LatentCRS: Latent Intent-enhanced Conversational Recommendation System with Large Langauge Models

ACL ARR 2025 May Submission1833 Authors

Abstract:

Conversational Recommender Systems (CRS) leverage interactive dialogues to deliver personalized recommendations, with large language models (LLMs) enhancing their natural language understanding and response generation. However, LLMs struggle to utilize collaborative information from user behavior, which is essential for accurate recommendations. To bridge this gap, we propose a Latent Intent-enhanced Conversational Recommendation System with Large Language Models (LatentCRS) that integrates LLMs with traditional recommendation models through latent user intents. Specifically, LatentCRS employs a variational expectation-maximization framework: a recommendation model infers the intent distribution from collaborative data, which then guides the refinement of behavioral and textual information to generate recommendations. Crucially, our approach avoids costly LLM fine-tuning, ensuring computational efficiency.
Extensive experiments demonstrate that LatentCRS consistently outperforms state-of-the-art baselines in both single-turn and multi-turn recommendation scenarios.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Dialogue and Interactive Systems,NLP Applications

Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 1833