Evaluating Word2Vec and FastText Embeddings in CNN-LSTM for Hotel Customer Satisfaction: Impact on Accuracy and Model Stability

Customer reviews on digital platforms such as Google Review are a valuable source of information for evaluating user satisfaction in the hospitality industry. This study aims to assess the effectiveness of a CNN-LSTM architecture for classifying hotel customer satisfaction based on textual reviews by comparing two word embedding techniques: Word2Vec (CBOW) and FastText. The dataset, which was labeled into two classes "satisfied" and "dissatisfied" underwent preprocessing, embedding, and training under two conditions: with and without the Early Stopping mechanism. Experimental results indicate that Word2Vec with Early Stopping achieved the highest accuracy of 91%, with balanced classification metrics and stable training-validation performance. Meanwhile, FastText yielded lower performance, particularly in recognizing minority class instances, especially when Early Stopping was enabled. Although disabling Early Stopping improved minority class detection, it introduced signs of overfitting. The findings demonstrate that Word2Vec embeddings, when combined with Early Stopping, offer better generalization and robustness in imbalanced sentiment classification tasks. This research contributes to the limited body of literature that systematically compares word embeddings within a CNN-LSTM framework for sentiment analysis. It also underscores the importance of selecting suitable embedding strategies and training configurations to improve model fairness and accuracy in real- world applications.