Fine-Tuned Large Language Models for Improved Clickbait Title Detection

The term clickbait is used to describe headlines or other content that is designed to attract attention and clicks, often by using sensationalized or misleading language. However, the task of detecting click-bait titles automatically can be challenging due to the wide variance in the wording and structure of titles. Rule-based approaches, machine learning models, deep learning models, and natural language processing techniques are some of the existing methods for click-bait detection. As good as these techniques are, they still need customization and large amounts of training data to achieve good accuracy. Since generative, pretrained transformer models have shown an inherent capability to understand the meanings of sentences really well, our research examined the use of GPT-3.5, and GPT-4 as zero-shot models. They resulted in a maximum test accuracy of 88.5%, lower than some of the custom techniques reported in previous research works of between 90% to 95%. This indicates the need to go beyond a generic LLM. We then fine-tune OpenAI's Large Language Model (LLM) Ada using an efficient dataset of only 1000 samples of clickbait and non-clickbait titles to get a test accuracy of up to 99.5 %. The research further shows that fine-tuning a GPT model can yield not only better accuracy but also consume smaller amount of training data, shorter development time, and be cost-effective. Our research is indicative of the broader possibility that fine-tuning a large language model is all you need for most targeted natural language processing tasks.