MG-SIN: Multigraph Sparse Interaction Network for Multitask Stance Detection.

Area of Science:

• Natural Language Processing
• Computational Social Science

Background:

• Stance detection identifies user support or opposition towards a target on social media.
• Existing methods often overlook word-level pragmatic dependencies, leading to performance degradation.
• Short social media texts exacerbate semantic sparsity issues in single-task learning.

Purpose of the Study:

• To propose a novel Multigraph Sparse Interaction Network (MG-SIN) for simultaneous stance detection and sentiment polarity classification.
• To address semantic sparsity and improve representation learning by leveraging multitask learning (MTL).
• To explore pragmatic dependency relationships between tasks at the word level.

Main Methods:

• Constructing two types of heterogeneous graphs: task-specific and task-related (tr-graphs).
• Developing a graph-aware module for adaptive information sharing between tasks.
• Implementing a novel sparse interaction mechanism among heterogeneous graphs within an MTL framework.

Main Results:

• MG-SIN achieved competitive improvements on two real-world datasets.
• Stance detection performance increased by up to 2.1% and 2.42%.
• Sentiment analysis performance improved by 5.26% and 3.93% compared to state-of-the-art baselines.

Conclusions:

• The proposed MG-SIN effectively enhances stance detection and sentiment analysis by leveraging MTL and graph-based modeling.
• Modeling word-level pragmatic dependencies through heterogeneous graphs boosts task-specific representations.
• The sparse interaction mechanism facilitates efficient information sharing, overcoming limitations of previous approaches.