto_inclusivo, etiqueta=0)]).assign(split_original=sp))
else:
out.append(d.rename(columns={"text": "texto", "label": "etiqueta", "country": "pais", "scenario": "escenario"}).assign(split_original=sp))
df = pd.concat(out, ignore_index=True)
df["texto"] = df.texto.astype(str).str.strip()
df["esqueleto"] = df.texto.map(skeleton)
return df
def baseline(tr, te, col="etiqueta"):
v = TfidfVectorizer(ngram_range=(1, 2), min_df=2, sublinear_tf=True)
m = LogisticRegression(max_iter=3000, class_weight="balanced").fit(v.fit_transform(tr.texto), tr[col])
return round(float(f1_score(te[col], m.predict(v.transform(te.texto)), average="macro")), 3)
def audit(name, group_col):
parts = splits_of(name)
df = to_long(name, parts)
rep = {"dataset": name, "filas_originales_por_split": {k: len(v) for k, v in parts.items()},
"textos_formato_largo": len(df), "por_etiqueta_original": df.etiqueta.value_counts().to_dict()}
# fuga en splits originales
tr0, te0 = df[df.split_original == "train"], df[df.split_original == "test"]
rep["fuga_original_test_en_train_exacta_pct"] = round(100 * te0.texto.isin(set(tr0.texto)).mean(), 1)
rep["fuga_original_test_en_train_esqueleto_pct"] = round(100 * te0.esqueleto.isin(set(tr0.esqueleto)).mean(), 1)
rep["baseline_split_original_f1"] = baseline(tr0, te0)
# limpieza
exact = df.drop_duplicates(["texto", "etiqueta"])
clean = exact.drop_duplicates(["esqueleto", "etiqueta"])
conf = clean.groupby("esqueleto").etiqueta.nunique(); ck = set(conf[conf > 1].index)
clean = clean[~clean.esqueleto.isin(ck)].reset_index(drop=True)
rep.update({"tras_duplicados_exactos": len(exact), "tras_casi_duplicados": len(clean),
"esqueletos_con_etiqueta_contradictoria": len(ck), "por_etiqueta_v1_1": clean.etiqueta.value_counts().to_dict(),
"unicos_por_etiqueta_pct_conservado": {int(k): round(100 * clean.etiqueta.eq(k).sum() / df.etiqueta.eq(k).sum(), 1) for k in [0, 1]}})
g1 = GroupShuffleSplit(1, test_size=0.3, random_state=SEED); a, b = next(g1.split(clean, groups=clean.esqueleto))
rest = clean.iloc[b]; g2 = GroupShuffleSplit(1, test_size=0.5, random_state=SEED); c, d = next(g2.split(rest, groups=rest.esqueleto))
sp = {"train": clean.iloc[a], "validation": rest.iloc[c], "test": rest.iloc[d]}
rep["splits_v1_1"] = {k: len(v) for k, v in sp.items()}
rep["baseline_split_v1_1_f1"] = baseline(sp["train"], sp["test"])
gs = GroupShuffleSplit(1, test_size=0.25, random_state=SEED); x, y = next(gs.split(clean, groups=clean[group_col]))
rep[f"baseline_{group_col}_no_vistas_f1"] = baseline(clean.iloc[x], clean.iloc[y])
rep["longitud_mediana"] = clean.groupby("etiqueta").texto.apply(lambda s: int(s.str.len().median())).to_dict()
os.makedirs(f"v11/{name}", exist_ok=True)
cols = [c for c in clean.columns if c not in ("esqueleto", "split_original")]
for k, v in sp.items():
v[cols].to_csv(f"v11/{name}/{k}.csv", index=False)
clean[cols].to_parquet(f"v11/{name}/completo.parquet", index=False)
json.dump(rep, open(f"v11/{name}/reporte_v1_1.json", "w"), ensure_ascii=False, indent=2, default=int)
print(json.dumps(rep, ensure_ascii=False, default=int))
if __name__ == "__main__":
audit("machismodetect", "subcategoria")
audit("sinodio_latam", "escenario")
|