"""Oswald's scoring, the same as the V3 evaluation (scripts/qwen3_asr/evaluate.py and parakeet_v3/evaluate_v3.py).

Every text, reference and hypothesis, is read back to its spoken form (itn.spoken: "1500 goud" -> "mil senksan
goud") and stripped of case, punctuation and accents (normalize.metric_form). So V1, which writes numbers in words,
and V2/V3, which write digits, are compared on the same words, and è/e never decides an error.
"""
import difflib
import html
import re

import jiwer

from itn import itn, spoken
from normalize import metric_form, normalize


def mform(text):
    return metric_form(spoken(text or ""))


def target(text):
    """The dataset's sentence as V2/V3 were taught to write it: lowercase, no punctuation, numbers as digits."""
    return itn(normalize(text or ""))


def rates(refs, hyps):
    pairs = [(mform(r), mform(h)) for r, h in zip(refs, hyps)]
    pairs = [(r, h) for r, h in pairs if r]
    if not pairs:
        return None, None
    return jiwer.wer([p[0] for p in pairs], [p[1] for p in pairs]), jiwer.cer([p[0] for p in pairs], [p[1] for p in pairs])


def clip_wer(ref, hyp):
    r = mform(ref)
    return jiwer.wer(r, mform(hyp)) if r else None


# Number words in metric form. A tens word carries its unit glued on (vennsenk, katrevendis).
_UNITS = {"zewo", "en", "de", "twa", "kat", "senk", "sis", "set", "uit", "nef", "dis", "onz", "douz", "trez",
          "katoz", "kenz", "sez", "diset", "dizuit", "diznef", "san", "mil", "milyon", "milya", "premye",
          "desan", "twasan", "katsan", "senksan", "sisan", "setsan", "uisan", "nefsan", "pwen", "vigil"}
_TENS = re.compile(r"^(ven|venn|vent|trant|trann|karant|karann|senkant|senkann|swasant|swasann|katreven)\w*$")
# These are also ordinary words (of, card, without, ...): alone they are not counted as a number.
_AMBIGUOUS = {"en", "de", "twa", "kat", "san", "sis", "set", "premye", "pwen", "vigil"}


def _is_num(w):
    return w in _UNITS or bool(_TENS.match(w)) or (w.endswith("yem") and len(w) > 4)


def number_spans(text):
    words, spans, cur = mform(text).split(), [], []
    for w in words + [""]:
        if w and _is_num(w):
            cur.append(w)
            continue
        if cur and (len(cur) > 1 or cur[0] not in _AMBIGUOUS):
            spans.append(" ".join(cur))
        cur = []
    return spans


def number_accuracy(refs, hyps):
    """Share of the reference's numbers found whole in the transcript ("de mil vennkat" as "de mil kat" is wrong)."""
    hit = total = 0
    for r, h in zip(refs, hyps):
        hyp = " " + " ".join(mform(h).split()) + " "
        for span in number_spans(r):
            total += 1
            hit += f" {span} " in hyp
    return (hit / total if total else None), total


def diff_html(ref, hyp):
    """The hypothesis as scored, word by word: missing words struck through in red, extra or wrong ones in orange."""
    r, h = mform(ref).split(), mform(hyp).split()
    out = []
    for op, a, b, c, d in difflib.SequenceMatcher(a=r, b=h, autojunk=False).get_opcodes():
        if op == "equal":
            out.append(html.escape(" ".join(h[c:d])))
            continue
        if b > a:
            out.append(f'<del class="miss">{html.escape(" ".join(r[a:b]))}</del>')
        if d > c:
            out.append(f'<ins class="extra">{html.escape(" ".join(h[c:d]))}</ins>')
    return " ".join(out)