This package contains 16,884 Arabic newspaper documents retained from the original 20,291-document Watan-2004 corpus compiled by Mourad Abbas from the Omani newspaper Alwatan. The compiler's corpus page describes six topics and intentional punctuation removal for language modeling; SourceForge records the original document count. The raw_text column means the supplied source text, which was already punctuation-stripped; it is not untouched newspaper prose. A source filename ending in .html does not guarantee that its contents are HTML.

This preparation applies exact deduplication before stratified splitting, with 3,407 documents dropped. Duplicate identity is SHA-256 of exact decoded UTF-8 raw text, without trimming or normalization; the lexicographically smallest document ID survives each same-label group. There were no cross-label exact-duplicate conflicts. Do not substitute the original corpus counts for this retained population.

Labels, retention and splits

Labels are the six original source directories. Class counts are:

Label	Original	Retained	Dropped
culture	2782	2746	36
economy	3468	3437	31
international	2035	2001	34
local	3596	3506	90
religion	3860	686	3174
sports	4550	4508	42
Total	20291	16884	3407

Religion has particularly heavy duplication (3,860 → 686), and the retained corpus is imbalanced. Deduplication changes class prevalence substantially; model comparisons should disclose this population and report results by class where appropriate.

The Hugging Face split corpus is the complete retained corpus, not a research training split. Stored research memberships in data/splits.json are fit 10,130, calibration 3,377 and validation 3,377: disjoint and exhaustive, stratified by the single label with seeds 20260824/20260825.

Representations and preprocessing

Each configuration reads one UTF-8 CSV and returns columns id, raw_text, vw_text. IDs follow watan2004/<topic>/<filename>. The VW string starts with this ID and has only the |@word namespace. noow preserves normalized lemma occurrences in source order, while bow aggregates the same occurrences as token:count. Neither VW representation preserves original prose.

Preprocessing uses Python 3.11.15, CAMeL Tools 1.6.0, simple_word_tokenize(split_digits=False) and MLEDisambiguator.pretrained("calima-msa-r13", top=1). Lemmas are canonicalized with NFKC, casefold and removal of Arabic combining diacritics. There is no stopword removal. The preparation contains 7,257,088 analyzed/final word occurrences and zero morphology fallbacks.

Access and loading

For non-commercial research or education, sign in on the dataset page and submit the access form if prompted. Once your request is approved, authenticate with the same account before downloading. Hugging Face already collects your username and email; the form asks only for your affiliation (or independent researcher status), intended use and acknowledgement of the use conditions.

Use the interactive Hugging Face login so your token stays out of commands and logs:

from huggingface_hub import login

login()

The default configuration is noow; bow contains aggregated counts. Both expose the complete population as the Hugging Face split corpus. Labels and the stored fit/calibration/validation memberships are auxiliary files, joined by document ID. data/labels.json directly maps each string ID to a list of string labels; data/splits.json directly maps each split name to an ordered ID list.

from datasets import load_dataset

repo = "TopicNet/Watan-2004"
corpus = load_dataset(repo, "noow", split="corpus", token=True)
bow = load_dataset(repo, "bow", split="corpus", token=True)

For local use, obtain the complete package including actual Git LFS contents. Set HF_DATASETS_OFFLINE=1 and HF_HUB_OFFLINE=1 before Python starts and choose a writable cache directory:

import json
from pathlib import Path
from datasets import load_dataset

package = Path("/path/to/Watan-2004")
corpus = load_dataset(str(package), "noow", split="corpus",
                      cache_dir="/writable/cache")
bow = load_dataset(str(package), "bow", split="corpus",
                   cache_dir="/writable/cache")
with (package / "data/labels.json").open(encoding="utf-8") as handle:
    labels = json.load(handle)
with (package / "data/splits.json").open(encoding="utf-8") as handle:
    splits = json.load(handle)
membership = {doc_id: split for split, ids in splits.items() for doc_id in ids}
annotated = corpus.map(lambda row: {"labels": labels[row["id"]],
                                   "research_split": membership[row["id"]]})
fit = annotated.filter(lambda row: row["research_split"] == "fit")

For remote metadata, download data/labels.json and data/splits.json with huggingface_hub.hf_hub_download(repo, filename, repo_type="dataset", token=True), then read the same direct mappings. Use the account with approved access for both CSV files and metadata.

Files and integrity

The package contains this card, .gitattributes, SHA256SUMS, two CSV representations, labels/splits metadata and standalone preprocessing code and instructions. Run sha256sum -c SHA256SUMS from the package root to check the listed files.

Rights and intended use

This package is intended for non-commercial research or education. Access approval does not grant redistribution rights. The data license is unknown; redistribution rights remain unresolved. The SourceForge project is labeled GPLv2, but that project metadata does not establish permission to redistribute the underlying newspaper articles. The compiler's corpus page includes a scientific-use note and an authorization condition for releasing derived resources/software; the corpus scope and applicability of that note need clarification with the rights holder.

These newspaper texts reflect historical and regional selection biases and may contain personal information. Deduplication substantially changes class proportions, and normalization does not restore missing punctuation. The articles may contain outdated claims. Assess the corpus's suitability for your application before relying on it for decisions.

Sources and citation

The SourceForge corpus description requests the Watan reference below; the compiler's page also lists the Alwatan-specific study:

M. Abbas, K. Smaili, D. Berkani (2011). Evaluation of Topic Identification Methods on Arabic Corpora. Journal of Digital Information Management, 9(5), 185–192.
M. Abbas, K. Smaili, D. Berkani (2011). Evaluation of Topic Identification Methods for Arabic Texts and their Combination by using a Corpus Extracted from the Omani Newspaper Alwatan. Arab Gulf Journal of Scientific Research, 29(3–4), 183–191.

When reporting this preparation, identify the 16,884-document deduplicated population, CAMeL Tools configuration and repository revision.

Downloads last month: 20

Total file size:

289 MB