scripts/fix_abvd_ipa.py

#!/usr/bin/env python3
"""Fix ABVD-sourced lexicon entries where Word == IPA (fake IPA).

Scans all TSV files in data/training/lexicons/ for entries sourced from
ABVD where the Word column is identical to the IPA column. For those
entries, applies G2P conversion (transphone if available, otherwise a
lightweight Austronesian-specific rule-based fallback) and recomputes
the SCA encoding from the new IPA.

TSV format: Word\tIPA\tSCA\tSource\tConcept_ID\tCognate_Set_ID
"""

from __future__ import annotations

import csv
import re
import sys
from pathlib import Path

ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT / "cognate_pipeline" / "src"))

from cognate_pipeline.normalise.sound_class import ipa_to_sound_class

LEXICON_DIR = ROOT / "data" / "training" / "lexicons"
HEADER = "Word\tIPA\tSCA\tSource\tConcept_ID\tCognate_Set_ID\n"


# ---------------------------------------------------------------------------
# G2P backend selection
# ---------------------------------------------------------------------------

_USE_TRANSPHONE = False

try:
    from transphone import read_ipa  # type: ignore[import-untyped]

    _USE_TRANSPHONE = True
except ImportError:
    pass


def _transphone_g2p(word: str) -> str:
    """Attempt G2P via transphone."""
    try:
        phones = read_ipa(word, "ind")  # Indonesian as default Austronesian
        if phones:
            return "".join(phones)
    except Exception:
        pass
    return ""


# ---------------------------------------------------------------------------
# Lightweight Austronesian rule-based G2P fallback
# ---------------------------------------------------------------------------

# Ordered list of (regex_pattern, ipa_replacement).
# Longer / multi-char patterns come first so they match before single chars.
_AUSTRONESIAN_RULES: list[tuple[str, str]] = [
    # Trigraphs
    (r"ngg", "ŋɡ"),
    (r"ngk", "ŋk"),

    # Digraphs (order matters: test longer sequences first)
    (r"ng", "ŋ"),
    (r"ny", "ɲ"),
    (r"sy", "ʃ"),
    (r"ch", "tʃ"),
    (r"sh", "ʃ"),
    (r"th", "t"),       # Austronesian <th> is typically /t/
    (r"dj", "dʒ"),
    (r"tj", "tʃ"),
    (r"nj", "ɲ"),
    (r"kh", "x"),
    (r"gh", "ɣ"),
    (r"ph", "p"),       # Austronesian <ph> is typically /p/
    (r"bh", "b"),

    # Single consonants
    (r"c", "tʃ"),
    (r"j", "dʒ"),
    (r"y", "j"),
    (r"q", "ʔ"),
    (r"x", "ks"),
    (r"'", "ʔ"),

    # Vowels — keep most as-is since Latin orthography is close to IPA
    # for Austronesian languages, but handle a few specifics
    (r"aa", "aː"),
    (r"ee", "eː"),
    (r"ii", "iː"),
    (r"oo", "oː"),
    (r"uu", "uː"),
]

# Compile the rules into a single pass-through regex + lookup
_RULE_PATTERNS = [(re.compile(pat, re.IGNORECASE), repl) for pat, repl in _AUSTRONESIAN_RULES]


def _austronesian_g2p(word: str) -> str:
    """Best-effort rule-based orthographic-to-IPA for Austronesian words.

    Applies substitution rules in priority order (longest match first),
    then passes through any remaining characters unchanged.
    """
    text = word.lower().strip()
    if not text:
        return ""

    # Apply rules iteratively: scan left-to-right, try longest match first
    result: list[str] = []
    i = 0
    while i < len(text):
        matched = False
        for pat, repl in _RULE_PATTERNS:
            m = pat.match(text, i)
            if m:
                result.append(repl)
                i = m.end()
                matched = True
                break
        if not matched:
            # Pass through: most Austronesian letters map trivially
            ch = text[i]
            # Skip non-alphabetic (hyphens, spaces, digits, etc.)
            if ch.isalpha():
                result.append(ch)
            i += 1

    return "".join(result)


def g2p(word: str) -> str:
    """Convert an orthographic word to IPA using the best available backend."""
    if _USE_TRANSPHONE:
        ipa = _transphone_g2p(word)
        if ipa:
            return ipa
    # Fall back to rule-based Austronesian G2P
    return _austronesian_g2p(word)


# ---------------------------------------------------------------------------
# Main fix logic
# ---------------------------------------------------------------------------

def fix_file(tsv_path: Path) -> tuple[int, int, int]:
    """Fix fake-IPA entries in a single TSV file.

    Returns (total_abvd_fake, fixed, failed).
    """
    rows: list[dict[str, str]] = []
    total_abvd_fake = 0
    fixed = 0
    failed = 0

    with open(tsv_path, encoding="utf-8", newline="") as f:
        reader = csv.DictReader(f, delimiter="\t")
        if reader.fieldnames is None:
            return 0, 0, 0
        for row in reader:
            source = row.get("Source", "")
            word = row.get("Word", "")
            ipa = row.get("IPA", "")

            if "abvd" in source and word == ipa and word:
                total_abvd_fake += 1
                new_ipa = g2p(word)
                if new_ipa and new_ipa != word:
                    row["IPA"] = new_ipa
                    row["SCA"] = ipa_to_sound_class(new_ipa)
                    fixed += 1
                else:
                    failed += 1

            rows.append(row)

    if fixed == 0:
        # Nothing changed — don't rewrite the file
        return total_abvd_fake, fixed, failed

    # Rewrite the file preserving exact column order
    with open(tsv_path, "w", encoding="utf-8", newline="") as f:
        f.write(HEADER)
        for row in rows:
            line = "\t".join([
                row.get("Word", ""),
                row.get("IPA", ""),
                row.get("SCA", ""),
                row.get("Source", ""),
                row.get("Concept_ID", ""),
                row.get("Cognate_Set_ID", ""),
            ])
            f.write(line + "\n")

    return total_abvd_fake, fixed, failed


def main() -> None:
    print("=" * 70)
    print("fix_abvd_ipa: Fix ABVD entries where Word == IPA")
    print("=" * 70)

    if _USE_TRANSPHONE:
        print("  G2P backend: transphone")
    else:
        print("  G2P backend: rule-based Austronesian fallback")

    if not LEXICON_DIR.exists():
        print(f"  ERROR: lexicon directory not found: {LEXICON_DIR}")
        sys.exit(1)

    tsv_files = sorted(LEXICON_DIR.glob("*.tsv"))
    print(f"  Lexicon directory: {LEXICON_DIR}")
    print(f"  TSV files found:   {len(tsv_files)}")
    print()

    total_files_scanned = 0
    total_files_modified = 0
    grand_total_fake = 0
    grand_total_fixed = 0
    grand_total_failed = 0

    for tsv_path in tsv_files:
        total_files_scanned += 1
        abvd_fake, fixed, failed = fix_file(tsv_path)

        grand_total_fake += abvd_fake
        grand_total_fixed += fixed
        grand_total_failed += failed

        if fixed > 0:
            total_files_modified += 1
            print(f"  {tsv_path.name}: {fixed} fixed, {failed} failed (of {abvd_fake} fake-IPA)")

    print()
    print("-" * 70)
    print("SUMMARY")
    print("-" * 70)
    print(f"  Files scanned:     {total_files_scanned}")
    print(f"  Files modified:    {total_files_modified}")
    print(f"  ABVD fake-IPA:     {grand_total_fake}")
    print(f"  Entries fixed:     {grand_total_fixed}")
    print(f"  Entries failed:    {grand_total_failed}")

    if grand_total_fake > 0:
        pct = grand_total_fixed / grand_total_fake * 100
        print(f"  Fix rate:          {pct:.1f}%")

    print("Done.")


if __name__ == "__main__":
    main()