Improving News Recommendations through Hybrid Sentiment Modelling and Reinforcement Learning

arXiv:2601.02372v1 Announce Type: new
Abstract: News recommendation systems rely on automated sentiment analysis to personalise content and enhance user engagement. Conventional approaches often struggle with ambiguity, lexicon inconsistencies, and limited contextual understanding, particularly in multi-source news environments. Existing models typically treat sentiment as a secondary feature, reducing their ability to adapt to users’ affective preferences. To address these limitations, this study develops an adaptive, sentiment-aware news recommendation framework by integrating hybrid sentiment analysis with reinforcement learning. Using the BBC News dataset, a hybrid sentiment model combines VADER, AFINN, TextBlob, and SentiWordNet scores to generate robust article-level sentiment estimates. Articles are categorised as positive, negative, or neutral, and these sentiment states are embedded within a Q-learning architecture to guide the agent in learning optimal recommendation policies. The proposed system effectively identifies and recommends articles with aligned emotional profiles while continuously improving personalisation through iterative Q-learning updates. The results demonstrate that coupling hybrid sentiment modelling with reinforcement learning provides a feasible, interpretable, and adaptive approach for user-centred news recommendation.