Exploration of Transformer Ensemble and Auto Regressive Approaches to Enhance Performance of Clickbait Title Detection
Md Adith Mollah, Shahrab Khan Sami · 2023
"ClickBait" is a term which refers to online content titles designed to attract the viewers to click, often with the intention of misleading or exaggerating the actual content. Clickbait can be used for various malicious purposes such as social engineering, phishing, misinformation spreading, click fraud etc. Therefore, it is important to design a mechanism for detecting clickbait titles. For this purpose, we have utilised an auto regressive model called XLNet-base-case and 3 different types of transformer embedding models such as BERT-base-en-case, ROBERTa-base-en-case and XLM-ROBERTa-multi-case. We individually trained these models using a large diversified dataset and also performed ensemble by majority voting of BERT-base-en-case, ROBERTa-base-en-case and XLM-ROBERTa-multi-case eventually achieving test accuracy of 90% on imbalanced class dataset and 89% on balanced class dataset where the auto regressive approach gave validation accuracy of about 90.7% and 88.19% respectively on imbalanced class and balanced class datasets.