TASTE: A Context-Aware Explainable Recommendation Model

ACL ARR 2026 January Submission5738 Authors

Keywords: ontext-aware recommendation, Explainable recommendation, Aspect-based sentiment analysis, Query-aware modeling, Large language models

Abstract:

Travel recommendation involves complex, context-dependent decision-making that goes beyond static user preferences. While recent large language model (LLM)–based conversational recommender systems enable natural language interaction and explanation generation, most existing approaches treat user preferences as fixed and utilize user queries only for surface-level explanation, without adapting the underlying recommendation logic. This results in a mismatch between conversational intent and actual recommendation criteria.

In this paper, we propose TASTE (Query-Aware Travel Recommendation via Aspect-based Sentiment Profiling), a novel recommendation framework that dynamically adapts aspect-level preference weights according to natural language user queries while preserving long-term user preference profiles. TASTE constructs structured user preference representations using Aspect-Based Sentiment Analysis (ABSA) applied to review texts and combines learned aspect attention with query-derived aspect importance through a controllable weighting mechanism. Large language models are employed exclusively for query interpretation and explanation generation, ensuring that recommendation decisions remain deterministic, interpretable, and faithful to model-internal reasoning.

We conduct extensive experiments on three real-world datasets spanning restaurant, travel, and hotel domains (Yelp Restaurant, Yelp Travel-related, and TripAdvisor Hotel). Quantitative results demonstrate that TASTE achieves competitive rating prediction performance compared to strong rating- and review-based baselines, without sacrificing accuracy for interpretability. Qualitative and quantitative analyses further show that TASTE effectively captures query-aware preference shifts, producing meaningful changes in recommendation rankings for the same user under different queries. Finally, explanation quality evaluation using an LLM-as-a-Judge protocol confirms that TASTE generates coherent, relevant, and aspect-aligned explanations across domains and languages.

Overall, TASTE provides a unified framework for accurate, query-aware, and explainable travel recommendation, addressing key limitations of existing conversational recommender systems.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: Explanation faithfulness, Feature attribution

Languages Studied: English, Korean

Submission Number: 5738