src/evaluate.py

import argparse
import csv
from collections import defaultdict
from pathlib import Path

import jiwer
from tabulate import tabulate


DEFAULT_GOLD = Path("data/gold.tsv")
DEFAULT_PREDICTIONS = Path("data/predictions.tsv")
PHONEME_CHARS = set("abdefhijklmnopstuvwzɡʁʃʒʔˈχ")
PHONEME_TRANSLATION = str.maketrans({"x": "χ", "r": "ʁ", "g": "ɡ"})


def normalize_phonemes(text):
    text = text.translate(PHONEME_TRANSLATION)
    return "".join(char for char in text if char in PHONEME_CHARS)


def read_tsv(path):
    if not path.exists() or path.stat().st_size == 0:
        return []
    with path.open("r", encoding="utf-8-sig", newline="") as f:
        return list(csv.DictReader(f, delimiter="\t"))


def read_predictions(path):
    if not path.exists() or path.stat().st_size == 0:
        return []
    with path.open("r", encoding="utf-8-sig", newline="") as f:
        return list(csv.DictReader(f, delimiter="\t"))


def read_predictions_header(path):
    if not path.exists() or path.stat().st_size == 0:
        return []
    with path.open("r", encoding="utf-8-sig", newline="") as f:
        return next(csv.reader(f, delimiter="\t"), [])


def parse_label(label):
    """Parse labels like: 1=ʔelˈajiχ 4=kibˈalt."""
    targets = {}
    for part in label.split():
        if "=" not in part:
            continue
        index, ipa = part.split("=", 1)
        targets[int(index)] = normalize_phonemes(ipa)
    return targets


def prediction_targets(prediction):
    """Allow either full token output or target-only index=IPA predictions."""
    if all("=" in part for part in prediction.split() if part):
        return parse_label(prediction)
    return None


def get_prediction_rows(gold_rows, pred_rows):
    if not pred_rows:
        raise ValueError("Predictions file is empty. Run a plan script first.")

    if len(pred_rows) != len(gold_rows):
        raise ValueError(
            "Predictions must have the same number of rows "
            f"as gold: got {len(pred_rows)}, expected {len(gold_rows)}."
        )
    return pred_rows


def target_prediction(row, pred_text):
    targets = parse_label(row["Label"])
    indexed_pred = prediction_targets(pred_text)

    if indexed_pred is not None:
        return " ".join(indexed_pred.get(i, "") for i in targets)

    pred_tokens = pred_text.split()
    return " ".join(normalize_phonemes(pred_tokens[i]) if i < len(pred_tokens) else "" for i in targets)


def score_rows(gold_rows, pred_rows, pred_col):
    by_category = defaultdict(lambda: {"refs": [], "hyps": [], "exact": []})
    all_refs = []
    all_hyps = []
    all_exact = []

    for gold, pred in zip(gold_rows, get_prediction_rows(gold_rows, pred_rows)):
        ref = " ".join(parse_label(gold["Label"]).values())
        hyp = target_prediction(gold, pred.get(pred_col, ""))
        exact = int(ref == hyp)

        all_refs.append(ref)
        all_hyps.append(hyp)
        all_exact.append(exact)

        bucket = by_category[gold["Category"]]
        bucket["refs"].append(ref)
        bucket["hyps"].append(hyp)
        bucket["exact"].append(exact)

    return all_refs, all_hyps, all_exact, by_category


def summarize(name, refs, hyps, exact):
    return {
        "Category": name,
        "Items": len(refs),
        "WER": jiwer.wer(refs, hyps),
        "CER": jiwer.cer(refs, hyps),
        "Exact": sum(exact) / len(exact) if exact else 0.0,
    }


def prediction_columns(predictions_path, requested):
    if requested:
        return requested

    header = read_predictions_header(predictions_path)
    columns = header
    if not columns:
        raise ValueError(
            "No prediction columns found. Expected data/predictions.tsv with "
            "one or more G2P columns."
        )
    return columns


def main():
    parser = argparse.ArgumentParser(description="Evaluate MILIM-Bench G2P predictions.")
    parser.add_argument("predictions", type=Path, nargs="?", default=DEFAULT_PREDICTIONS)
    parser.add_argument("--gold", type=Path, default=DEFAULT_GOLD)
    parser.add_argument(
        "--prediction-column",
        action="append",
        help="Prediction column to evaluate. May be passed more than once. Defaults to all non-Text columns.",
    )
    args = parser.parse_args()

    gold_rows = read_tsv(args.gold)
    pred_rows = read_predictions(args.predictions)
    columns = prediction_columns(args.predictions, args.prediction_column)
    table = []

    for column in columns:
        refs, hyps, exact, by_category = score_rows(gold_rows, pred_rows, column)
        model_summary = [summarize("ALL", refs, hyps, exact)]
        model_summary.extend(
            summarize(category, data["refs"], data["hyps"], data["exact"])
            for category, data in sorted(by_category.items())
        )

        for row in model_summary:
            table.append(
                [
                    column,
                    row["Category"],
                    row["Items"],
                    f"{1 - row['WER']:.4f}",
                    f"{1 - row['CER']:.4f}",
                    f"{row['Exact']:.4f}",
                ]
            )

    print(tabulate(table, headers=["model", "category", "items", "WER ↑", "CER ↑", "Exact ↑"]))


if __name__ == "__main__":
    main()