File size: 8,299 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 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 | """
Validación externa de los datasets sintéticos de DignityAI contra datos reales anotados por personas.
A. Reconocimiento externo: un modelo entrenado con comentarios reales anotados por humanos
(piuba-bigdata/beto-contextualized-hate-speech, UBA) clasifica muestras de nuestros textos.
B. Línea base del mismo modelo en su propio test real (para saber qué tan sensible es).
C. Evaluación externa de SinOdio v3 en HateCheck-ES (casos funcionales escritos por expertos).
D. Transferencia: ¿un clasificador entrenado con nuestros datos detecta odio real?
¿y agregar nuestros datos a datos reales ayuda o estorba?
"""
import json, glob, sys, time
import numpy as np, pandas as pd, torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score, precision_score, recall_score, accuracy_score
RNG = 42
DEV = "mps" if torch.backends.mps.is_available() else "cpu"
OUT = {}
def log(*a): print(time.strftime("%H:%M:%S"), *a, flush=True)
# ---------- datos propios
def own():
r = pd.read_parquet("racismo/v1.1/completo.parquet")
c = pd.read_parquet("clasismo/v1.1/completo.parquet")
m = pd.concat([pd.read_parquet(f) for f in glob.glob("data/machismodetect__*.parquet")])
m = pd.concat([
m[["texto_machista", "nivel", "pais"]].rename(columns={"texto_machista": "texto", "nivel": "categoria"}).assign(etiqueta=1),
m[["texto_inclusivo", "nivel", "pais"]].rename(columns={"texto_inclusivo": "texto", "nivel": "categoria"}).assign(etiqueta=0),
]).drop_duplicates("texto")
s = pd.concat([pd.read_parquet(f) for f in glob.glob("data/sinodio_latam__*.parquet")])
s = s.rename(columns={"text": "texto", "label": "etiqueta", "scenario": "categoria", "country": "pais"}).drop_duplicates("texto")
return {"racismo": r, "clasismo": c, "machismo": m, "sinodio": s}
def sample(df, n_pos=1500, n_neg=1000):
p = df[df.etiqueta == 1]; n = df[df.etiqueta == 0]
return pd.concat([p.sample(min(n_pos, len(p)), random_state=RNG), n.sample(min(n_neg, len(n)), random_state=RNG)])
# ---------- modelos
def load(rid):
tok = AutoTokenizer.from_pretrained(rid)
mod = AutoModelForSequenceClassification.from_pretrained(rid).to(DEV).eval()
return tok, mod
@torch.no_grad()
def logits(tok, mod, texts, ctx=None, bs=64):
out = []
for i in range(0, len(texts), bs):
a = list(texts[i:i + bs])
b = list(ctx[i:i + bs]) if ctx is not None else None
enc = tok(a, b, truncation=True, max_length=128, padding=True, return_tensors="pt").to(DEV)
out.append(mod(**enc).logits.float().cpu())
return torch.cat(out)
UBA_GROUPS = ["WOMEN", "LGBTI", "RACISM", "CLASS", "POLITICS", "DISABLED", "APPEARANCE", "CRIMINAL"]
TARGET = {"racismo": "RACISM", "clasismo": "CLASS", "machismo": "WOMEN", "sinodio": None}
def main():
data = own()
tok, mod = load("piuba-bigdata/beto-contextualized-hate-speech")
id2l = [mod.config.id2label[i] for i in range(len(mod.config.id2label))]
log("UBA labels", id2l)
gi = [id2l.index(g) for g in UBA_GROUPS]
# B. Línea base UBA en su test real (con contexto, como fue entrenado)
real = pd.read_parquet("data/piuba_real__test.parquet")
rs = pd.concat([real[real[g] == 1].sample(min(400, int((real[g] == 1).sum())), random_state=RNG) for g in ["RACISM", "CLASS", "WOMEN"]]
+ [real[real.HATEFUL == 0].sample(1000, random_state=RNG)]).drop_duplicates("id")
for use_ctx in [True, False]:
lg = logits(tok, mod, rs.text.tolist(), rs.context_tweet.fillna("").tolist() if use_ctx else None)
pr = (torch.sigmoid(lg) > 0.5).numpy()
res = {g: round(float(pr[(rs[g] == 1).values, id2l.index(g)].mean()), 3) for g in ["RACISM", "CLASS", "WOMEN"]}
res["falsas_alarmas_no_odio"] = round(float(pr[(rs.HATEFUL == 0).values][:, gi].any(1).mean()), 3)
OUT[f"uba_en_real_{'con' if use_ctx else 'sin'}_contexto"] = res
log("UBA real", use_ctx, res)
# A. UBA sobre nuestros textos (sin contexto)
sin_tok, sin_mod = load("antonn-dromundo/SinOdio-BETO-HateSpeech-Detector-v3")
sid2l = {int(k): v for k, v in sin_mod.config.id2label.items()}
log("SinOdio labels", sid2l)
hate_idx = 1
for name, df in data.items():
s = sample(df)
pr = (torch.sigmoid(logits(tok, mod, s.texto.tolist())) > 0.5).numpy()
sp = torch.softmax(logits(sin_tok, sin_mod, s.texto.tolist()), -1)[:, hate_idx].numpy() > 0.5
pos, neg = (s.etiqueta == 1).values, (s.etiqueta == 0).values
r = {
"n_discriminatorios": int(pos.sum()), "n_contraste": int(neg.sum()),
"uba_detecta_odio_en_discriminatorios": round(float(pr[pos][:, gi].any(1).mean()), 3),
"uba_falsas_alarmas_en_contraste": round(float(pr[neg][:, gi].any(1).mean()), 3),
"sinodio_detecta_odio_en_discriminatorios": round(float(sp[pos].mean()), 3),
"sinodio_falsas_alarmas_en_contraste": round(float(sp[neg].mean()), 3),
}
if TARGET[name]:
r[f"uba_etiqueta_{TARGET[name]}_en_discriminatorios"] = round(float(pr[pos, id2l.index(TARGET[name])].mean()), 3)
s = s.assign(uba_odio=pr[:, gi].any(1), sinodio_odio=sp)
r["uba_por_categoria"] = s[pos].groupby("categoria").uba_odio.mean().round(3).to_dict()
r["uba_etiquetas_mas_asignadas"] = dict(pd.Series([id2l[j] for row in pr[pos] for j in np.where(row)[0]]).value_counts().head(5))
OUT[f"externo_{name}"] = r
log(name, json.dumps(r, ensure_ascii=False, default=int))
# C. SinOdio v3 y UBA en HateCheck-ES
hc = pd.read_parquet("data/hatecheck_es__test.parquet")
y = (hc.label_gold == "hateful").astype(int).values
sp = (torch.softmax(logits(sin_tok, sin_mod, hc.test_case.tolist()), -1)[:, hate_idx].numpy() > 0.5).astype(int)
up = (torch.sigmoid(logits(tok, mod, hc.test_case.tolist())) > 0.5).numpy()[:, gi].any(1).astype(int)
hcr = {}
for nm, p in [("sinodio_v3", sp), ("uba_beto", up)]:
f = hc.assign(ok=(p == y))
hcr[nm] = {"exactitud": round(accuracy_score(y, p), 3), "f1_macro": round(f1_score(y, p, average="macro"), 3),
"recall_odio": round(recall_score(y, p), 3), "precision_odio": round(precision_score(y, p), 3),
"peores_funcionalidades": f.groupby("functionality").ok.mean().sort_values().head(6).round(3).to_dict()}
OUT["hatecheck_es"] = hcr
log("hatecheck", json.dumps(hcr, ensure_ascii=False))
# D. Transferencia sintético -> real (TF-IDF + LR)
tr_real = pd.read_parquet("data/piuba_real__train.parquet"); te_real = real
tr_neg = tr_real[tr_real.HATEFUL == 0].sample(8000, random_state=RNG)
trans = {}
for name in ["racismo", "clasismo", "machismo"]:
g = TARGET[name]
te = pd.concat([te_real[te_real[g] == 1], te_real[te_real.HATEFUL == 0].sample(3000, random_state=RNG)])
yte = (te[g] == 1).astype(int).values
syn = data[name][["texto", "etiqueta"]]
rp = tr_real[tr_real[g] == 1]
sets = {
"solo_sintetico": (syn.texto, syn.etiqueta),
"solo_real": (pd.concat([rp.text, tr_neg.text]), np.r_[np.ones(len(rp)), np.zeros(len(tr_neg))]),
}
sets["real_mas_sintetico"] = (pd.concat([sets["solo_real"][0], syn.texto]), np.r_[sets["solo_real"][1], syn.etiqueta])
rr = {"n_test_positivos": int(yte.sum()), "n_test_negativos": int((1 - yte).sum())}
for k, (X, Y) in sets.items():
v = TfidfVectorizer(ngram_range=(1, 2), min_df=2, sublinear_tf=True, max_features=200000)
m = LogisticRegression(max_iter=3000, class_weight="balanced").fit(v.fit_transform(X), Y)
p = m.predict(v.transform(te.text))
rr[k] = {"f1": round(f1_score(yte, p), 3), "recall": round(recall_score(yte, p), 3), "precision": round(precision_score(yte, p, zero_division=0), 3)}
trans[name] = rr
log("transfer", name, rr)
OUT["transferencia_a_real_uba"] = trans
json.dump(OUT, open("resultados_validacion.json", "w"), ensure_ascii=False, indent=2, default=int)
log("LISTO")
if __name__ == "__main__":
main()
|