Phenotype Extraction Based on Word Embedding to Sentence Embedding Cascaded Approach.

Area of Science:

• Bioinformatics
• Computational Biology
• Plant Science

Background:

• Phenotypic descriptions in biomedical literature are crucial for named entity recognition but often use complex semantics.
• Existing methods struggle to accurately identify diverse and complicated phenotypic expressions.
• There is a need for advanced computational approaches to extract plant phenotype information effectively.

Purpose of the Study:

• To propose and evaluate a novel approach for identifying plant phenotypes using a cascaded word and sentence embedding method.
• To accurately recognize a variety of complicated phenotypic expressions in biomedical text.
• To compare the performance of different word representation models for this task.

Main Methods:

• A cascaded approach combining word embedding and sentence embedding techniques was developed.
• High-frequency phenotypes were identified using a word embedding method (skip-gram with negative sampling) with original sentences as input.
• The approach was evaluated on a dataset of 56,748 PubMed abstracts for the model organism Arabidopsis thaliana.

Main Results:

• The proposed cascaded approach demonstrated superior performance in identifying plant phenotypes.
• Skip-gram with negative sampling was identified as the best-performing word representation model.
• The approach achieved a 2.588-fold increase in the number of new phenotypic descriptions identified compared to the original phenotype ontology.

Conclusions:

• The cascaded word and sentence embedding approach is effective for accurate plant phenotype identification.
• This method enhances the discovery of novel phenotypic descriptions from biomedical literature.
• The findings contribute to improved named entity recognition and data mining in plant science.