BioWordVec, improving biomedical word embeddings with subword information and MeSH

Area of Science:

• Biomedical Natural Language Processing (BioNLP)
• Computational Linguistics
• Bioinformatics

Background:

• Distributed word representations are crucial for BioNLP, text mining, and information retrieval.
• Traditional word embeddings are computed from unlabeled text, often overlooking internal word structure and structured resources like ontologies.
• Leveraging these overlooked data sources can potentially enhance word representation quality.

Purpose of the Study:

• To introduce BioWordVec, a novel set of biomedical word vectors/embeddings.
• To combine subword information from unlabeled biomedical text with the Medical Subject Headings (MeSH) controlled vocabulary.
• To evaluate the effectiveness of BioWordVec in improving biomedical NLP tasks.

Main Methods:

• Developed BioWordVec by integrating subword information from biomedical corpora with MeSH vocabulary.
• Generated biomedical word embeddings using a combination of unsupervised and ontology-guided approaches.
• Assessed the performance of BioWordVec across multiple biomedical NLP tasks, including text classification and named entity recognition.

Main Results:

• BioWordVec embeddings demonstrated superior performance compared to existing state-of-the-art methods in various biomedical NLP tasks.
• The integration of subword information and MeSH significantly improved the quality and utility of word representations.
• Benchmarking confirmed the validity and effectiveness of the proposed word embeddings.

Conclusions:

• BioWordVec offers an improved approach to generating biomedical word embeddings by incorporating subword and ontological information.
• The developed embeddings provide a valuable resource for advancing BioNLP, text mining, and information retrieval.
• This work highlights the potential of integrating diverse data sources for enhanced domain-specific word representations.