Review text in recommender systems provides rich insights into user preferences and experiences that cannot be fully captured by numerical ratings alone. While recent studies have increasingly leveraged review text to enhance recommendation accuracy, most have primarily focused on improving model performance, with limited attention to quantitatively examining how specific textual elements influence rating prediction. To address this gap, this study empirically investigates the impact of review text characteristics on prediction performance in review-based recommender systems. Specifically, we employ the Unstructured Context-Aware Model (UCAM), where contextual information is replaced with review text embedded using a pre-trained BERT model. Three key textual factors are examined: review length, aspect, and emotion type. Review length is divided into quartiles, and results show that removing shorter reviews significantly degrades performance, indicating their critical role. For analysis, reviews are categorized into food, service, price, atmosphere, and location, with service and food contributing most to performance improvements, while location shows relatively low influence. Emotion types are classified based on Plutchik’s framework, revealing that removing joy, trust, and anticipation reduces performance, whereas excluding sadness slightly improves it. Overall, this study highlights the differential importance of textual features and demonstrates their potential for enhancing recommender system design.

With the rapid advancement of e-commerce platforms and the overwhelming amount of information available online, recommender systems have emerged as essential tools for helping users efficiently discover items aligned with their preferences. Online reviews, in particular, offer rich information for understanding both user preferences and item features, and have been widely utilized in existing multimodal recommender systems. However, most prior approaches have not sufficiently addressed modality-specific noise in multimodal reviews, thereby constraining their capacity to precisely represent users’ experiences. Furthermore, by merely extracting and fusing features from texts and images without considering the structural heterogeneity between modalities, these approaches often fall short in accurately predicting user preferences. To address these limitations, this study proposes multimodal aspect summarization for recommender systems (MAS-Rec), a novel model that leverages large language models (LLMs) to summarize multimodal reviews by aspect, thereby enhancing preference prediction. The proposed model is structured as follows: First, LLMs are employed to independently summarize user preferences and item features into aspects from textual and visual data, respectively. Second, these aspect-based summarizations are then encoded using a bidirectional encoder representations from transformers (BERT) to extract semantic features. Finally, a co-attention mechanism is applied to model the semantic interactions between user preferences and item features. Experimental results on two review datasets provided by Amazon.com demonstrate that MAS-Rec outperforms existing unimodal and multimodal models. These findings demonstrate that summarizing images by aspect using the LLM effectively captures the semantic connections between texts and images, thereby enabling more accurate prediction of user preferences compared to traditional approaches.