🤖 AI Summary
This work investigates the effectiveness of large language models (LLMs) for headline-level clickbait detection under zero-shot and few-shot settings. We systematically evaluate mainstream LLMs—including GPT and ChatGLM—across multiple Chinese and English benchmark datasets. Our empirical study reveals, for the first time, that LLMs relying solely on headline text—without fine-tuning—achieve substantially lower performance than task-specific, fine-tuned pretrained language models (PLMs), exposing a critical structural limitation in prompt-based inference for fine-grained content safety tasks. The core contributions are threefold: (1) a cross-lingual, multi-benchmark evaluation framework for LLM-based clickbait detection; (2) empirical validation that unimodal headline input imposes a fundamental information bottleneck, constituting a primary source of performance degradation; and (3) evidence-based delineation of the applicability boundaries of LLMs in content safety, challenging the assumption that LLMs can directly replace fine-tuned PLMs in such domains.
📝 Abstract
Clickbait, which aims to induce users with some surprising and even thrilling headlines for increasing click-through rates, permeates almost all online content publishers, such as news portals and social media. Recently, Large Language Models (LLMs) have emerged as a powerful instrument and achieved tremendous success in a series of NLP downstream tasks. However, it is not yet known whether LLMs can be served as a high-quality clickbait detection system. In this paper, we analyze the performance of LLMs in the few-shot and zero-shot scenarios on several English and Chinese benchmark datasets. Experimental results show that LLMs cannot achieve the best results compared to the state-of-the-art deep and fine-tuning PLMs methods. Different from human intuition, the experiments demonstrated that LLMs cannot make satisfied clickbait detection just by the headlines.
