dwablimol/wikipedia-20231101-wordcount

Wordcounts for the English Wikipedia dump (2023-11-01), including words that occur at least 10 times in the corpus. Created using the following script:

import re
import duckdb
from collections import Counter
from datasets import load_dataset
from tqdm.auto import tqdm

conn = duckdb.connect(":memory:")

def ensure_db(conn: duckdb.DuckDBPyConnection):
    conn.execute("""
        CREATE TABLE IF NOT EXISTS wc (
            word TEXT PRIMARY KEY,
            c BIGINT
        );
    """)

ensure_db(conn)

def merge_batch(conn: duckdb.DuckDBPyConnection, counts: Counter):
    if not counts:
        return
    df = pd.DataFrame({"word": list(counts.keys()), "c": list(map(int, counts.values()))})
    # Register the batch dataframe as a view, then MERGE (UPSERT)
    conn.register("batch_df", df)
    conn.execute("""
        MERGE INTO wc AS t
        USING batch_df AS s
        ON t.word = s.word
        WHEN MATCHED THEN UPDATE SET c = t.c + s.c
        WHEN NOT MATCHED THEN INSERT (word, c) VALUES (s.word, s.c);
    """)
    conn.unregister("batch_df")

TOKEN_RE = re.compile(r"[a-z]+(?:'[a-z]+)?")   # keep internal apostrophes

def tokenize_en_lower(text: str):
    if not text:
        return []
    return TOKEN_RE.findall(text.lower())

batch_size = 500
limit = 0

ds_iter = load_dataset("wikimedia/wikipedia", "20231101.en", split="train", streaming=True)
buf = Counter()
n = 0
pbar = tqdm(desc="Processing (streaming)", unit="art")
for ex in ds_iter:
    buf.update(tokenize_en_lower(ex.get("text", "")))
    n += 1
    if n % batch_size == 0:
        merge_batch(conn, buf); buf.clear()
        pbar.update(batch_size)
    if limit and n >= limit:
        break
if buf:
    merge_batch(conn, buf); pbar.update(n % batch_size)
pbar.close()