creation_script.py

import json
import os
from collections import defaultdict
from pathlib import Path
from typing import List, Union, Tuple, Set, Any

from bs4 import BeautifulSoup, Tag
from tqdm import tqdm

NKJP_PATH = "NKJP-PodkorpusMilionowy-1.2"


def get_split() -> Tuple[Set[str], Set[str]]:
    with open("data/split/train.txt", "r") as f:
        train = list(f.readlines())
    with open("data/split/test.txt", "r") as f:
        test = list(f.readlines())
    train = [x.strip() for x in train]
    test = [x.strip() for x in test]
    train_set = set(train)
    test_set = set(test)
    assert len(train_set) == len(train)
    assert len(test_set) == len(test)
    return train_set, test_set


def parse_sentence(sentence_tag: Tag) -> dict[str, Any]:
    sentence = defaultdict(list)
    for seg in sentence_tag.find_all("seg"):  # słowo
        [f_orth] = seg.find_all("f", attrs={"name": "orth"})
        sentence["tokens"].append(f_orth.getText().strip())
        [f_orth] = seg.find_all("f", attrs={"name": "disamb"})
        sentence["pos_tags"].append(f_orth.getText().strip().split(":")[1])
    assert len(sentence["tokens"]) == len(sentence["pos_tags"])
    return dict(sentence)


def parse_tei_file(path: Path) -> List[dict[str, Union[List[str], str]]]:
    with open(path, "r") as tei:
        soup = BeautifulSoup(tei, "lxml")

    result = []
    for p in soup.find_all("p"):
        for s in p.find_all("s"):
            example = parse_sentence(s)
            example["id"] = f"{path.parent.name}_{s['xml:id']}"
            result.append(example)
    return result


train_names, test_names = get_split()
train, test = [], []
for entry in tqdm(list(os.scandir(NKJP_PATH))):
    if entry.is_dir():
        file_data = parse_tei_file(Path(entry.path) / "ann_morphosyntax.xml")
        if entry.name in train_names:
            train += file_data
        elif entry.name in test_names:
            test += file_data
        else:
            raise ValueError(f"Couldn't find file in splits: {entry.name}")

with open("data/train.jsonl", "w") as f:
    for item in train:
        f.write(json.dumps(item, ensure_ascii=False) + "\n")
with open("data/test.jsonl", "w") as f:
    for item in test:
        f.write(json.dumps(item, ensure_ascii=False) + "\n")