metrics/bleu.py

# metrics/bleu.py
"""
BLEU metric wrappers using sacreBLEU and file_utils.
"""
from utils.file_utils import  *
from sacrebleu.metrics import BLEU

# Instância global de BLEU com tokenização 'intl', lowercase e smoothing 'exp'
_bleu_scorer = BLEU(tokenize='intl', lowercase=True, smooth_method='exp')

def section_bleu(gen_txt: str, ref_txt: str) -> float:
    """
    Calcula BLEU para um par de strings (seção), retornando score de 0 a 100.
    """
    if not gen_txt.strip() and not ref_txt.strip():
        return 100.0
    if (not gen_txt.strip()) ^ (not ref_txt.strip()):
        return 0.0
    return _bleu_scorer.sentence_score(gen_txt, [ref_txt]).score

def full_bleu(gen_raw: str, ref_raw: str) -> float:
    """
    Calcula BLEU global para strings completas, retornando score de 0 a 100.
    """
    gen = normalize_and_flatten(gen_raw)
    ref = normalize_and_flatten(ref_raw)
    if not gen and not ref:
        return 100.0
    if (not gen) ^ (not ref):
        return 0.0
    return _bleu_scorer.sentence_score(gen, [ref]).score


def compute_bleu_single(reference: str, prediction: str) -> str:
    """
    Compute and format BLEU score for a single pair.
    """
    if not reference or not prediction:
        return "Please provide both texts."
    score = full_bleu(prediction, reference) / 100.0
    return f"BLEU Score: {score:.4f}"