Integrating Vision-Centric Text Understanding for Conversational Recommender Systems

TL;DR

CRSs often struggle to infer user preferences from long, noisy textual contexts enriched with external information. STARCRS addresses this by a dual-path architecture that simultaneously processes full-context screen text via a vision-centric encoder and concise salient text via a standard text encoder, fused through knowledge-anchored cross-attention and adaptive gating. The method uses multi-path entity representations (KG, text, vision) with contrastive alignment and prompt learning to optimize both recommendation and conversation generation. Extensive experiments on ReDial and INSPIRED show state-of-the-art performance in both recommendation accuracy and response quality, highlighting STARCRS' robustness to noise and its practical viability without relying on external image data.

Abstract

Conversational Recommender Systems (CRSs) have attracted growing attention for their ability to deliver personalized recommendations through natural language interactions. To more accurately infer user preferences from multi-turn conversations, recent works increasingly expand conversational context (e.g., by incorporating diverse entity information or retrieving related dialogues). While such context enrichment can assist preference modeling, it also introduces longer and more heterogeneous inputs, leading to practical issues such as input length constraints, text style inconsistency, and irrelevant textual noise, thereby raising the demand for stronger language understanding ability. In this paper, we propose STARCRS, a Screen-Text-AwaRe Conversational Recommender System that integrates two complementary text understanding modes: (1) a screen-reading pathway that encodes auxiliary textual information as visual tokens, mimicking skim reading on a screen, and (2) an LLM-based textual pathway that focuses on a limited set of critical content for fine-grained reasoning. We design a knowledge-anchored fusion framework that combines contrastive alignment, cross-attention interaction, and adaptive gating to integrate the two modes for improved preference modeling and response generation. Extensive experiments on two widely used benchmarks demonstrate that STARCRS consistently improves both recommendation accuracy and generated response quality.

Figures (4)

• Figure 1: Problems in CRS's textual materials.
• Figure 2: The framework of STARCRS.
• Figure 3: Prompt template and entity description example.
• Figure 4: The impact of important hyper-parameters on INSPIRED. Similar trends can be observed on other datasets.