Generating, retrieving persona and generating responses for long-term open-domain dialogue.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Open-domain dialogue systems excel in short conversations but struggle with long-term context due to input length limitations.
• Existing systems often lose crucial information from earlier dialogues in multi-session chat (MSC), impacting response coherence.
• Maintaining context is vital for natural and consistent dialogue generation in extended interactions.

Purpose of the Study:

• To propose a novel framework, GRGPerDialogue, for generating contextually relevant and consistent responses in long-term conversations.
• To address the challenge of information loss in multi-session chat (MSC) by incorporating speaker persona.
• To improve the performance of dialogue systems in extended conversational scenarios.

Main Methods:

• Developed a three-stage GRGPerDialogue framework: persona generation (using Llama 2 ICL), persona retrieval (using a trained Facebook Dense Passage Retrieval model on the PUP dataset), and response generation (using GPT-2 and BART).
• Introduced a new dataset, Persona-Utterance Pair (PUP), to facilitate persona-relevant utterance retrieval.
• Leveraged In-Context Learning (ICL) for real-time persona generation and Dense Passage Retrieval (DPR) for efficient persona sentence retrieval.

Main Results:

• The GRGPerDialogue framework demonstrated superior performance over baseline models, achieving an approximate 0.6% to 1% improvement in Rouge-1 scores on a long-term dialogue dataset.
• Human evaluations corroborated the effectiveness of the proposed framework in generating fluent, consistent, and contextually relevant responses.
• The framework successfully mitigated the issue of context loss in long-term dialogues.

Conclusions:

• GRGPerDialogue effectively preserves and utilizes historical dialogue information through persona generation and retrieval.
• The proposed framework significantly enhances the quality of responses in long-term conversational AI.
• GRGPerDialogue represents a promising advancement for open-domain dialogue systems handling extended interactions.