dataset/clean_dataset.py

import csv
import re
import unicodedata
from pathlib import Path

# ── Paths ──────────────────────────────────────────────────────────────────────
BASE = Path(__file__).parent
INPUT_FILES = {
    "train": BASE / "train.csv",
    "test":  BASE / "test.csv",
}
OUTPUT_FILES = {
    "train": BASE / "train_cleaned.csv",
    "test":  BASE / "test_cleaned.csv",
}

# ── Metadata patterns to strip ─────────────────────────────────────────────────
METADATA_PATTERNS = [
    r"Pass\s*Key\s*:\s*\d+",
    r"PRINTED\s+BY[\w\s,\.]*",
    r"Report\s+Date\s*:\s*[\d\-/]+\s*[\d:]+\s*[AP]M",
    r"THIS\s+REPORT\s+HAS\s+BEEN[^.]*\.",
    r"Reg\.?\s*No\.?\s*:\s*[\w/\-]+",
    r"Thanks\s+for\s+reference[\w\s,\.]*",
    r"={3,}.*",                          # === separators and everything after
    r"\*{3,}.*",                         # *** separators
    r"DISCLAIMER[^.]*\.",
    r"This\s+report\s+is\s+confidential[^.]*\.",
    r"For\s+clinical\s+use\s+only[^.]*\.",
]

# ── Word formation fixes (merged words) ───────────────────────────────────────
WORD_FIXES = {
    r"foraminalnarrowing":          "foraminal narrowing",
    r"papyraceaon":                 "papyracea on",
    r"Bilateralfronto":             "Bilateral fronto",
    r"bilateralfronto":             "bilateral fronto",
    r"Ethmoidappears":              "Ethmoid appears",
    r"ethmoidappears":              "ethmoid appears",
    r"thislevel":                   "this level",
    r"\bL(\d)L(\d)\b":             r"L\1-L\2",   # L3L4 → L3-L4
    r"\bL(\d)S(\d)\b":             r"L\1-S\2",   # L5S1 → L5-S1
}

# ── Non-ASCII normalisation map ────────────────────────────────────────────────
CHAR_MAP = {
    "×": "x",    # × multiplication sign
    "–": "-",    # – en-dash
    "—": "-",    # — em-dash
    "’": "'",    # ' right single quote
    "‘": "'",    # ' left single quote
    "“": '"',    # " left double quote
    "”": '"',    # " right double quote
    "…": "...",  # … ellipsis
    "°": " degrees",  # ° degree sign
}

# ── HTML / URL strip ───────────────────────────────────────────────────────────
HTML_PATTERN   = re.compile(r"<[^>]+>")
URL_PATTERN    = re.compile(r"https?://\S+|www\.\S+")
MULTI_SPACE    = re.compile(r"  +")
MULTI_DOT      = re.compile(r"\.{2,}")
MULTI_SLASH    = re.compile(r"/{2,}")


def strip_metadata(text: str) -> str:
    for pat in METADATA_PATTERNS:
        text = re.sub(pat, "", text, flags=re.IGNORECASE)
    return text


def fix_word_formation(text: str) -> str:
    for pat, replacement in WORD_FIXES.items():
        text = re.sub(pat, replacement, text)
    return text


def normalise_chars(text: str) -> str:
    for char, replacement in CHAR_MAP.items():
        text = text.replace(char, replacement)
    # drop any remaining non-ASCII that slipped through
    text = text.encode("ascii", errors="ignore").decode("ascii")
    return text


def clean_text(text: str) -> str:
    text = strip_metadata(text)
    text = HTML_PATTERN.sub("", text)
    text = URL_PATTERN.sub("", text)
    text = fix_word_formation(text)
    text = normalise_chars(text)
    text = MULTI_DOT.sub(".", text)
    text = MULTI_SLASH.sub("/", text)
    text = MULTI_SPACE.sub(" ", text)
    text = text.lower()
    return text.strip()


def process(split: str):
    src  = INPUT_FILES[split]
    dst  = OUTPUT_FILES[split]

    stats = {
        "total":           0,
        "duplicates":      0,
        "metadata_fixed":  0,
        "word_fix":        0,
        "char_normalised": 0,
        "too_short":       0,
        "too_long":        0,
        "kept":            0,
    }

    seen_transcriptions: set[str] = set()
    rows_out: list[dict] = []
    orphaned_audio: list[Path] = []

    with src.open(newline="", encoding="utf-8") as f:
        reader = csv.DictReader(f)
        rows = list(reader)

    stats["total"] = len(rows)

    for row in rows:
        original = row["transcription"]
        cleaned  = clean_text(original)

        # track what changed
        if cleaned != original:
            if any(re.search(p, original, re.IGNORECASE) for p in METADATA_PATTERNS):
                stats["metadata_fixed"] += 1
            if any(re.search(p, original) for p in WORD_FIXES):
                stats["word_fix"] += 1
            if any(c in original for c in CHAR_MAP):
                stats["char_normalised"] += 1

        word_count = len(cleaned.split())

        # skip empty after cleaning
        if not cleaned:
            orphaned_audio.append(BASE / row["file_name"])
            continue

        # skip duplicates (keep first occurrence)
        if cleaned in seen_transcriptions:
            stats["duplicates"] += 1
            orphaned_audio.append(BASE / row["file_name"])
            continue
        seen_transcriptions.add(cleaned)

        # flag but keep short / long (let user decide manually)
        if word_count < 5:
            stats["too_short"] += 1
        if word_count > 300:
            stats["too_long"] += 1

        rows_out.append({"file_name": row["file_name"], "transcription": cleaned})
        stats["kept"] += 1

    with dst.open("w", newline="", encoding="utf-8") as f:
        writer = csv.DictWriter(f, fieldnames=["file_name", "transcription"])
        writer.writeheader()
        writer.writerows(rows_out)

    # delete audio files for removed rows
    deleted = 0
    missing = 0
    for audio_path in orphaned_audio:
        if audio_path.exists():
            audio_path.unlink()
            deleted += 1
        else:
            missing += 1

    print(f"\n{'='*55}")
    print(f"  {split.upper()} -- {src.name}  ->  {dst.name}")
    print(f"{'='*55}")
    print(f"  Total rows          : {stats['total']}")
    print(f"  Duplicates removed  : {stats['duplicates']}")
    print(f"  Metadata stripped   : {stats['metadata_fixed']}")
    print(f"  Word formation fixed: {stats['word_fix']}")
    print(f"  Chars normalised    : {stats['char_normalised']}")
    print(f"  Short (<5 words)    : {stats['too_short']}  (kept, review manually)")
    print(f"  Long (>300 words)   : {stats['too_long']}  (kept, review manually)")
    print(f"  Final rows kept     : {stats['kept']}")
    print(f"  Audio files deleted : {deleted}")
    if missing:
        print(f"  Audio not found     : {missing}  (already missing)")
    print(f"{'='*55}")


if __name__ == "__main__":
    print("Cleaning dataset — originals unchanged.")
    process("train")
    process("test")
    print("\nDone. Cleaned files saved.")