Vector representations of multi-word terms for semantic relatedness

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Accurate semantic similarity and relatedness are crucial for the biomedical domain.
• Distributional context vectors are widely used for representing biomedical terms.
• Effective aggregation of multi-word terms into vectors remains a challenge.

Purpose of the Study:

• To compare various multi-word term aggregation methods for distributional context vectors.
• To evaluate these methods on semantic similarity and relatedness tasks in the biomedical domain.
• To assess the impact of dimensionality reduction techniques on vector performance.

Main Methods:

• Compared summation, mean aggregation, compoundify, and MetaMap for multi-word term vectors.
• Evaluated baseline co-occurrence vectors against dimensionality-reduced vectors (SVD, word2vec CBOW, skip-gram).
• Determined optimal vector dimensionalities for each technique.

Main Results:

• No statistically significant difference was found between the tested multi-word term aggregation methods.
• Dimensionality reduction techniques like SVD and word2vec embeddings showed comparable performance.
• Optimal vector dimensionalities varied across methods and datasets.

Conclusions:

• Flexibility exists in choosing multi-word term aggregation methods for biomedical NLP tasks.
• The findings suggest that computationally expensive corpus preprocessing may be avoidable.
• The study achieved state-of-the-art results on standard evaluation datasets.