Consistency-Aware Multi-Task Learning for Joint Sentiment Classification and Rating Prediction

Abstract

Online product reviews carry two related but usually separately modelled signals: textual sentiment and a numerical star rating. Training independent models for each signal ignores their semantic connection and can produce outputs in which the predicted sentiment and the predicted rating contradict one another. This study addresses that gap through a systematic empirical comparison built on a Consistency-Aware Multi-Task Learning (CAMTL) setup that jointly performs three-class sentiment classification (negative, neutral, positive) and five-class star-rating prediction (1–5) from a single shared text encoder, benchmarking three encoder architectures, Bidirectional Long Short-Term Memory (BiLSTM), Temporal Convolutional Network (TCN), and a Hybrid BiLSTM-TCN, each of which was trained both with and without a probability-level consistency penalty. The framework was trained and evaluated on 568,453 Amazon Fine Food Reviews using a 72/8/20 stratified split. Among the six resulting configurations, TCN without the consistency penalty obtained the highest weighted accuracy on both tasks (89.21% sentiment, 75.43% rating; macro-AUC 0.941 and 0.895) while training roughly 7.2 times faster than BiLSTM and 9.6 times faster than the slowest Hybrid configuration. BiLSTM produced the highest inter-task Consistency Rate (97.39%), and adding the consistency penalty shifted accuracy by at most 0.24 percentage points while slightly lowering the Consistency Rate for every architecture. Because sentiment labels were derived from the same numerical scores used as rating targets, the reported Consistency Rates should be interpreted cautiously rather than as proof of independently learned semantic alignment. Future work should replace this dependency with independently annotated labels and extend the comparison to macro-averaged metrics, cross-domain corpora, and pre-trained encoders.