A Clickbait-Filtered Topic Modeling Approach for Real Time Vietnamese News Clustering

Abstract

The proliferation of online journalism in Vietnam has led to a high volume of clickbait headlines designed to maximize pageviews. While real-time news clustering systems are widely used to aggregate related articles and track event streams, their performance is degraded by the stylistic and syntactic noise introduced by clickbait. This paper investigates the quantitative impact of clickbait filtering on neural topic modeling. We evaluate a two-stage framework: first, news articles are classified and filtered using a fine-tuned PhoBERT classifier; second, the remaining clean stream is clustered using BERTopic and labeled using the Gemini API. Experimental evaluation on a crawled dataset of 12,450 Vietnamese news articles shows that filtering clickbait improves the Topic Coherence (𝐶𝑣) of identified events from 0.4128 to 0.5471 (+32.5%) and increases Topic Diversity from 0.6842 to 0.8219 (+20.1%). Human evaluations confirm that the event labels generated by the LLM achieve higher objectivity (4.72 vs. 2.85 out of 5) and accuracy (4.55 vs. 3.42 out of 5) when clickbait is removed prior to clustering.