Projects

A Conversation is Worth A Thousand Recommendations: A Survey of Holistic Conversational Recommendation Systems Link to heading

Paper, Github

Conversational recommendation systems (CRS) generate recommendations through an interactive process. However, not all CRS approaches use human conversations as their source of interaction data; the majority of prior CRS work simulates interactions by exchanging entity-level information. As a result, claims of prior CRS work do not generalise to real-world settings where conversations take unexpected turns, or where conversational and intent understanding is not perfect. To tackle this challenge, the research community has started to examine holistic CRS, which are trained using conversational data collected from real-world scenarios. We present a comprehensive survey of holistic CRS methods.

Zero-shot Dialogue State Tracking with Unlabelled Data Link to heading

In previous zero-shot DST, transferring learning methods are adopted while the unlabelled data in the target domain is ignored. We leverage the unlabelled data in the zero-shot DST by transforming the zero-shot problem into a few-shot problem with a two-step training strategy. Our proposed methods outperform the baseline by 8%.

Incorporating External Knowledge and Goal Guidance for LLM-based Conversational Recommender Systems (ChatCRS) Link to heading

Paper, Github

Even the most advanced LLMs struggle to generate grounded, recommendation-oriented responses, or to proactively steer a conversation through different dialogue goals. We introduce ChatCRS, a framework that decomposes conversational recommendation into a knowledge retrieval agent, which reasons over external knowledge bases through a tool-augmented approach, and a goal-planning agent that predicts the next dialogue goal. Combining knowledge grounding with explicit goal planning improves informativeness by 17% and proactivity by 27%, and boosts recommendation accuracy by up to 10x on multi-goal recommendation datasets.

Improving Conversational Recommendation with Contextual Adaptation of External Recommenders and LLM-Based Reranking (CARE) Link to heading

Paper, Github

LLMs used directly as conversational recommenders frequently suggest items outside the target catalogue, and lean almost entirely on dialogue context for content-based suggestions — neglecting the collaborative signals among item sequences that traditional recommenders capture well. CARE combines an external recommender system, acting as a domain expert grounded in collaborative signals, with an LLM functioning as a reranker over its candidates using the full conversational context. This contextual adaptation improves recommendation accuracy by an average of 54% on ReDial and 25% on INSPIRED over using either component alone, even in low-resource settings.