Clickbait detection in indonesian news headlines using various prompting strategies in large language models

Abstract

Clickbait detection in news headlines is a critical task in Natural Language Processing (NLP) related to information quality and the credibility of digital journalism. While traditional machine learning and deep learning approaches have demonstrated impressive performance in clickbait detection, they are limited by a heavy reliance on extensive annotated datasets and significant computational requirements for model training. Unlike previous methods, Large Language Models (LLMs) do not require massive amounts of annotated data. LLMs allow classification tasks to be solved through zero-shot and few-shot prompting without additional retraining. However, the effectiveness of these models can depend significantly on prompting configuration. Despite this, linguistically-enriched prompting strategies have not been widely evaluated for non-English domains such as Indonesian news headlines. This study aims to analyze the performance of various LLM prompting strategies in detecting Indonesian-language clickbait headlines. For this purpose, this study evaluated Llama 4 on the CLICK-ID dataset using multiple combinations of plain and linguistically-enriched prompts (zero-shot and few-shot) alongside advanced inference techniques (self-consistency, weighted self-consistency, and Self-Refine). Performance was measured via Accuracy and Macro F1-scores against a fine-tuned IndoBERT as baseline model. The results show that the prompting approach in the Large Language Models (LLMs) can be used to classify Indonesian clickbait effectively. The use of linguistic prompting and few-shot managed to provide the best performance, which achieved an accuracy of 0.90 and a Macro F1-score of 0.89.