A Stack-Propagation Framework With Slot Filling for Multi-Domain Dialogue State Tracking

Area of Science:

• Artificial Intelligence
• Natural Language Processing

Background:

• Dialogue state tracking (DST) is crucial for task-oriented dialogue systems.
• Current end-to-end DST models often omit spoken language understanding (SLU) and slot filling.
• Slot information from SLU is vital for updating dialogue states.

Purpose of the Study:

• To improve DST performance by explicitly integrating slot filling as a subtask.
• To propose a novel stack-propagation framework for joint slot filling and DST.
• To enhance DST by utilizing key slot semantic knowledge.

Main Methods:

• Developed an end-to-end joint model integrating slot filling and DST.
• Introduced a stack-propagation framework to jointly model both tasks.
• Designed a slot-masked attention mechanism for focused information retrieval.
• Implemented a slot-value softcopy mechanism to emphasize key slot information.

Main Results:

• The proposed approach significantly outperforms previous methods in DST.
• The model demonstrates outstanding performance on two benchmark datasets.
• Explicit integration of slot filling enhances DST accuracy.

Conclusions:

• The novel stack-propagation framework effectively integrates slot filling and DST.
• Leveraging slot semantic knowledge through joint modeling improves dialogue understanding.
• This approach offers a promising direction for advancing task-oriented dialogue systems.