[Paper Review] Machine Generation and Detection of Arabic Manipulated and Fake News

El Moatez Billah Nagoudi, AbdelRahim Elmadany+3 more|arXiv (Cornell University)|Nov 5, 2020

Misinformation and Its ImpactsSocial Sciences47 references41 citations

TL;DR

The paper proposes a simple method to automatically generate Arabic manipulated (and potentially fake) news using POS-tagged data and word embeddings, introduces AraNews+, a large POS-tagged Arabic news dataset, builds manipulation and fake-news detectors, and achieves state-of-the-art results on Arabic fake news detection.

ABSTRACT

Fake news and deceptive machine-generated text are serious problems threatening modern societies, including in the Arab world. This motivates work on detecting false and manipulated stories online. However, a bottleneck for this research is lack of sufficient data to train detection models. We present a novel method for automatically generating Arabic manipulated (and potentially fake) news stories. Our method is simple and only depends on availability of true stories, which are abundant online, and a part of speech tagger (POS). To facilitate future work, we dispense with both of these requirements altogether by providing AraNews, a novel and large POS-tagged news dataset that can be used off-the-shelf. Using stories generated based on AraNews, we carry out a human annotation study that casts light on the effects of machine manipulation on text veracity. The study also measures human ability to detect Arabic machine manipulated text generated by our method. Finally, we develop the first models for detecting manipulated Arabic news and achieve state-of-the-art results on Arabic fake news detection (macro F1=70.06). Our models and data are publicly available.

Motivation & Objective

• Motivate detection of fake and machine-generated Arabic news amidst data scarcity.
• Introduce AraNews, a large POS-tagged Arabic news dataset to support off-the-shelf research.
• Propose a simple automatic manipulation method that substitutes tokens using word embeddings.
• Create manipulated-news datasets ATB+ and AraNews+ to train detection models.
• Evaluate manipulated-text detection (MTD) and fake-news detection (FND) models, achieving strong results and showing data augmentation benefits.

Proposed method

• Use ATB and AraNews as seed true stories.
• POS-tag the data (ATB already tagged; AraNews tagged via MADAMIRA).
• Train a 300-dim (300-size) AraNewsEmb word embedding on Arabic corpora.
• Identify candidate tokens by POS: N_PROP, N_NUM, ADJ, ADJ_COMP, ADJ_NUM, NEG_PART.
• Replace each candidate token with a k-nearest token from AraNewsEmb, using a 50% character-level similarity threshold.
• Remove negation and replace digits with random numbers for manipulated text.

Experimental results

Research questions

• RQ1 Can automatic Arabic text manipulation generate plausible manipulated/fake news without external checks?
• RQ2 How well can humans detect machine-manipulated Arabic text generated by the proposed method?
• RQ3 Can detectors trained on generated manipulated text improve fake-news detection in Arabic?
• RQ4 What is the impact of using AraNews+ and ATB+ data for manipulated-text detection and fake-news detection?

Key findings

• Best manipulated-text detection (MTD) performance on ATB+ and AraNews+ using large multilingual models reaches macro F1 up to 89.25 with AraBERT on AraNews+ (test set).
• MTD results show ATB+ is harder to detect manipulated text than AraNews+ (F1: 83.20 vs 89.25 on respective test sets).
• On fake-news detection (FND) using Khouja as external gold, data augmentation with generated data (AraNews+ doubled) yields F1 of 70.06, outperforming baselines.
• Zero-shot FND with generated data alone achieves notable F1 (up to 52.71) indicating utility of generated data for detection tasks without gold training data.
• The authors release AraNews, ATB+, AraNews+, and detection models publicly for research use.