LARGE-SCALE ASPECT-BASED SENTIMENT ANALYSIS OF HOSPITAL PATIENT REVIEWS USING AN ATTENTION-ENHANCED BIDIRECTIONAL LONG SHORT-TERM MEMORY NETWORK

Abstract

Patient reviews on public portals provide much more detail of hospital care than the satisfaction surveys do, and a single polar score eradicates the information about which services satisfy or disappoint patients. Aspect-based sentiment analysis (ABSA) is a promising way to actionable health-care analytics as it associates opinions with named service aspects. The typical number of annotated items in existing hospital studies is a limited number of thousands, omitting the neutral class or the least common features, or requiring intensive computation, which are usually required. This paper addresses the challenges associated with large-scale ABSA on the hospital reviews corpus consisting of 87,310 cleaned reviews (from 105,636 raw reviews) each labeled with a three-class sentiment class and a multi-label annotation of the 10 service aspects and a no-aspect. We design a two-branch bidirectional long short-term memory (BiLSTM) model that consists of an attention-pooled sentiment classifier trained using class-balanced weight and a dedicated aspect detector trained using positive weight damped and validated decision thresholds. The opinion profile for aspects is derived from the predicted sentiments by taking the average of the detected sentiments. The feature-selection stage consists of vocabulary selection, sequence masking and TF-IDF feature limits. On a held-out test set of 8,763 reviews, it achieves 95.77% accuracy with a Matthews correlation coefficient of 0.833 and a macro-F1 score of 0.880 outpacing a unidirectional LSTM, a plain BiLSTM, and two TF-IDF linear baselines. The dedicated aspect detector achieves 0.962 micro-F1 and 0.937 macro-F1, while auxiliary heads within the sentiment networks only achieve 0.790 and 0.327. The greatest negative shares are for cost (40.6%), waiting time (32.2%) and test procedures (30.3%). The network contains approximately three million parameters and trains in about 6 minutes on one CPU core. Constraints are the one run training and only supervision at the review level. The results validate use of lightweight recurrent models as a viable foundation for continuous monitoring of hospital feedback; the Python code released enables the model to be replicated on other review collections.