HYBRID MULTI-DOMAIN CONVERSATIONAL FRAMEWORK USING LARGE LANGUAGE MODELS FOR ADAPTIVE DIALOGUE MANAGEMENT
DOI:
https://doi.org/10.70917/ijcisim-2026-2101Keywords:
Multi-Domain Dialogue System, Natural Language Processing (NLP), Large Language Models (LLMs), Dialogue Generation, Retrieval-Augmented Generation (RAG), Conversational AI, Context-Aware Systems, BLEU, METEOR, BERTScore, ROUGE Metrics.Abstract
The last few years have seen an explosive growth in large language models which has translated to significant improvements in the field of conversational AI. However, models that are purely generative have a tendency to hallucinate and are non-factual when answering questions of a specific domain. In order to develop a solution to this problem, the authors of this paper are proposing a Hybrid Retrieval-Augmented Generation (RAG) approach to developing a multi-domain conversational assistant based on the MultiWOZ 2.1 dataset. In order to simplify the scope of the developments and also help with the accuracy of the retrieval and the correctness of the responses, the conversational assistant is limited to the hotel and train domains. This architecture is supplemented with the following: dense semantic retrieval using ChromaDB; sparse probabilistic retrieval using BM25; cross-encoder re-ranking; structured slot extraction from large language models; conversational memory management; and in order to keep the slot extraction decoupled from the retrieval, a grounded (factual) response generation capacity, a 2 stage (interleaved) generation approach is utilized. In order to measure the accuracy and coherence of the responses that the model generates, the authors have chosen to measure the following: BLEU, ROUGE, METEOR, and F1 score. This paper shows that the proposed LLM model has a BLEU score of 30.91%, ROUGE score of 52.37%, and METEOR score of 45.12%, which shows that there is a good degree of linguistic alignment to the reference responses. In the evaluation of the experiments conducted, the authors claim that there is a significant increase in retrieval accuracy, response relevancy, and the stability of the overall AI system when using the hybrid RAG architecture over standard generative techniques. The suggested structure offers a scalable and interpretable framework for developing dependable multi-domain conversational agents.
