MTL-DAS: Automatic Text Summarization for Domain Adaptation.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation for text summarization is challenging due to scarce annotated data in target domains.
• Existing methods often focus on single low-resource domain adaptation, limiting practical application.
• The need for robust models adaptable to multiple domains is critical.

Purpose of the Study:

• To propose a unified model, Multitask Learning for Multidomain Adaptation Summarization (MTL-DAS), for effective multidomain adaptive text summarization.
• To enhance model generalization capabilities across diverse low-resource domains using multitask learning.
• To improve the detection of summary-worthy content and acquire domain-specific knowledge and generation styles.

Main Methods:

• Utilizing the BART architecture combined with a multitask learning approach.
• Employing text reconstruction and text classification tasks to adapt to target domains.
• Evaluating the model on the AdaptSum dataset, comprising six low-resource domains.

Main Results:

• The unified MTL-DAS model demonstrates superior performance compared to separately trained models.
• The proposed method achieves effective domain adaptation in low-resource scenarios.
• MTL-DAS proves to be time-efficient and requires fewer computational resources.

Conclusions:

• MTL-DAS offers a practical and efficient solution for multidomain adaptive text summarization.
• Multitask learning significantly enhances model generalization across various domains.
• The model successfully adapts to low-resource domains, addressing a key limitation in previous approaches.