Dialog Generation with Conversational Agent in Task-Oriented Context using a Transformer Architecture

The use of conversational agents has become increasingly popular in recent years due to their ability to mimic humanlike interactions in Human-Computer Interaction(HCI) and provide personalized assistance to users. However, creating effective dialogues between humans and conversational agents remains a challenging task, particularly in the context of task-oriented applications. This is because such applications require agents to understand complex user requests and generate appropriate responses that take into account the user’s goals, preferences, and constraints. To address this challenge, we propose a new model, MegaT, an adaptation of the LongT5 (Long Text-ToText-Transfer Transformer) architecture, a transformer-based language processing model well known for its performance in a lot of Natural Language Processing (NLP) tasks. This involves designing and implementing a task-oriented conversational agent trained on annotated dialogues related to specific tasks. The agent’s performance is evaluated and analyzed using metrics such as belief accuracy, belief loss, response accuracy, and response loss. Experimental results demonstrate that MegaT comparably outperforms the T5-based agent in terms of generating accurate, fluent, and coherent responses to user queries, as well as handling longer sequences of text and producing more informative and engaging responses. Among the improvement factors, the most significant one comes from reducing the Belief Loss and Response loss by at least around 50 points across the board. This Study provides insights into the development of more effective conversational agents by leveraging the LongT5 model for generating high-quality task-oriented dialogues.