scripts/stoicheia_parse_hypotactic.py

#!/usr/bin/env python3
"""Parse reconstructed Hypotactic sentences and crop their trees to verse lines."""

from __future__ import annotations

import argparse
from collections import defaultdict
import json
import os
from pathlib import Path
import re
import sys
import time
import unicodedata

import torch
from huggingface_hub import snapshot_download
from transformers import AutoModel

try:
    from scripts.build_dataset import (
        curated_hypotactic_line,
        crop_conllu_to_normalized_span,
        hypotactic_line_key,
        load_hypotactic,
        normalize,
    )
except ModuleNotFoundError:
    from build_dataset import (
        curated_hypotactic_line,
        crop_conllu_to_normalized_span,
        hypotactic_line_key,
        load_hypotactic,
        normalize,
    )


MODEL_ID = "Ericu950/Stoicheia-tagger-parser"
MODEL_REVISION = "cd8ae1658c364874c3b6f4df37bd1d46313e3cea"
SENTENCE_FINAL = re.compile(r"[.!?;\u037e][\]\)}\u00bb\u2019\u201d\"']*$")
MAX_WORDS = 350
MAX_CHARS = 1900


def retained_line(stem: str, line: dict) -> dict | None:
    curated, reason = curated_hypotactic_line(stem, line)
    if reason:
        return None
    return {
        "key": list(hypotactic_line_key(line)),
        "stem": stem,
        "line_sequence": int(line["line_sequence"]),
        "poem_sequence": line["poem_sequence"],
        "book": line["book"],
        **curated,
    }


def _capped_sentence(entries: list[dict]) -> list[tuple[list[dict], bool]]:
    """Split an overlong punctuation sentence only at metrical-line boundaries."""
    chunks = []
    start = 0
    while start < len(entries):
        words = chars = 0
        end = start
        last_line_boundary = None
        while end < len(entries):
            token = entries[end]["token"]
            next_words = words + 1
            next_chars = chars + len(token)
            if next_words > MAX_WORDS or next_chars > MAX_CHARS:
                break
            words, chars = next_words, next_chars
            end += 1
            if end == len(entries) or entries[end]["key"] != entries[end - 1]["key"]:
                last_line_boundary = end
        if end == len(entries):
            chunks.append((entries[start:end], start > 0))
            break
        if last_line_boundary is None or last_line_boundary <= start:
            raise ValueError("one metrical line exceeds Stoicheia's safe context window")
        chunks.append((entries[start:last_line_boundary], True))
        start = last_line_boundary
    return chunks


def reconstruct_sentences(retained: list[dict]) -> list[dict]:
    """Create punctuation-delimited sentences across consecutive metrical lines."""
    sentences = []
    current = []
    previous_group = None
    previous_sequence = None

    def finish() -> None:
        nonlocal current
        if not current:
            return
        if not any(normalize(entry["token"]) for entry in current):
            # Standalone editorial punctuation may follow a sentence-final word.
            # Stoicheia intentionally emits no word row for such a segment.
            current = []
            return
        for entries, forced in _capped_sentence(current):
            sentence_number = len(sentences) + 1
            spans = []
            cursor = 0
            for entry in entries:
                length = len(normalize(entry["token"]))
                if not length:
                    # Editorial punctuation contributes no normalized text and
                    # therefore cannot define a non-empty line crop.
                    continue
                if spans and spans[-1]["key"] == entry["key"]:
                    spans[-1]["end"] += length
                else:
                    spans.append({
                        "key": entry["key"], "start": cursor, "end": cursor + length,
                    })
                cursor += length
            if not cursor:
                raise ValueError("reconstructed sentence has no Greek token content")
            sentences.append({
                "sentence_id": f"hypotactic-context-{sentence_number:06d}",
                "tokens": [entry["token"] for entry in entries],
                "text": " ".join(entry["token"] for entry in entries),
                "spans": spans,
                "line_count": len({tuple(span["key"]) for span in spans}),
                "forced_context_split": forced,
            })
        current = []

    for line in retained:
        group = (line["stem"], line["poem_sequence"], line["book"])
        if (
            previous_group is not None
            and (group != previous_group or line["line_sequence"] != previous_sequence + 1)
        ):
            finish()
        for token in line["tokens"]:
            current.append({"key": line["key"], "token": token})
            if SENTENCE_FINAL.search(token):
                finish()
        previous_group = group
        previous_sequence = line["line_sequence"]
    finish()
    return sentences


def repair_tree(words: list[dict]) -> list[dict]:
    """Make the greedy biaffine decode a single rooted, acyclic UD tree."""
    if not words:
        raise ValueError("Stoicheia decoded no Greek tokens")
    n = len(words)
    original = [(int(word.get("head") or 0), word.get("deprel")) for word in words]
    heads = [int(word.get("head") or 0) for word in words]
    roots = [
        index for index, (word, head) in enumerate(zip(words, heads), 1)
        if head == 0 and word["upos"] != "PUNCT"
    ]
    primary = roots[0] if roots else next(
        (index for index, word in enumerate(words, 1) if word["upos"] != "PUNCT"), 1,
    )
    for index, word in enumerate(words, 1):
        head = heads[index - 1]
        if index == primary:
            heads[index - 1] = 0
            word["deprel"] = "root"
        elif head < 0 or head > n or head in {0, index}:
            heads[index - 1] = primary
            word["deprel"] = "punct" if word["upos"] == "PUNCT" else "dep"
        elif word.get("deprel") == "root":
            word["deprel"] = "dep"

    changed = True
    while changed:
        changed = False
        for token_id in range(1, n + 1):
            trail = []
            cursor = token_id
            while cursor:
                if cursor in trail:
                    cycle = trail[trail.index(cursor):]
                    break_id = min(cycle)
                    heads[break_id - 1] = 0 if break_id == primary else primary
                    words[break_id - 1]["deprel"] = (
                        "root" if break_id == primary else "dep"
                    )
                    changed = True
                    break
                trail.append(cursor)
                cursor = heads[cursor - 1]
            if changed:
                break
    for word, head, (old_head, old_deprel) in zip(words, heads, original):
        word["head"] = head
        word["head_repair"] = head != old_head or word.get("deprel") != old_deprel
    return words


def restore_alphabetic_non_greek_tokens(
    tokens: list[str], words: list[dict],
) -> list[dict]:
    """Restore alphabetic tokens the character model intentionally omits."""
    old_to_new = {}
    merged = []
    decoded_index = 0
    for token in tokens:
        has_greek = any("GREEK" in unicodedata.name(char, "") for char in token)
        if has_greek:
            if decoded_index >= len(words) or words[decoded_index]["form"] != token:
                raise ValueError("Stoicheia form order differs from Hypotactic tokens")
            decoded_index += 1
            old_to_new[decoded_index] = len(merged) + 1
            merged.append(dict(words[decoded_index - 1]))
        elif normalize(token):
            merged.append({
                "form": token, "lemma": token, "upos": "X", "xpos": "x--------",
                "feats": {}, "head": 0, "deprel": "dep", "_restored": True,
            })
    if decoded_index != len(words):
        raise ValueError("Stoicheia returned unexpected extra forms")
    for word in merged:
        if word.pop("_restored", False):
            continue
        if word.get("head"):
            word["head"] = old_to_new[word["head"]]
    return merged


def encode_conllu(text: str, tokens: list[str], words: list[dict]) -> str:
    words = restore_alphabetic_non_greek_tokens(tokens, words)
    words = repair_tree(words)
    lines = [f"# text = {text}"]
    for index, word in enumerate(words, 1):
        feats = word.get("feats") or {}
        feat_text = "|".join(f"{key}={feats[key]}" for key in sorted(feats)) or "_"
        lines.append("\t".join([
            str(index), word["form"], word.get("lemma") or "_",
            word.get("upos") or "X", word.get("xpos") or "_", feat_text,
            str(word["head"]), word.get("deprel") or "dep", "_",
            "HeadRepair=Yes" if word["head_repair"] else "_",
        ]))
    conllu = "\n".join(lines) + "\n\n"
    forms = "".join(word["form"] for word in words)
    if normalize(forms) != normalize(text):
        raise ValueError("Stoicheia forms do not cover the Hypotactic line")
    return conllu


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--hypotactic", type=Path, required=True)
    parser.add_argument("--output", type=Path, required=True)
    parser.add_argument("--batch-size", type=int, default=64)
    parser.add_argument("--log-every", type=int, default=1024)
    args = parser.parse_args()

    all_lines = load_hypotactic(args.hypotactic, all_files=True)
    retained = [
        row
        for stem, lines in sorted(all_lines.items())
        for line in lines
        if (row := retained_line(stem, line)) is not None
    ]
    keys = [tuple(row["key"]) for row in retained]
    if len(keys) != len(set(keys)):
        raise ValueError("Hypotactic line keys are not unique")
    sentences = reconstruct_sentences(retained)

    args.output.parent.mkdir(parents=True, exist_ok=True)
    print(
        f"inventory retained_lines={len(retained)} reconstructed_sentences={len(sentences)} "
        f"multiline_sentences={sum(row['line_count'] > 1 for row in sentences)} "
        f"forced_context_splits={sum(row['forced_context_split'] for row in sentences)} "
        f"model={MODEL_ID}@{MODEL_REVISION}",
        flush=True,
    )

    local = snapshot_download(
        MODEL_ID, revision=MODEL_REVISION,
        allow_patterns=["*.json", "*.txt", "*.py", "*.model", "*.safetensors"],
    )
    sys.path.insert(0, local)
    from processing_char_bert_joint import CharBertJointProcessor

    device = torch.device("cuda")
    model = AutoModel.from_pretrained(
        local, trust_remote_code=True, dtype=torch.bfloat16,
    ).to(device).eval()
    processor = CharBertJointProcessor.from_pretrained(local)
    print(
        f"loaded model device={device} dtype={next(model.parameters()).dtype} "
        f"batch_size={args.batch_size}",
        flush=True,
    )

    started = time.monotonic()
    fragments = defaultdict(list)
    temporary = args.output.with_suffix(args.output.suffix + ".incomplete")
    with temporary.open("w", encoding="utf-8", buffering=1) as handle:
        for start in range(0, len(sentences), args.batch_size):
            items = sentences[start:start + args.batch_size]
            batch = processor([item["tokens"] for item in items])
            model_batch = {
                key: value.to(device, non_blocking=True)
                for key, value in batch.items() if not key.startswith("_")
            }
            with torch.inference_mode():
                output = model(**model_batch)
            decoded = processor.decode(output, batch, ud=True)
            if len(decoded) != len(items):
                raise RuntimeError("Stoicheia returned the wrong batch cardinality")
            for item, words in zip(items, decoded):
                sentence_conllu = encode_conllu(item["text"], item["tokens"], words)
                for span in item["spans"]:
                    fragments[tuple(span["key"])].append({
                        "conllu": crop_conllu_to_normalized_span(
                            sentence_conllu, span["start"], span["end"],
                        ),
                        "sentence_id": item["sentence_id"],
                        "line_count": item["line_count"],
                        "forced_context_split": item["forced_context_split"],
                    })
            done = start + len(items)
            if done == len(sentences) or done % args.log_every < args.batch_size:
                elapsed = time.monotonic() - started
                rate = done / elapsed if elapsed else 0
                print(
                    f"progress sentences={done}/{len(sentences)} "
                    f"rate={rate:.1f}_sentences_s elapsed={elapsed:.1f}s",
                    flush=True,
                )
        if set(fragments) != set(keys):
            raise RuntimeError(
                f"sentence cropping did not cover every retained line: "
                f"missing={len(set(keys) - set(fragments))} "
                f"extra={len(set(fragments) - set(keys))}"
            )
        retained_by_key = {tuple(row["key"]): row for row in retained}
        for key in keys:
            item = retained_by_key[key]
            pieces = fragments[key]
            conllu = "".join(piece["conllu"] for piece in pieces)
            result = {
                **{name: item[name] for name in ("key", "author", "work", "work_id", "text")},
                "conllu": conllu,
                "context_sentence_ids": [piece["sentence_id"] for piece in pieces],
                "context_sentence_line_counts": [piece["line_count"] for piece in pieces],
                "context_forced_splits": [piece["forced_context_split"] for piece in pieces],
                "context_parsing": "sentence_then_crop_v1",
                "model": MODEL_ID,
                "model_revision": MODEL_REVISION,
            }
            forms = "".join(
                columns[1]
                for line in conllu.splitlines()
                if not line.startswith("#") and line
                for columns in [line.split("\t")]
                if len(columns) == 10 and columns[0].isdigit()
            )
            if normalize(forms) != normalize(item["text"]):
                raise ValueError(f"cropped sentence trees do not cover line {key}")
            handle.write(json.dumps(result, ensure_ascii=False, sort_keys=True) + "\n")
        handle.flush()
        os.fsync(handle.fileno())
    os.replace(temporary, args.output)
    print(
        f"completed lines={len(keys)} sentence_fragments={sum(map(len, fragments.values()))} "
        f"output={args.output}", flush=True,
    )


if __name__ == "__main__":
    main()