Semi-supervised learning of the hidden vector state model for extracting protein-protein interactions

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• The Hidden Vector State (HVS) model, an extension of discrete Markov models, effectively encodes context using stack-oriented state vectors for protein-protein interaction extraction.
• Statistical models like HVS typically require large annotated corpora for reliable parameter estimation, which are often challenging to acquire in practice.

Purpose of the Study:

• To develop novel semi-supervised learning approaches for training the HVS model.
• To address the challenge of limited annotated data in practical applications of HVS models.
• To improve the performance and efficiency of HVS models in protein-protein interaction extraction.

Main Methods:

• Two semi-supervised learning techniques were developed: one classification-based and one expectation-maximization-based.
• These methods enable training the HVS model using a combination of annotated and unannotated corpora.
• The approaches aim to reduce the dependency on extensive manually annotated datasets.

Main Results:

• Experimental results demonstrated improved performance compared to a baseline system.
• The HVS model trained with the proposed semi-supervised methods outperformed a model trained solely on annotated data.
• The findings validate the feasibility and efficiency of the developed semi-supervised learning strategies.

Conclusions:

• The novel semi-supervised learning approaches effectively train the HVS model for protein-protein interaction extraction.
• These methods significantly enhance model performance, even with limited annotated data.
• The study confirms the practical utility and efficiency of semi-supervised learning for HVS model parameter estimation.