BanglaBait: Semi-Supervised Adversarial Approach for Clickbait Detection on Bangla Clickbait Dataset

Proceedings of Recent Advances in Natural Language Processing, pages 748–758 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_081 748 BanglaBait: Semi-Supervised Adversarial Approach for Clickbait Detection on Bangla Clickbait Dataset Md. Motahar Mahtab BRAC University Dhaka, Bangladesh mahtab27672767@gmail.com Monirul Haque BRAC University Dhaka, Bangladesh monirul.haque.mail@gmail.com Mehedi Hasan BRAC University Dhaka, Bangladesh mehedi.hasan@g.bracu.ac.bd Farig Sadeque BRAC University Dhaka, Bangladesh farig.sadeque@bracu.ac.bd Abstract Intentionally luring readers to click on a partic- ular content by exploiting their curiosity de- fines a title as clickbait. Although several studies focused on detecting clickbait titles in English articles, low-resource language like Bangla has not been given adequate attention. To tackle clickbait titles in Bangla, we have constructed the first Bangla clickbait detection dataset containing 15,056 labeled news articles and 65,406 unlabelled news articles extracted from clickbait-dense news sites. Each article has been labeled by three expert linguists and includes an article’s title, body, and other meta- data. By incorporating labeled and unlabelled data, we finetune a pre-trained Bangla trans- former model in an adversarial fashion using Semi-Supervised Generative Adversarial Net- works (SS-GANs). The proposed model acts as a good baseline for this dataset, outperform- ing traditional neural network models (LSTM, GRU, CNN) and linguistic feature-based mod- els. We expect that this dataset and the detailed analysis and comparison of these clickbait de- tection models will provide a fundamental basis for future research into detecting clickbait ti- tles in Bengali articles. We have released the corresponding code and dataset 1. 1 Introduction Due to the widespread usage of the internet, the news industry has progressively evolved into an online news industry leading to the explosion of clickbait titles in recent years. As the concept of clickbait can be hazy to grasp, the classification of 1https://github.com/mdmotaharmahtab/ BanglaBait clickbait is a highly subjective endeavor. Biyani et al. (2016) suggests that clickbait titles can be roughly categorized into eight types. Table 1 dis- plays this different clickbait categories2 and their corresponding Bangla articles. There are an estimated 11.4 million internet users in Bangladesh3 who receive their daily news mostly from online news sites. However, no re- search has been conducted on tackling the increas- ing number of clickbait titles on these sites and other news websites. For English articles, Pot- thast et al. (2018a) built the first large-scale an- notated clickbait corpus (Webis Clickbait Corpus 2017) containing 338,517 articles. In the Bangla language, the lack of an annotated clickbait-rich dataset is hindering the progress of Bangla Click- bait Detection. We construct the first Bangla Clickbait Corpus, which contains an article’s ti- tle, content, and other metadata collected from various clickbait-rich websites upon which future researchers can build an effective Bangla click- bait detection model. The effectiveness of Semi- Supervised Generative Adversarial networks (SS- GANs; Salimans et al., 2016) have been shown for text classification tasks in Croce et al. (2020). From our experiments, it is evident that fine-tuning a Bangla ELECTRA model in this setup improves clickbait detection performance outperforming all other model types. The main contributions of this paper can be sum- marized …

Content truncated to 500 words