Automated Social Text Annotation With Joint Multilabel Attention Networks

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Automated social text annotation aids document management and user experience on social media.
• Current methods face challenges in accurately suggesting relevant tags for shared content.

Purpose of the Study:

• To propose a novel deep learning-based method for automated social text annotation.
• To enhance document representation and leverage semantic relationships among labels for improved tagging.

Main Methods:

• Developed an attention-based neural network with semantic-based regularization, termed the joint multilabel attention network (JMAN).
• The network models document titles and content separately, using title-guided attention and semantic loss regularizers (similarity and subsumption).

Main Results:

• JMAN significantly outperformed state-of-the-art models like Bi-GRU and HAN in F1 score across four real-world datasets.
• JMAN demonstrated faster convergence and training speed compared to baseline models.
• Semantic-based loss regularizers also improved the performance of existing HAN and Bi-GRU models.

Conclusions:

• The proposed JMAN model offers a superior approach to automated social text annotation.
• Semantic-based regularization effectively leverages label correlations, enhancing tagging accuracy.
• Further research into dynamic label semantic matrix updates (JMANd) may yield performance gains but requires memory optimization.