온라인 댓글의 주제 분석을 위한 토픽 모델링 : 이슈 포착과 분류에 활용 가능한 LDA와 BTM의 비교와 검증

초록

Using computers to rapidly and efficiently build a model to organize massive volumes of textual data, topic modeling is an unsupervised machine learning technique that can be used to classify texts into related themes or to analyze the nature and distribution of topics. However, topic modeling's usage in media research has recently come under fire for failing to take into account reliable and valid measures of theoretically defined concepts. This means that topic modeling needs methodological validation and reliability in order to be employed in media research as a tool for investigating and summarizing massive volumes of textual material. Additionally, it is helpful to be able to group online comments into "issues" and list their important points in order to quickly identify social issues and monitor discourse patterns in real-time on digital platforms. For this reason, attempts to diagnose the methodological validity of topic modeling for analyzing the topics of comments are of great significance. Therefore, this study validates topic modeling for analyzing online comments by verifying its performance as follows. First, we discussed why topic analysis of comments is necessary and what the implications are through the conceptualization of "issues" in the context of online comments. Then, with an emphasis on the Latent Dirichlet Allocation (LDA) model, we reviewed the principle of topic modeling to estimate the topic of text and the assumptions of statistical models that affect topic estimation. Additionally, we contrasted the merits and drawbacks of LDA and the Biterm Topic Model (BTM) to suggest topic modeling as a means of identifying the subject of comments and categorizing them as "issues." Based on the above theoretical discussion, we applied topic modeling to analyze 9,000 online news comments on articles covering nine social issues and validate whether the topics are useful to classify comments according to the "issues" of the news. The results are as follows. First, compared to BTM, LDA is highly dependent on the hyperparameter, , with lower values leading to better model performance. Second, both BTM and LDA were able to estimate the optimal number of topics (K ), but BTM showed less variation in performance with value selection than LDA, and performance degradation was worse when the value was lower than the optimal K than when it was higher. Third, both BTM and LDA performed better when adding bigrams along with unigrams to the vocabulary, but the difference was more pronounced for LDA. Based on these validation results, we assessed the validity of topic modeling for analysis of comments and discussed its implications.

키워드