Deep neural model with self-training for scientific keyphrase extraction.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Information Retrieval

Background:

• Scientific information extraction is vital for understanding research.
• Keyphrase extraction identifies and categorizes important terms in articles.

Purpose of the Study:

• To develop a neural network approach for scientific keyphrase extraction.
• To classify extracted keyphrases into predefined categories.
• To address the challenge of limited annotated data using self-training.

Main Methods:

• Utilized a neural network model incorporating bidirectional Long Short-Term Memory (LSTM) for sentence representation.
• Employed Conditional Random Fields (CRF) for predicting label sequences.
• Integrated a self-training method to leverage unlabeled scientific articles.

Main Results:

• Achieved promising performance on the ScienceIE and ACL keyphrase corpora.
• Demonstrated effectiveness without relying on hand-designed features or external knowledge.
• Showcased competitive results compared to state-of-the-art systems through efficient incorporation of unlabeled data.

Conclusions:

• The proposed neural model offers an effective solution for scientific keyphrase extraction.
• Self-training significantly enhances model performance by utilizing unlabeled data.
• The approach provides a robust and efficient method for keyphrase identification in scientific literature.