Abstract:
Customer sentiment based on hotel reviews is crucial to the quality improvement process in the digital hospitality context, but informal and multilingual feedback is difficult toanalyze. This paper deals with the issue of correctly identifying positive and negativeguest reviews based on real life hotel reviews in Dhaka, which was gathered through a Google API scraper in order to obtain quantity and authenticity. The suggested solution is a hybrid one with the use of Long Short-Term Memory (LSTM) neural networks and Term Frequency-Inverse Document Frequency (TF IDF) features. The methodology includes the comprehensive data cleaning, tokenization, feature engineering, and a dual- branch neural network: sequential LSTM modeling of contextual cues and TF IDF encoding of global lexical detail. The model achieves a high classification accuracy of about 95 percent and the precision and recall are balanced between the two levels of sentiment. The findings validate the usefulness of sequence-based deep learning with sparse lexical features in robust sentiment interpretation of diverse, automatically- collected hotel review datasets.
