Advancing Faithfulness of Large Language Models in Goal-Oriented Dialogue Question Answering

Goal-oriented dialogue systems, such as assistant chatbots and conversational AI systems, have gained prominence for their question-answering capabilities, often utilizing large language models (LLMs) as knowledge bases. However, these systems face limitations when knowledge outside their intrinsic scope is required. In this paper we address these limitations by designing more faithful and useful systems that can accurately respond to users based on external information. Guided by reference-free evaluation metrics instead of traditional word-overlap metrics, we present two novel methods to prompt LLMs which surpass the baselines in accuracy, linguistic quality, and faithfulness. The first method employs a reranking technique using LLMs to rank document relevance without the need for fine-tuning. The second system builds upon the ReAct framework by incorporating a self-reflection mechanism, ensuring answers are grounded in retrieved content. Overall, our methods advance few-shot prompting as a way to learn to condition on external evidence, and significantly reduce hallucinations.