Leveraging Large Language Models for Domain‐Agnostic Aspect‐Aware Recommendation

The growth of e‐commerce and the corresponding increase in the number of online reviews have driven the demand for recommender systems that capture user preferences. Online reviews contain detailed evaluations of item aspects, and recent studies have employed aspect‐based sentiment analysis (ABSA) to incorporate fine‐grained user sentiment for improved personalisation. However, existing ABSA‐based recommendation approaches rely on predefined aspect categories, which prevents them from capturing user preferences for specific aspect terms and limits their generalisation across diverse domains. To address these limitations, this study proposes a novel ABSA‐based recommendation framework, domain‐agnostic aspect‐aware recommendation (DAAR), which leverages large language models (LLMs) to extract aspect terms and integrate them with sentiment information. First, aspect terms that represent the targets of user evaluation are extracted from review texts using LLMs. Then, semantic representations of the aspect terms are generated using Phrase‐BERT based on bidirectional encoder representations from transformers (BERT), and sentiment scores for each aspect term are obtained through a sentiment analysis model based on decoding‐enhanced BERT with disentangled attention (DeBERTa). Finally, the two representations are combined into a unified representation, and an attention mechanism is applied to capture the relative importance among aspect terms. The resulting representation is then integrated with user–item interactions to predict the final rating. Experimental results on the Yelp and Amazon review datasets demonstrate that DAAR outperforms the baseline models and shows strong generalisation capabilities. This indicates that the proposed framework alleviates domain dependency and more precisely captures user preferences by integrating users' sentiment towards aspect terms.