MAPLE: Enhancing Review Generation with Multi-Aspect Prompt LEarning in Explainable Recommendation

TL;DR

MAPLE addresses key gaps in explainable recommendations by introducing Multi-Aspect Prompt Learner, a two-stage training framework that leverages aspect categories as auxiliary signals to generate personalized, diverse, and factually grounded explanations. Stage 1 optimizes explanation generation with loss $L_T$, while Stage 2 trains an aspect recommender using distribution-balanced loss $L_{DB}$, combining them into $L = L_T + \alpha L_{DB}$ to enable ID-based selection. It renews Yelp restaurant review data with enriched aspect inventories and demonstrates that MAPLE improves item-wise feature coverage, factuality, and sentence-level as well as corpus-level diversity, while enabling MAPLE to function as a discrete retriever within a retriever-reader pipeline. The approach yields practical benefits for personalized explanations in recommender systems and suggests a scalable path for integrating multi-aspect controls into explainable AI, with potential applicability beyond the restaurant domain.

Abstract

The Explainable Recommendation task is designed to receive a pair of user and item and output explanations to justify why an item is recommended to a user. Many models approach review generation as a proxy for explainable recommendations. While these models can produce fluent and grammatically correct sentences, they often lack precision and fail to provide personalized, informative recommendations. To address this issue, we propose a personalized, aspect-controlled model called Multi-Aspect Prompt LEarner (MAPLE), which integrates aspect category as another input dimension to facilitate memorizing fine-grained aspect terms. Experiments conducted on two real-world review datasets in the restaurant domain demonstrate that MAPLE significantly outperforms baseline review-generation models. MAPLE excels in both text and feature diversity, ensuring that the generated content covers a wide range of aspects. Additionally, MAPLE delivers good generation quality while maintaining strong coherence and factual relevance. The code and dataset used in this paper can be found here https://github.com/Nana2929/MAPLE.git.

Figures (7)

• Figure 1: Overview of explainable recommendation: A simple illustration of the "diversity-factuality dilemma".
• Figure 2: Automated sentiment-analysis pipeline for Multi-Aspect Review Segmentation.
• Figure 3: The proposed MAPLE architecture. MAPLE is trained on Stage 1: Explanation Generation and Stage 2: Aspect Recommendation as a unified architecture. During inference, MAPLE exploits the trained Aspect-Recommendation Model (Rec. Model) in Stage 2 to predict an aspect distribution $\hat{c}_{u, i}$ and obtain an aspect signal $\hat{c}$, and then feed it back to Stage 1 for generation.
• Figure 4: The t-SNE results on MAPLE trained on Yelp19 (fold 1) with S@3 strategy. The inverted triangle symbol indicates the aspect category prompt; the dots with the same color indicate the feature words belonging to the aspect category. We plot the 20 nearest feature words for each category.
• Figure 5: The long-tail dataset distribution of Yelp19 and Yelp23.