Abstract
Smart cities need the involvement of their residents to enhance quality oflife. Conversational query-answering is an emerging approach for userengagement. There is an increasing demand of an advanced conversationalquestion-answering that goes beyond classic systems. Existing approaches haveshown that LLMs offer promising capabilities for CQA, but may struggle tocapture the nuances of conversational contexts. The new approach involvesunderstanding the content and engaging in a multi-step conversation with theuser to fulfill their needs. This paper presents a novel method to elevate theperformance of Persian Conversational question-answering (CQA) systems. Itcombines the strengths of Large Language Models (LLMs) with contextual keywordextraction. Our method extracts keywords specific to the conversational flow,providing the LLM with additional context to understand the user's intent andgenerate more relevant and coherent responses. We evaluated the effectivenessof this combined approach through various metrics, demonstrating significantimprovements in CQA performance compared to an LLM-only baseline. The proposedmethod effectively handles implicit questions, delivers contextually relevantanswers, and tackles complex questions that rely heavily on conversationalcontext. The findings indicate that our method outperformed the evaluationbenchmarks up to 8% higher than existing methods and the LLM-only baseline.