LLM-Guided Review Condensation for Text-Convolutional Recommendation
Yoonhyuk Choi, Chong-kwon Kim · IEEE Access · 2026
Review-based recommender systems use natural-language feedback to alleviate sparse interaction data and to explain user preferences. A common design concatenates user and item reviews, embeds the resulting text, and applies two-dimensional text convolution before collaborative prediction. This design is effective but can be sensitive to long and noisy review documents because convolutional filters are shared across all local windows and can smooth informative phrases together with generic expressions. This article presents an expanded study of LLM-guided review condensation for text-convolutional recommendation. The proposed framework uses a large language model as an offline, domain-aware text condenser that retains preference-bearing review content before the conventional convolutional encoder is trained. The method is model-agnostic and can be attached to single-domain and cross-domain review-based recommenders without changing their collaborative filtering objectives. The study gives an explicit condensation protocol, leakage-control rules, reproducibility details, theoretical sufficient conditions, and a setting-level consistency analysis. Experiments on Amazon review benchmarks show that the proposed preprocessing consistently improves seven representative recommendation models. Across 45 paired method-dataset settings, condensation improves NDCG by 18.1 percent and HR by 12.6 percent on average. All paired settings show positive changes for both metrics, and additional aggregate and domain-sensitivity analyses show that the effect is strongest for sparse target domains and for older convolutional encoders. These results support the view that LLM-guided condensation acts as a reusable noise-reduction layer for review-based recommendation.