Uncertainty-based Self-training for Biomedical Keyphrase Extraction.

Area of Science:

• Computational linguistics
• Bioinformatics
• Information retrieval

Background:

• Automated methods are crucial for managing the growing volume of digital documents.
• Keyphrase extraction identifies salient concepts, aiding search and discovery.
• Supervised keyphrase extraction models outperform unsupervised methods but require large labeled datasets.

Purpose of the Study:

• To develop an improved keyphrase extraction method leveraging abundant unlabeled data.
• To address the limitation of small labeled datasets in supervised learning for keyphrase extraction.

Main Methods:

• Introduced a self-learning model for keyphrase extraction.
• Incorporated uncertainty estimation to select informative instances from unlabeled data.
• Augmented small labeled training sets with selected unlabeled data.

Main Results:

• The self-learning model demonstrated improved keyphrase extraction performance.
• Performance gains were observed on a publicly available biomedical dataset.
• The method surpassed existing state-of-the-art models in keyphrase extraction.

Conclusions:

• Self-learning with uncertainty estimation is effective for enhancing keyphrase extraction.
• This approach effectively utilizes large unlabeled datasets to improve model performance.
• The method offers a promising solution for scalable information extraction in large literature corpora.