Hyun-Jung Park, Tae-Min Lee, Heui-Seok Lim, Journal of Internet Computing and Services, Vol. 26, No. 1, pp. 19-33, Feb. 2025
10.7472/jksii.2025.26.1.19, Full Text:
Keywords: Topic Model, Text Mining, Coherence, diversity, 1~N-gram Topic Representation, LDA, BERTopic

Abstract

Topic models are text mining techniques actively applied in various domains to explore the underlying themes in large-scale text data. However, despite their high popularity, there is a significant lack of performance comparison studies on these models. To provide a comprehensive performance comparison of topic models, this study focuses on 1~N-gram topic representations in addition to unigrams. While most existing studies use unigram topic representations, 1~N-gram topic representations have the advantage of enhancing topic interpretability. However, it is challenging to calculate topic coherence for 1~N-gram topic representations using existing general-purpose Python libraries. Therefore, we first identify the causes of this difficulty, proposing and implementing solutions. Next, we model unigram and 1~3-gram topic representations of major traditional BOW-based and recent deep learning-based topic models on Korean online articles and KCI paper data, comparing and analyzing their coherence and diversity. As a result, we provide multifaceted insights for effective topic modeling: BERTopic and semi-supervised BERTopic generally outperform NMF and LDA in terms of coherence and diversity, both for unigrams and 1~3-grams. The coherence of BERTopic and semi-supervised BERTopic for 1~3-gram topic representations tends to increase with a larger number of topics, longer window sizes, and longer given texts compared to LDA or NMF, etc.

Statistics
Cite this article