Enhancing Multiparty Dialog Discourse Parsing With Dynamic Task-Adaptive Graph Transformer and Difficulty-Aware Task...

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Artificial Intelligence

Background:

• Multiparty dialog discourse parsing (MDDP) is crucial for understanding conversational structure.
• Existing methods face data sparsity challenges and rely on costly manual annotations.
• Multitask learning has been explored but often requires fine-grained, specific annotations.

Purpose of the Study:

• To propose an innovative multitask learning approach, DTGT-DTS, for MDDP.
• To enhance discourse parsing by leveraging related tasks without additional annotations.
• To address data sparsity and improve practical applicability in MDDP.

Main Methods:

• Developed a dynamic task-adaptive graph transformer (DTGT) to capture shared and private links between tasks.
• Implemented a difficulty-aware task scheduling (DTS) strategy for optimized multitask training.
• Utilized neighboring tasks like addressee recognition and speaker identification.

Main Results:

• DTGT-DTS achieved significant performance improvements on the Molweni and STAC datasets.
• A 6.07% and 5.31% increase in link identification accuracy was observed.
• Relation recognition accuracy improved by 7.27% and 6.02% respectively.

Conclusions:

• The proposed DTGT-DTS model effectively enhances multiparty dialog discourse parsing.
• Leveraging neighboring tasks and dynamic scheduling overcomes data sparsity without extra annotations.
• This approach offers a more practical and efficient solution for discourse parsing.