Optimizing aspect-based sentiment analysis using sentence embedding transformer, bayesian search clustering, and sparse attention mechanism

Abstract

The aspect-based sentiment analysis of hotel reviews represents a critical yet underexplored area of research. Such reviews encompass a range of expressions that extend beyond simple positive, negative, or neutral sentiments, delving into specific aspects discussed within the review. For instance, a statement like 'the location is comfortable, the food is good, but unfortunately, the service is bad' encapsulates multiple aspects, including location, food, and service. Existing research in Aspect-Based Sentiment Analysis (ABSA) frequently employs word embedding techniques for feature extraction, utilizing both Term and Transformer models. This study introduces an innovative approach by transitioning from word to sentence embedding, employing the SBERT transformer model for this transformation. This study further innovates label creation and topic modeling by integrating Bayesian search clustering with Inverse Document Frequency (IDF) calculations. This approach enables the identification of relevant topics, categorizing IDF scores based on occurrence frequencies. Scores ranging between 4–6 are marked as having low value but high informative potential, making them suitable candidates for label-specific datasets, particularly those without pre-existing labels. Additionally, this study incorporates the BigBird model to enhance this classification methodology. Applying this comprehensive framework to hotel review datasets, the result of the ABSA evaluation achieved an impressive accuracy score of 99%, indicating the efficacy of the proposed methodology in this nuanced field.