Arabic_Translated_LAIR_WELFAKE_dataset

Arabic Translated WELFake and LIAR Dataset for Fake News Detection
This dataset is a unified Arabic headline corpus constructed for the purpose of binary fake news detection in Arabic. It was produced by translating two established English fake news benchmarks — WELFake and LIAR — into Arabic, followed by a systematic preprocessing and cleaning pipeline, and merging the results into a single resource.
Source Datasets
The WELFake dataset (Verma et al., 2021) originally contains 72,134 English news articles aggregated from four sources (Kaggle, McIntire, Reuters, and BuzzFeed Political), labeled as fake (0) or real (1). Only the headline (title) field was retained for this work, discarding full article text. The LIAR dataset (Wang, 2017) contains 12,745 short political statements collected from PolitiFact, originally annotated on a six-level veracity scale. The six labels were collapsed into a binary scheme: a true class (half-true, mostly-true, true) and a false class (pants-fire, false, barely-true).
Translation
All English headlines and statements were translated into Arabic using the Google Translate API via the googletrans Python library. Due to throughput constraints, WELFake was processed in 72 chunk-batches of approximately 1,000 headlines each. For LIAR, the NLLB-200 distilled-600M multilingual model (facebook/nllb-200-distilled-600M) was used as a complementary translation backbone whenever Google Translate produced unstable outputs.
Preprocessing Pipeline
Both datasets were subjected to an identical Arabic text cleaning pipeline consisting of: (1) URL and hyperlink removal; (2) Arabic diacritic (Tashkeel) removal; (3) Tatweel (kashida) removal; (4) normalization of Arabic letter variants (e.g., أ / إ / آ → ا, ى → ي, ة → ه); (5) removal of Latin-script tokens; (6) removal of Western-Arabic and Eastern-Arabic-Indic digits; (7) removal of all characters outside the Unicode Arabic block; (8) collapsing of repeated characters (e.g., جدااا → جدا); (9) Arabic stop word removal using a curated list covering prepositions, conjunctions, pronouns, demonstratives, and high-frequency reporting verbs; and (10) whitespace normalization and trimming. At the row level, translation-error markers (خطا في الترجمه), empty or single-character instances, exact duplicates, and label-conflicting instances (same text with both fake and real labels) were removed.
Final Dataset Composition
After merging and applying cross-dataset deduplication and conflict removal, the final unified corpus comprises 74,559 instances.
File Structure
The dataset is provided as a single UTF-8 encoded CSV file (Arabic_Translated_Welfak_and_Liar_Dataset.csv) with three columns:

text — the preprocessed Arabic headline or statement
label — binary class label: 1 = Fake, 0 = Real
source — origin dataset: WELFake or LIAR