Ar-WELFake-LIAR: An Arabic Machine-Translated Headline Corpus for Fake News Detection (74,559 Instances)
Description
Description This dataset is an Arabic headline-level corpus for binary fake news detection. It was built by cross-lingual transfer of two established English benchmarks, WELFake and LIAR. It is the translated resource described in the article “Cross-Lingual Data Transfer for Arabic Headline Misinformation Detection: Translation, Multi-Source Evaluation, and Ensemble Learning” (Albtoush, Gan, et al.) 1. Source datasets WELFake (Verma et al., 2021). 72,134 English news articles (35,028 real, 37,106 fake) merged from Kaggle, McIntire, Reuters and BuzzFeed Political. Only the title field was kept. Records with empty or non-textual titles, or that produced a translation error, were removed, leaving 62,154 titles (27,818 fake, 34,336 real). The removals fell mostly on fake items, because empty titles are concentrated in the BuzzFeed Political and McIntire subsets. LIAR (Wang, 2017). 12,745 short political statements from PolitiFact, labelled on a six-level scale. The labels were collapsed into two classes: Fake (pants-fire, false, barely-true) and Real (half-true, mostly-true, true). This gave 5,641 fake and 7,104 real statements. 2. Translation WELFake titles were translated into Arabic with Google Translate via the googletrans Python library. Because of rate limits, they were translated in 72 chunks of about 1,000 titles each and then merged back in the original order. LIAR statements were translated with the open-source Helsinki-NLP/opus-mt-en-ar model (MarianMT), run locally in batch. Translation quality was evaluated on a stratified sample of 200 items: 100 from WELFake and 100 from LIAR, each with 50 fake and 50 real. Both engines translated the same items, and the outputs were compared with LaBSE cosine similarity, chrF, BLEU (sacreBLEU), 1,000 bootstrap replicates and a numeric-preservation check. Mean LaBSE: 0.888 for Google and 0.826 for Helsinki. Agreement between the two engines was lower for WELFake headlines (chrF 47.6, BLEU 19.3) than for LIAR statements (chrF 63.1, BLEU 36.1). Numbers were preserved in 93.5% of Google outputs and 95.5% of Helsinki outputs. LaBSE did not differ systematically between classes (fake 0.8845, real 0.8911). 3. Arabic cleaning and normalisation Both translated sources went through the same pipeline: removal of Arabic diacritics and tatweel letter normalisation removal of URLs, Latin-script tokens and digits removal of Arabic stop-words collapsing three or more repeated characters whitespace normalisation Rows were then removed if they: carried the translation-error marker were one character or less were exact duplicates (the first occurrence was kept) appeared with both a fake and a real label 4. Final composition Source Fake Real Total Mean words Ar-WELFake 27,651 34,204 61,855 9.87 Ar-LIAR 5,617 7,087 12,704 12.48 Total 33,268 41,291 74,559 no rows were removed when they were merged. All CSV files are UTF-8 encoded, and every file uses the same label coding: 1 = Fake, 0 = Real.
Files
Steps to reproduce
Obtain the source data. Download WELFake (Verma et al., 2021; 72,134 articles) and LIAR (Wang, 2017; 12,745 statements). Keep only the title field from WELFake and the statement field from LIAR, each with its veracity label. Binarise the LIAR labels. Map pants-fire, false and barely-true to Fake, and half-true, mostly-true and true to Real. This gives 5,641 fake and 7,104 real statements. Filter the WELFake titles. Remove records whose title is empty or non-textual, or that produced a translation error. This leaves 62,154 titles (27,818 fake, 34,336 real). Translate into Arabic. WELFake: Google Translate via the googletrans Python library (source = en, target = ar), in 72 chunks of about 1,000 titles. Merge the chunks back in their original order with labels preserved. LIAR: Helsinki-NLP/opus-mt-en-ar (MarianMT) through Hugging Face Transformers, run locally on GPU in batches. Evaluate translation quality. Draw a stratified random sample of 200 items with a fixed seed: 100 WELFake and 100 LIAR, each with 50 fake and 50 real. Translate the sample with both engines, then compute: LaBSE cosine similarity between source and translation (sentence-transformers) chrF and BLEU between the two engine outputs (sacreBLEU) 95% bootstrap confidence intervals (1,000 replicates) the share of numbers preserved Clean and normalise the Arabic text. Apply the same steps to both sources: remove diacritics and tatweel normalise letters (أ/إ/آ → ا, ى → ي, ة → ه) remove URLs, Latin-script tokens and digits remove Arabic stop-words collapse three or more repeated characters normalise whitespace Filter rows. Drop rows that: carry the marker "خطأ في الترجمة" have one character or less after cleaning are exact duplicates (keep the first occurrence) appear with both labels This gives Ar-WELFake = 61,855 (27,651 fake / 34,204 real) and Ar-LIAR = 12,704 (5,617 fake / 7,087 real). Merge and encode labels. Combine the two sources with a source column and code labels as 1 = Fake, 0 = Real. The final corpus has 74,559 instances (33,268 fake / 41,291 real), and no rows are lost at the merge. Verify the files. Check each CSV against the SHA-256 checksums in README.txt. Software and environment: Python 3, pandas, NumPy, googletrans, Hugging Face Transformers (Helsinki-NLP/opus-mt-en-ar), sentence-transformers (LaBSE), sacreBLEU, run in Google Colab Pro+ on an NVIDIA A100 GPU.
Institutions
- Universiti Sains MalaysiaPenang, George Town