HADSF: Aspect Aware Semantic Control for Explainable Recommendation

TL;DR

HADSF addresses the challenges of explainable review-based recommendation by introducing a two-stage, aspect-aware extraction framework that constrains LLM outputs to a compact corpus-level vocabulary. It couples Stage I semantic vocabulary construction with Stage II dynamic, user-aware extraction to produce structured aspect–opinion triples, while introducing ADR and OFR to quantify hallucination and grounding. Across ~3 million reviews and LLMs from $1.5\mathrm{B}$ to $70\mathrm{B}$ parameters, HADSF yields consistent rating-prediction improvements and enables smaller models to approach larger-model performance under practical deployment constraints. The work also provides an open-source toolkit, enabling reproducible research on hallucination-aware, LLM-enhanced explainable recommendation and suggesting guidelines for prompt design and model selection.

Abstract

Recent advances in large language models (LLMs) promise more effective information extraction for review-based recommender systems, yet current methods still (i) mine free-form reviews without scope control, producing redundant and noisy representations, (ii) lack principled metrics that link LLM hallucination to downstream effectiveness, and (iii) leave the cost-quality trade-off across model scales largely unexplored. We address these gaps with the Hyper-Adaptive Dual-Stage Semantic Framework (HADSF), a two-stage approach that first induces a compact, corpus-level aspect vocabulary via adaptive selection and then performs vocabulary-guided, explicitly constrained extraction of structured aspect-opinion triples. To assess the fidelity of the resulting representations, we introduce Aspect Drift Rate (ADR) and Opinion Fidelity Rate (OFR) and empirically uncover a nonmonotonic relationship between hallucination severity and rating prediction error. Experiments on approximately 3 million reviews across LLMs spanning 1.5B-70B parameters show that, when integrated into standard rating predictors, HADSF yields consistent reductions in prediction error and enables smaller models to achieve competitive performance in representative deployment scenarios. We release code, data pipelines, and metric implementations to support reproducible research on hallucination-aware, LLM-enhanced explainable recommendation. Code is available at https://github.com/niez233/HADSF

Figures (12)

• Figure 1: Comparison of aspect–opinion extraction across traditional neural models, mainstream LLM methods and our HADSF. Color‐coded highlights denote different aspect–opinions.
• Figure 2: Framework overview. In the above section, reviews are compressed, global aspects are extracted, and the Top-N personalized aspects are selected; In the below section, the selected aspects and the target review are fed into an LLM to extract fine-grained aspect–opinion pairs and dynamically update the user’s aspect memory.
• Figure 3: Comparison of feature independence of semantic aspects before and after clustering.
• Figure 4: Identification process of Aspect Drift and Opinion Fidelity.
• Figure 5: Ablation of the HADSF.