Topic Modeling of Instructor Comments During Captain Upgrade Training Using Latent Dirichlet Allocation

Analyzing instructor comments, which provide feedback to applicants to enhance their growth, using Latent Dirichlet Allocation, we obtained interpretable topics and their differences in frequency across groups stratified by total training period length and relative timing during captain upgrade training conducted by an airline company. Representative themes were extracted by evaluating coherence scores, perplexity, and topic stability across different numbers of topics. The five-topic model, which yielded the highest coherence score, produced four topics that remained consistently identifiable, with the top words preserved in the eight- and ten-topic models. From these results, five core topics, labelled "policy," "weather," "look," "adequate," and "stable," were extracted. We then analyzed the temporal and group-level distributions of these topics, revealing upward, downward, V-shaped, and inverted V-shaped trends. The most frequent topic labelled "look" showed an upward trend across phases and a downward trend across categories, reflecting training-period length, whereas the other topics exhibited diverse patterns.