src/ls_import_ws.py

# /// script
# requires-python = ">=3.9"
# dependencies = ["python-crfsuite"]
# ///
"""Convert AL Cycle 0 top-500 TSV + BIO files to Label Studio JSON.

Reads al_cycle0_top500.tsv to identify which sentences to annotate,
looks up their silver BIO tags from udd-ws-v1.1-{dev,test}.txt,
and produces a Label Studio import JSON with pre-annotations.

When --model is provided, computes per-span confidence scores from CRF
marginal probabilities. Each word span gets score = min confidence across
its syllables, so words with any uncertain boundary get a low score.

When --dict is provided (standalone or auto-detected from model dir),
adds meta.dict to each span (shown in Region Details when selected).

Usage:
    uv run src/ls_import_ws.py [--validate]
    uv run src/ls_import_ws.py --dict path/to/dictionary.txt
    uv run src/ls_import_ws.py --model path/to/model.crfsuite
"""

import argparse
import json
import sys
import unicodedata
from pathlib import Path


def parse_bio_file(filepath):
    """Parse BIO file into list of sentences.

    Returns list of dicts with keys: sent_id, text, syllables, tags.
    Reused from fix_ws_errors.py.
    """
    sentences = []
    current = {"sent_id": "", "text": "", "syllables": [], "tags": []}

    with open(filepath, "r", encoding="utf-8") as f:
        for line in f:
            line = line.rstrip("\n")
            if line.startswith("# sent_id = "):
                current["sent_id"] = line.split("= ", 1)[1]
                continue
            if line.startswith("# text = "):
                current["text"] = line.split("= ", 1)[1]
                continue
            if line.startswith("#"):
                continue
            if not line:
                if current["syllables"]:
                    sentences.append(dict(current))
                    current = {"sent_id": "", "text": "", "syllables": [], "tags": []}
                continue
            parts = line.split("\t")
            if len(parts) == 2:
                current["syllables"].append(parts[0])
                current["tags"].append(parts[1])

    if current["syllables"]:
        sentences.append(dict(current))

    return sentences


def parse_tsv(filepath):
    """Parse al_cycle0_top500.tsv.

    Returns list of dicts with keys: rank, file, sent_idx, score, text.
    """
    rows = []
    with open(filepath, "r", encoding="utf-8") as f:
        header = f.readline().rstrip("\n").split("\t")
        for line in f:
            line = line.rstrip("\n")
            if not line:
                continue
            fields = line.split("\t")
            row = dict(zip(header, fields))
            rows.append({
                "rank": int(row["rank"]),
                "file": row["file"],
                "sent_idx": int(row["sent_idx"]),
                "score": float(row["score"]),
                "text": row["text"],
            })
    return rows


def nfc(text):
    """Normalize text to NFC for consistent Unicode handling."""
    return unicodedata.normalize("NFC", text)


def confidence_to_label(conf):
    """Map confidence score to label: WH (>=0.8), WM (0.6-0.8), WL (<0.6)."""
    if conf >= 0.8:
        return "WH"
    elif conf >= 0.6:
        return "WM"
    else:
        return "WL"


def bio_to_spans(syllables, tags, text, confidences=None):
    """Convert BIO tags to Label Studio span annotations.

    Args:
        syllables: list of syllable strings
        tags: list of BIO tags (B-W / I-W)
        text: space-joined text string
        confidences: optional list of per-syllable confidence floats

    Returns:
        list of span dicts for Label Studio predictions
    """
    spans = []
    span_start = None
    span_syllables = []
    span_syl_indices = []
    char_pos = 0

    def _close_span():
        span_text = " ".join(span_syllables)
        # Choose label based on confidence: WH/WM/WL for predictions, W if no confidence
        if confidences is not None:
            span_conf = min(confidences[j] for j in span_syl_indices)
            label = confidence_to_label(span_conf)
        else:
            label = "W"
        span_dict = {
            "id": f"s{len(spans)}",
            "from_name": "label",
            "to_name": "text",
            "type": "labels",
            "value": {
                "start": span_start,
                "end": span_start + len(span_text),
                "text": span_text,
                "labels": [label],
            },
        }
        if confidences is not None:
            span_dict["score"] = round(span_conf, 4)
        spans.append(span_dict)

    for i, (syl, tag) in enumerate(zip(syllables, tags)):
        # Verify syllable matches text at expected position
        expected = text[char_pos:char_pos + len(syl)]
        if nfc(expected) != nfc(syl):
            raise ValueError(
                f"Syllable mismatch at pos {char_pos}: "
                f"expected {syl!r}, got {expected!r} in text {text!r}"
            )

        if tag == "B-W":
            # Close previous span
            if span_syllables:
                _close_span()
            span_start = char_pos
            span_syllables = [syl]
            span_syl_indices = [i]
        elif tag == "I-W":
            span_syllables.append(syl)
            span_syl_indices.append(i)
        else:
            raise ValueError(f"Unknown tag {tag!r} for syllable {syl!r}")

        char_pos += len(syl)
        # Skip space between syllables
        if i < len(syllables) - 1:
            char_pos += 1  # space separator

    # Close final span
    if span_syllables:
        _close_span()

    return spans


def add_dict_meta(spans, dictionary):
    """Add dictionary lookup metadata to each span.

    When the annotator selects a region, LS shows the meta fields in
    the Region Details panel. This provides instant dictionary feedback.

    For multi-syllable spans: shows dict status + up to 5 similar entries.
    For single-syllable spans: shows dict status only.
    """
    for span in spans:
        span_text = span["value"]["text"]
        normalized = nfc(span_text.lower().strip())
        n_syllables = span_text.count(" ") + 1
        in_dict = normalized in dictionary

        if n_syllables == 1:
            span["meta"] = {"dict": "✓" if in_dict else "—"}
        else:
            if in_dict:
                span["meta"] = {"dict": "✓ in dict"}
            else:
                # Find similar entries (prefix match on first syllable)
                first_syl = normalized.split()[0]
                similar = sorted(
                    e for e in dictionary
                    if e.startswith(first_syl + " ") and e != normalized
                )[:5]
                meta = {"dict": "✗ not found"}
                if similar:
                    meta["similar"] = ", ".join(similar)
                span["meta"] = meta


def validate_spans(spans, text, syllables):
    """Validate that spans cover every character in text without gaps/overlaps."""
    text_len = len(text)
    covered = [False] * text_len

    for span in spans:
        start = span["value"]["start"]
        end = span["value"]["end"]
        for j in range(start, end):
            if covered[j]:
                raise ValueError(f"Overlap at char {j} in text {text!r}")
            covered[j] = True

    # Check that all non-space characters are covered
    for j, ch in enumerate(text):
        if ch != " " and not covered[j]:
            raise ValueError(f"Gap at char {j} ({ch!r}) in text {text!r}")

    # Check span count matches word count
    n_words = sum(1 for t in syllables if t == "B-W")  # tags not syllables
    # Actually count from spans
    if len(spans) != n_words:
        raise ValueError(
            f"Span count {len(spans)} != word count {n_words} in text {text!r}"
        )


def load_dictionary(dict_path):
    """Load dictionary file into a set of NFC-normalized entries.

    Returns set of lowercase NFC-normalized dictionary entries.
    """
    dictionary = set()
    with open(dict_path, encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if line:
                dictionary.add(nfc(line.lower()))
    print(f"Dictionary: {len(dictionary)} entries from {dict_path}")
    return dictionary


def load_crf_scorer(model_path):
    """Load CRF tagger for confidence scoring.

    Returns tagger object.
    """
    import pycrfsuite

    tagger = pycrfsuite.Tagger()
    tagger.open(str(model_path))
    return tagger


def compute_confidences(tagger, syllables, dictionary=None):
    """Compute per-syllable confidence from CRF marginal probabilities.

    Confidence = max(P(B|x,i), P(I|x,i)) for each syllable position.
    Uses feature extraction from al_score_ws.py.
    """
    # Import feature extraction (same as al_score_ws.py)
    sys.path.insert(0, str(Path(__file__).resolve().parent))
    from al_score_ws import sentence_to_features

    xseq = sentence_to_features(syllables, dictionary)
    tagger.set(xseq)

    confidences = []
    for i in range(len(syllables)):
        p_b = tagger.marginal("B", i)
        p_i = tagger.marginal("I", i)
        confidences.append(max(p_b, p_i))

    return confidences


def main():
    parser = argparse.ArgumentParser(
        description="Convert AL top-500 + BIO to Label Studio JSON"
    )
    parser.add_argument(
        "--tsv",
        default="al_cycle0_top500.tsv",
        help="Path to ranked TSV file (default: al_cycle0_top500.tsv)",
    )
    parser.add_argument(
        "--bio-dir",
        default=".",
        help="Directory containing udd-ws-v1.1-{dev,test}.txt (default: .)",
    )
    parser.add_argument(
        "--output",
        default="ls_import_cycle1.json",
        help="Output JSON path (default: ls_import_cycle1.json)",
    )
    parser.add_argument(
        "--validate",
        action="store_true",
        help="Run validation checks on all tasks",
    )
    parser.add_argument(
        "--model",
        default=None,
        help="Path to CRF model (.crfsuite) for per-span confidence scores",
    )
    parser.add_argument(
        "--dict",
        default=None,
        help="Path to dictionary.txt (auto-detected from model dir if not set)",
    )
    args = parser.parse_args()

    root = Path(args.bio_dir)

    # Load dictionary (standalone or auto-detected from model dir)
    dictionary = None
    if args.dict:
        dictionary = load_dictionary(args.dict)
    elif args.model:
        auto_dict = Path(args.model).parent / "dictionary.txt"
        if auto_dict.exists():
            dictionary = load_dictionary(auto_dict)

    # Load CRF scorer if model provided
    tagger = None
    if args.model:
        model_path = Path(args.model)
        print(f"Loading CRF model: {model_path}")
        tagger = load_crf_scorer(model_path)
        print("Per-span confidence scoring enabled")

    # Parse TSV
    tsv_rows = parse_tsv(args.tsv)
    print(f"Loaded {len(tsv_rows)} rows from {args.tsv}")

    # Parse BIO files and build text-based index
    bio_by_text = {}
    for split in ("dev", "test", "train"):
        bio_path = root / f"udd-ws-v1.1-{split}.txt"
        if not bio_path.exists():
            continue
        sentences = parse_bio_file(bio_path)
        print(f"Loaded {len(sentences)} sentences from {bio_path}")
        for sent in sentences:
            key = nfc(" ".join(sent["syllables"]))
            bio_by_text[key] = sent

    print(f"Text index: {len(bio_by_text)} unique sentences")

    # Build tasks
    tasks = []
    errors = []

    for row in tsv_rows:
        tsv_text = nfc(row["text"])

        sent = bio_by_text.get(tsv_text)
        if sent is None:
            errors.append(f"Rank {row['rank']}: text not found in BIO files")
            continue

        text = tsv_text

        # Compute per-syllable confidence if CRF model available
        confidences = None
        if tagger is not None:
            confidences = compute_confidences(
                tagger, sent["syllables"], dictionary
            )

        # Convert BIO to spans
        try:
            spans = bio_to_spans(
                sent["syllables"], sent["tags"], text, confidences
            )
        except ValueError as e:
            errors.append(f"Rank {row['rank']}: span conversion error: {e}")
            continue

        # Validate span coverage
        if args.validate:
            try:
                validate_spans(spans, text, sent["tags"])
            except ValueError as e:
                errors.append(f"Rank {row['rank']}: validation error: {e}")
                continue

        # Task-level prediction score = 1 - AL uncertainty score
        pred_score = round(1.0 - row["score"], 6)

        # Add dictionary metadata to spans
        if dictionary:
            add_dict_meta(spans, dictionary)

        task = {
            "data": {
                "text": text,
                "sent_id": sent["sent_id"],
                "rank": row["rank"],
            },
            "predictions": [{
                "model_version": "silver_crf_v1.1",
                "score": pred_score,
                "result": spans,
            }],
        }
        tasks.append(task)

    # Report
    if errors:
        print(f"\n{len(errors)} errors:")
        for e in errors:
            print(f"  - {e}")

    print(f"\nGenerated {len(tasks)} tasks")

    if tagger is not None:
        # Report confidence statistics
        all_scores = []
        for t in tasks:
            for r in t["predictions"][0]["result"]:
                if "score" in r:
                    all_scores.append(r["score"])
        if all_scores:
            low = sum(1 for s in all_scores if s < 0.8)
            print(f"Span confidence: min={min(all_scores):.4f}, "
                  f"mean={sum(all_scores)/len(all_scores):.4f}, "
                  f"<0.8: {low}/{len(all_scores)}")

    # Write output
    output_path = Path(args.output)
    with open(output_path, "w", encoding="utf-8") as f:
        json.dump(tasks, f, ensure_ascii=False, indent=2)
    print(f"Written to {output_path}")

    if errors:
        sys.exit(1)


if __name__ == "__main__":
    main()