A dynamic goal adapted task oriented dialogue agent.

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Natural Language Processing

Background:

• Existing virtual agents (VAs) are limited to information retrieval or static goal-driven interactions.
• Real-world users often have dynamic goals that change during a task, which current VAs cannot handle.
• There is a lack of dialogue datasets capturing dynamic goal adjustments.

Purpose of the Study:

• To develop a virtual agent capable of handling dynamic user goals in real-time.
• To enhance user satisfaction by allowing goal adaptation during task completion.
• To address the limitations of static goal-driven dialogue systems.

Main Methods:

• Created a new conversational dataset (DevVA) with annotated intents, slots, and sentiment.
• Developed a Dynamic Goal Driven Dialogue Agent (DGDVA) using deep reinforcement learning.
• Incorporated a Dynamic Goal Driven Module (GDM) that uses user sentiment as feedback for goal adjustment.

Main Results:

• The DGDVA demonstrated a high task success rate (0.88) and maximum user satisfaction.
• User sentiment effectively guided the agent to resolve goal discrepancies.
• The agent successfully adapted to users' dynamic goal-setting behaviors.

Conclusions:

• The proposed sentiment-aware VA effectively handles dynamic user goals, leading to increased task success.
• DGDVA offers a more satisfactory user experience by allowing goal evolution.
• This represents a novel approach to dynamic goal adaptation in task-oriented dialogue systems.