Advancing conversational information system: enhanced context understanding, knowledge grounding, and synergistic framework

Abstract

The increasing prevalence of conversational interfaces necessitates the development of robust Conversational Information Systems (CIS) capable of understanding context-dependent user needs and providing accurate, grounded information through interactive dialogue. However, designing effective CIS faces significant challenges, including resolving linguistic ambiguity, managing evolving context, ensuring knowledge grounding (especially with Large Language Models), developing effective interaction strategies, and creating suitable evaluation methodologies. This thesis addresses these challenges across core CIS domains and explores the extension of underlying principles to complex information processing tasks mediated by collaborating intelligent agents. ☐ We first investigate ambiguity resolution in Conversational Text Retrieval (CTR). We propose a mixed-initiative framework leveraging user feedback for clarification and introduce ZeQR, a novel zero-shot query reformulation technique that reframes context understanding as a machine reading comprehension task, eliminating the need for conversational training data. Subsequently, we focus on Conversational Recommender Systems (CRS). We introduce Behavior Alignment, a novel metric for evaluating the strategic alignment of LLM-based CRS with human interaction patterns, addressing shortcomings in existing evaluation methods. We also develop ReGeS, a synergistic Retrieval-Augmented Generation (RAG) framework that enhances knowledge grounding and recommendation accuracy through reciprocal interaction between retrieval and generation components. ☐ Finally, recognizing that the principles of iterative, context-aware, and collaborative information processing extend beyond human-machine dialogue, we explore their application in multi-agent systems. We present DocAgent, a multi-agent collaborative system employing dependency-aware processing to generate grounded documentation for complex source code. This serves as a case study demonstrating how CIS interaction paradigms can be adapted for orchestrating agent-agent collaboration on sophisticated, structured information generation tasks. ☐ Collectively, this research advances the state-of-the-art by providing novel methods for context understanding, knowledge grounding, system evaluation, and synergistic design in CIS. It highlights the evolution of conversational paradigms towards multi-agent collaboration and contributes to building more intelligent, reliable, and effective systems for interactive information access and processing.