data/ncc_parliament/create.py

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "datasets>=3.2.0"
# ]
# ///

import inspect
import logging
import re
from collections import defaultdict
from collections.abc import Callable
from datetime import datetime
from pathlib import Path

import pandas as pd
from datasets import Dataset, load_dataset

logger = logging.getLogger(__name__)
########## edit manually for each source
hf_path = "NbAiLab/NCC"
source = "ncc_parliament"
doc_type_searchword = "parliament"
license = "other"
domain = "Legal"
num_proc = 10
##########
today = datetime.now().strftime("%Y-%m-%d")

# stop words taken from spaCy
# https://github.com/explosion/spaCy/blob/master/spacy/lang/da/stop_words.py
# Source: Handpicked by Jens Dahl Møllerhøj.
spacy_sw = set(
    """
af aldrig alene alle allerede alligevel alt altid anden andet andre at

bag begge blandt blev blive bliver burde bør

da de dem den denne dens der derefter deres derfor derfra deri dermed derpå derved det dette dig din dine disse dog du

efter egen eller ellers en end endnu ene eneste enhver ens enten er et

flere flest fleste for foran fordi forrige fra få før først

gennem gjorde gjort god gør gøre gørende

ham han hans har havde have hel heller hen hende hendes henover her herefter heri hermed herpå hun hvad hvem hver hvilke hvilken hvilkes hvis hvor hvordan hvorefter hvorfor hvorfra hvorhen hvori hvorimod hvornår hvorved

i igen igennem ikke imellem imens imod ind indtil ingen intet

jeg jer jeres jo

kan kom kommer kun kunne

lad langs lav lave lavet lidt lige ligesom lille længere

man mange med meget mellem men mens mere mest mig min mindre mindst mine mit må måske

ned nemlig nogen nogensinde noget nogle nok nu ny nyt nær næste næsten

og også om omkring op os over overalt

på

samme sammen selv selvom senere ses siden sig sige skal skulle som stadig synes syntes så sådan således

temmelig tidligere til tilbage tit

ud uden udover under undtagen

var ved vi via vil ville vore vores vær være været

øvrigt
""".split()
)


def word_tokenize(text: str) -> list[str]:
    """
    Tokenizes a string into words, splitting on whitespace and punctuation.

    Example:
        >>> word_tokenize("Hello, world!")
        ['Hello', ',', 'world', '!']
        >>> word_tokenize("This is a test.")
        ['This', 'is', 'a', 'test', '.']
        >>> word_tokenize("Many spaces    between words.")
        ['Many', 'spaces', 'between', 'words', '.']
    """

    punkt = [",", ".", "!", "?", ":", ";", "(", ")", "[", "]", "{", "}", '"', "'"]
    for p in punkt:
        text = text.replace(p, f" {p} ")
    return text.split()


def alpha_ratio(text: str | list[str]) -> float:
    """
    If not split already to words, splits text with word_tokenize()
    Calculates ratio of words with only alphabetical characters

    """
    if type(text) is str:
        text = word_tokenize(text)
    else:
        pass

    alpha_ratio = 1 - sum(not word.isalpha() for word in text) / len(text)

    return alpha_ratio


def count_min_target(given_list: list, target_list: list, min: int) -> bool:
    """
    Iterates through given list, until at least min items match any items from target list

    """
    c_item = 0
    given_list_iter = iter(given_list)
    while c_item < min:
        try:
            current_item = next(given_list_iter)
            if current_item in target_list:
                c_item += 1
        except StopIteration:
            break

    return c_item == min


def dynaword_format(
    meta_document: dict[str, str | int],
) -> dict[str, str | dict[str, str]]:
    """Reformats data to fit dynaword standards"""

    text = meta_document.get("text")
    id = meta_document.get("id")
    date = meta_document.get("publish_year")
    doc_type = meta_document.get("doc_type")

    newdata = {
        "text": text,
        "source": source,
        "id": id,
        "added": today,
        "created": f"{date}-01-01, {date}-12-31",
        "license": license,
        "domain": domain,
        "metadata": {
            "source-pretty": f"Norwegian Colossal Corpus ({re.sub('ncc_', '', source)})",
            "source-type": doc_type,
        },
    }

    return newdata


def log_pre_filter_lang_data(
    lang_metadata: dict[str, dict[str, int]], filtered_ds: Dataset
):
    """
    Function for logging changes in a large dataset,
    based on the metadata pre filering and the filtered dataset,
    used for language filtering
    """
    all_docs = sum(lang_metadata[source].values())
    no_docs = lang_metadata[source].get("no")
    da_docs = lang_metadata[source].get("da")
    no_perc = round(no_docs / all_docs * 100, 4)
    da_perc = round(da_docs / all_docs * 100, 4)

    f_length = len(filtered_ds)
    f_perc = round(f_length / da_docs * 100, 4)
    f_total_perc = round(f_length / all_docs * 100, 4)

    logger.info(f"Documents of {source}:")
    logger.info(f"NO: {no_docs}, {no_perc}% ; DA: {da_docs}, {da_perc}%")
    logger.info("After language confidence filtering:")
    logger.info(f"DA: {f_length}, lost: {100 - f_perc}%")
    logger.info("Total document change:")
    logger.info(f"{all_docs} -> {f_length}, loss: {100 - f_total_perc}%")


def get_var_name(var):
    """outputs the variable name"""
    callers_local_vars = inspect.currentframe().f_back.f_back.f_back.f_locals.items()
    return [var_name for var_name, var_val in callers_local_vars if var_val is var]


def filter_with_changelog(
    filter_func: Callable[[Dataset], Dataset], dataset: Dataset
) -> Dataset:
    """
    Function, which takes a filter and a dataset.
    Counts text docs and tokens before and after filtering,
    Saves filtering changes to log.
    """

    filter_name = get_var_name(filter_func)
    pre_filter_docs = len(dataset)
    pre_filter_tokens = sum(len(word_tokenize(i["text"])) for i in dataset)

    dataset = dataset.filter(filter_func, num_proc=num_proc)

    post_filter_docs = len(dataset)
    post_filter_tokens = sum(len(word_tokenize(i["text"])) for i in dataset)
    tokens_removed = round((1 - (post_filter_tokens / pre_filter_tokens)) * 100, 2)
    docs_removed = round((1 - (post_filter_docs / pre_filter_docs)) * 100, 2)

    logger.info(f"FILTER: {filter_name}")
    logger.info(
        f"TOKENS: pre: {pre_filter_tokens}, post: {post_filter_tokens}, loss: {tokens_removed}%"
    )
    logger.info(
        f"DOCUMENTS: pre: {pre_filter_docs}, post: {post_filter_docs}, loss: {docs_removed}%"
    )

    return dataset


source_filter = lambda ds: doc_type_searchword in ds["doc_type"]  # noqa
length_filter = lambda ds: len(word_tokenize(ds["text"])) >= 10  # noqa
too_long_filter = lambda ds: len(word_tokenize(ds["text"])) > 1e5  # noqa
alpha_filter = lambda ds: alpha_ratio(ds["text"]) >= 0.7  # noqa
stop_word_filter = lambda ds: count_min_target(word_tokenize(ds["text"]), spacy_sw, 2)  # noqa

samples_pr_source: dict = defaultdict(lambda: defaultdict(int))


def language_filter_with_desc_stats(ds: Dataset) -> bool:
    """
    Language filtering in a streamed dataset while logging all languages
    """
    s = source
    language = ds["lang_fasttext"]
    samples_pr_source[s][language] += 1

    language_filter = (
        ds["lang_fasttext"] == "da" and float(ds["lang_fasttext_conf"]) >= 0.75
    )

    return language_filter


def quality_checks(ds: Dataset) -> Dataset:
    """
    Quality checks for:
    - no duplicate ids
    - no duplicate texts
    - logs texts > 1e5 tokens
    """
    # convert to pandas for the drop_duplicates()
    df = pd.DataFrame(ds)
    # remove duplicate ids
    len_df = len(df)
    df = df.drop_duplicates(subset=["id"])
    logger.info(f"Removed {len_df - len(df)} duplicate ids")
    # remove rows with duplicate text
    len_df = len(df)
    df = df.drop_duplicates(subset=["text"])
    logger.info(f"Removed {len_df - len(df)} rows with duplicate text")
    # reconvert and remove index
    ds_f = Dataset.from_pandas(df, preserve_index=False)
    try:
        ds_f["__index_level_0__"]
        ds_f = ds_f.remove_columns("__index_level_0__")
    except KeyError:
        pass

    assert len(set(ds_f["id"])) == len(ds_f), "IDs are not unique"
    assert len(set(ds_f["text"])) == len(ds_f), "Texts are not unique"

    long_texts = ds_f.filter(too_long_filter, num_proc=None)
    if len(long_texts["id"]) > 0:
        logger.info(f"{len(long_texts['id'])} Long texts (>~1e5 tokens) found")
        for id in long_texts["id"]:
            logger.info(f"id: {id}")
    else:
        logger.info("No long texts (>~1e5 tokens) found")

    return ds_f


def main():
    # load all splits
    logger.info(f"Loading data from: {hf_path}")

    danish_data = load_dataset(
        hf_path, streaming=False, split="train+validation", num_proc=num_proc
    )
    danish_data.cleanup_cache_files()

    # filter by metadata
    logger.info(f"Processing source: {source}")
    danish_data = danish_data.filter(source_filter, num_proc=num_proc)

    logger.info("Processing language")
    danish_data = danish_data.filter(language_filter_with_desc_stats, num_proc=None)

    # log language changes
    log_pre_filter_lang_data(samples_pr_source, danish_data)

    # convert to dynaword format
    danish_data = danish_data.map(dynaword_format)
    danish_data = danish_data.select_columns(
        ["text", "source", "id", "added", "created", "license", "domain", "metadata"]
    )

    # filter and log changes
    danish_data = filter_with_changelog(length_filter, danish_data)
    danish_data = filter_with_changelog(alpha_filter, danish_data)
    danish_data = filter_with_changelog(stop_word_filter, danish_data)

    # Quality checks
    danish_data = quality_checks(danish_data)

    ### saving
    save_path = Path(__file__).parent / f"{source}.parquet"
    danish_data.to_parquet(save_path)


if __name__ == "__main__":
    log_path = Path(__file__).parent / f"{source}.log"
    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s - %(levelname)s - %(message)s",
        handlers=[
            logging.StreamHandler(),
            logging.FileHandler(log_path),
        ],
    )
    main()