A Comparative Evaluation of LDA and BERTopic for Topic Modeling of Traffic Complaints in Short Social Media Texts

This study compares the performance of two topic modeling techniques, Latent Dirichlet Allocation (LDA) and BERTopic, in analyzing short, unstructured traffic complaint texts from Twitter. Evaluations were conducted under two preprocessing conditions (with and without stemming) using two coherence metrics, C_V and C_NPMI. The results showed that BERTopic consistently achieved higher average coherence scores across all scenarios. For non-stemmed data, BERTopic reached 0.3517 (C_V) and −0.0986 (C_NPMI), outperforming LDA's 0.3107 and −0.1724, respectively. For stemmed data, BERTopic scored 0.3902 and −0.1391, compared to LDA's 0.3201 and −0.1830, respectively. The BERTopic also demonstrated greater flexibility in determining the optimal number of topics. Paired t-tests confirmed the statistical significance of these differences in two of the four scenarios. These findings suggest that BERTopic is more effective for modeling noisy, short-form texts on social media. Future work should explore multilingual traffic complaints and human-in-the-loop validation to further enhance topic interpretability in real-world applications.