Position-aware attentional neural network for review-based recommendation.

In e-commerce, user reviews have become one of the most effective sources to improve recommendation accuracy, as they provide rich semantic signals about user preferences and item characteristics. However, most existing review-based systems recommend items mainly by exploiting lexical and semantic cues, while overlooking the positional distribution of tokens and sentences. In practice, reviews exhibit clear spatial regularities, and incorporating such patterns can enhance the ability to capture key information. To address this limitation, we propose Position-aware Attentional Neural Network (PAAN), which explicitly treats position as a central signal in review. Specifically, PAAN introduces a position-aware excitation module that integrates explicit learnable positional embeddings with contextual features to recalibrate attention and highlight informative tokens. By modeling where important information appears in reviews, PAAN improves the extraction of discriminative preference signals, especially under sparse user-item interactions where review content provides crucial auxiliary evidence. Moreover, a dual cross-attention mechanism further aligns interactions between user and item representations through enhanced review features, ensuring that relevance is learned in a structurally coherent and personalized manner. Extensive experiments demonstrate that our approach consistently outperforms existing review-based recommendation methods in rating prediction tasks. Compared to baseline methods, our model achieves an average improvement of 2.49% in MSE and 4.86% in MAE across five highly sparse datasets: Digital Music, Health and Personal Care, Home and Kitchen, Movies and TV, and Yelp. The code is available at https://github.com/YuanpengJiang/PAAN_Recommender.