CaT: A Category-oriented Transformer for Text and Sentiment Analysis in Japanese Hotel Overall Ratings Prediction

• Published in: International Symposium on Computing and Networking (Online) pp. 162 - 168
• Main Authors: Kusunoki, Zaku, Jiang, Shuai, Hou, Shengzhe, Kamei, Sayaka, Morimoto, Yasuhiko
• Format: Conference Proceeding
• Language: English
• Published: IEEE 26.11.2024
• Subjects: Analytical models; BERT; Deep learning; Dictionaries; Evaluation Prediction; Feature extraction; Fuses; Knowledge graphs; Natural Language Processing; Predictive models; Reviews; Sentiment analysis; Transformer; Transformers
• ISSN: 2379-1896
• DOI: 10.1109/CANDAR64496.2024.00027

Guest reviews on hotel booking websites typically include textual comments and numerical ratings. However, missing ratings and inconsistencies between textual comments and numerical ratings often confuse site users. Consequently, extracting a comprehensive numerical score from textual reviews has become critical. Specifically, inconsistency problems with hotel online ratings in the Japanese context have not been studied well. We proposed a data-driven method utilizing CSPD (Category-oriented Sentiment Polarity Dictionaries), which is automatically edited for each category using the review database from Rakuten Travel in Japanese. For each token in the review, the text features are first learned using BERT, while the sentiment features are learned by assigning sentiment polarity values from CSPD via an encoder. These outputs are then input into a transformer to learn the interaction between text and sentiment features. Our method predicts the comprehensive evaluation based on the interaction learned between these features. We conducted evaluation experiments using the Rakuten Travel review dataset from 2014 to 2019. The results of the experiments showed that our method achieved higher accuracy than methods that learn only text features, demonstrating the effectiveness of using sentiment features as auxiliary information in deep learning. The ablation experiment also proves the effectiveness of each module of the model.