A framework for semisupervised feature generation and its applications in biomedical literature mining

Area of Science:

• Machine Learning
• Natural Language Processing
• Bioinformatics

Background:

• Feature representation is crucial for machine learning and text mining.
• Supervised learning methods often struggle with sparse features in labeled data.

Purpose of the Study:

• To develop a novel framework for generating informative features from unlabeled data.
• To enhance the utility of sparse features in supervised learning models.
• To improve performance in biomedical literature mining tasks.

Main Methods:

• Introduced the feature coupling generalization (FCG) framework.
• Identified example-distinguishing features (EDFs) and class-distinguishing features (CDFs).
• Generalized EDFs into higher-level features using their coupling with CDFs in unlabeled data (over 20 GB of PubMed abstracts).

Main Results:

• FCG effectively utilizes sparse features often ignored by traditional supervised learning.
• Significant performance improvements were observed: 7.8% in gene NER, 5.0% in PPIE, and 5.8% in GO annotation TC.
• Achieved high F-scores of 89.1 and 64.5, and a normalized utility of 60.1 on benchmark datasets.

Conclusions:

• The FCG framework successfully enriches sparse features by leveraging unlabeled data.
• This approach offers a powerful method for incorporating new information and boosting performance in various text mining applications, particularly in the biomedical domain.