app/application/threshold_mapper.py

from dataclasses import dataclass
import re
import unicodedata

from app.domain.entities import ClassificationLabel, ClassificationResult, ModelScores


HATER_PATTERNS = [
    r"\bmu[eé]rete\b",
    r"\bmu[eé]ranse\b",
    r"\bmuere\b",
    r"\bque\s+te\s+mueras\b",
    r"\bojal[aá]\s+(?:que\s+)?(?:te\s+)?mueras\b",
    r"\bte\s+voy\s+a\s+matar\b",
    r"\bdeber[ií]as\s+morir\b",
    r"\bmaldito(?:s|as)?\b",
    r"\bescoria\b",
]

CRITICAL_PATTERNS = [
    r"\bhorrible\b",
    r"\basqueros[oa]s?\b",
    r"\bapesta\b",
    r"\binfumable\b",
    r"\bbasura\b",
    r"\bp[eé]sim[oa]\b",
    r"\bmal[ií]sim[oa]\b",
    r"\bmalo\s+totalmente\b",
    r"\bno\s+me\s+gusta\b",
    r"\bdecepcionante\b",
    r"\bdesastre\b",
    r"\bvergonzoso\b",
]


@dataclass(frozen=True)
class ThresholdConfig:
    hater_hs: float = 0.70
    hater_ag: float = 0.80
    hater_borderline_hs: float = 0.55
    hater_borderline_ag: float = 0.55
    critical_hs: float = 0.25
    critical_ag: float = 0.25
    critical_tr: float = 0.25


class HateSpeechThresholdMapper:
    def __init__(self, config: ThresholdConfig):
        self._config = config

    def map(self, scores: ModelScores) -> ClassificationResult:
        hs = self._clamp(scores.hs)
        tr = self._clamp(scores.tr)
        ag = self._clamp(scores.ag)
        text_match = self._match_text_rules(scores.text)

        hater_score = max(hs, ag)
        critical_score = max(hs, ag, tr)
        neutral_score = self._clamp(1 - critical_score)

        if text_match and text_match["label"] == ClassificationLabel.HATER:
            label = ClassificationLabel.HATER
            confidence = max(hater_score, 0.95)
            hater_score = max(hater_score, 0.95)
            critical_score = max(critical_score, 0.95)
            neutral_score = self._clamp(1 - hater_score)
        elif text_match and text_match["label"] == ClassificationLabel.CRITICAL:
            label = ClassificationLabel.CRITICAL
            confidence = max(critical_score, 0.85)
            critical_score = max(critical_score, 0.85)
            neutral_score = self._clamp(1 - critical_score)
        elif self._is_hater(hs, ag):
            label = ClassificationLabel.HATER
            confidence = hater_score
        elif self._is_critical(hs, tr, ag):
            label = ClassificationLabel.CRITICAL
            confidence = critical_score
        else:
            label = ClassificationLabel.NEUTRAL
            confidence = neutral_score

        return ClassificationResult(
            predicted_label=label,
            confidence=round(confidence, 4),
            hater_score=round(hater_score, 4),
            critical_score=round(critical_score, 4),
            neutral_score=round(neutral_score, 4),
            raw_response={
                "model_scores": {"HS": hs, "TR": tr, "AG": ag},
                "model_output": scores.raw,
                "thresholds": self._config.__dict__,
                "rule_match": self._serialize_rule_match(text_match),
            },
        )

    def _is_hater(self, hs: float, ag: float) -> bool:
        return (
            hs >= self._config.hater_hs
            or ag >= self._config.hater_ag
            or (
                hs >= self._config.hater_borderline_hs
                and ag >= self._config.hater_borderline_ag
            )
        )

    def _is_critical(self, hs: float, tr: float, ag: float) -> bool:
        return (
            hs >= self._config.critical_hs
            or ag >= self._config.critical_ag
            or tr >= self._config.critical_tr
        )

    @staticmethod
    def _clamp(value: float) -> float:
        return max(0.0, min(1.0, float(value or 0.0)))

    @classmethod
    def _match_text_rules(cls, text: str):
        normalized = cls._normalize_text(text)
        if not normalized:
            return None

        for pattern in HATER_PATTERNS:
            if re.search(pattern, normalized):
                return {
                    "label": ClassificationLabel.HATER,
                    "pattern": pattern,
                    "source": "text_rule",
                }

        for pattern in CRITICAL_PATTERNS:
            if re.search(pattern, normalized):
                return {
                    "label": ClassificationLabel.CRITICAL,
                    "pattern": pattern,
                    "source": "text_rule",
                }

        return None

    @staticmethod
    def _normalize_text(text: str) -> str:
        text = unicodedata.normalize("NFKD", text or "")
        text = "".join(char for char in text if not unicodedata.combining(char))
        text = text.lower()
        return re.sub(r"\s+", " ", text).strip()

    @staticmethod
    def _serialize_rule_match(match):
        if not match:
            return None
        return {
            "label": match["label"].value,
            "pattern": match["pattern"],
            "source": match["source"],
        }