Projects on Victor|Li Chuang's Personal Webpage

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

Thu, 01 Jan 2026 00:00:00 +0000

When large language models are used directly as conversational recommenders, they frequently recommend items that fall outside the target item catalogue, since they have no explicit grounding in it. They also tend to rely almost entirely on the dialogue context for content-based suggestions, neglecting the collaborative relationships among item sequences that traditional recommender systems are specifically designed to capture.

CARE (Contextual Adaptation of Recommenders) addresses this by combining an external recommender system, which acts as a domain expert grounded in collaborative signals, with an LLM that functions as a reranker: the external recommender proposes a candidate set from the target catalogue, and the LLM reranks and selects among them using the full conversational context. This division of labour lets each component do what it is best at, significantly improving recommendation accuracy of LLM-based conversational recommenders by an average of 54% on ReDial and 25% on INSPIRED, even in low-resource settings.

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

Wed, 01 Jan 2025 00:00:00 +0000

Paper, Github

Large language models, even the most advanced ones, still struggle in conversational recommendation: they find it difficult to generate grounded, recommendation-oriented responses, or to proactively steer the conversation through different dialogue goals rather than passively responding to the user. As a result, LLM-based conversational recommenders often produce generic suggestions that are not well grounded in domain knowledge, and fail to guide the conversation strategically towards a successful recommendation.

We introduce ChatCRS, a framework that decomposes conversational recommendation into two specialised components: a knowledge retrieval agent, which uses a tool-augmented approach to reason over external knowledge bases, and a goal-planning agent, which predicts the next dialogue goal to keep the conversation moving productively towards a recommendation. By combining explicit knowledge grounding with goal-oriented planning, ChatCRS improves informativeness by 17% and proactivity by 27%, and boosts recommendation accuracy by up to 10x on multi-goal recommendation datasets.

Zero-shot Dialogue State Tracking with Unlabelled Data

Mon, 01 Jan 2024 00:00:00 +0000

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%.

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

Sun, 01 Jan 2023 00:00:00 +0000

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. Despite the emergence of such systems, such holistic approaches are still largely under-explored.