"""
GLEU (Generalized Language Evaluation Understanding) score.
Preferred over BLEU for grammatical error correction tasks.
Also computes BERTScore for semantic similarity evaluation.
"""

import sacrebleu
from bert_score import score as bert_score_fn
from typing import List, Tuple
from loguru import logger


class GLEUScorer:
    """Computes GLEU and BERTScore metrics for GEC evaluation."""

    def compute_gleu(
        self,
        predictions: List[str],
        references: List[str],
    ) -> float:
        """Corpus-level GLEU score (0-100).

        GLEU is the geometric mean of n-gram precisions and recall,
        preferred over BLEU for GEC because it equally penalises
        both under-correction and over-correction.
        """
        if not predictions or not references:
            return 0.0

        # sacrebleu expects references as a list of lists
        refs = [references]

        # Use BLEU with smoothing as GLEU approximation
        # sacrebleu doesn't have a native GLEU, so we use smoothed BLEU
        bleu = sacrebleu.corpus_bleu(
            predictions,
            refs,
            smooth_method="exp",
            smooth_value=0.1,
        )
        return bleu.score

    def compute_bert_score(
        self,
        predictions: List[str],
        references: List[str],
        lang: str = "en",
    ) -> Tuple[float, float, float]:
        """Returns (precision, recall, F1) as averages over the batch."""
        if not predictions or not references:
            return (0.0, 0.0, 0.0)

        try:
            P, R, F1 = bert_score_fn(
                predictions,
                references,
                lang=lang,
                verbose=False,
                device="cpu",  # CPU-optimised
            )
            return (
                P.mean().item(),
                R.mean().item(),
                F1.mean().item(),
            )
        except Exception as e:
            logger.warning(f"BERTScore computation failed: {e}")
            return (0.0, 0.0, 0.0)