In the context of a recommendation system, a user session is a list of user actions or operations that occur during a specific period while using a website, application, or online service. These sessions are used in session-based recommendation systems to learn user preferences, and are applied to item recommendations. However, because user preferences change dynamically, different topic elements are combined in a single session. To accurately capture changes in user preferences, it is important to divide sessions by topic and analyze the separate subsessions. Therefore, the selection of the best session-segmentation method is an important issue. In this study, we investigate cosine similarity, k-means clustering, and classifier including Light Gradient Boosting Machine (LightGBM), Support Vector Machine (SVM), and Logistic Regression to segmentation methods. In addition, Item2Vec, Word2Vec, and OpenAI were used for item embedding. We conducted a comparative evaluation using annotation data to assess the combination of the item embedding and segmentation methods. The evaluation metrics used were the F1-score, PR-AUC, and ROC-AUC. The combination of cosine similarity and Item2Vec exhibited the best performance. The values achieved were an F1-score of 0.714 and PR-AUC of 0.794.