Embedding of semantic predications.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Artificial Intelligence

Background:

• Generating distributed vector representations of biomedical concepts is crucial for knowledge extraction.
• Existing methods like Predication-based Semantic Indexing (PSI) use structured knowledge (semantic predications).
• Neural-probabilistic language models excel at generating term representations from free text.

Purpose of the Study:

• To evaluate if insights from neural-probabilistic language models can enhance vector representations of biomedical concepts.
• To introduce a novel neural-probabilistic approach, Embedding of Semantic Predications (ESP), for encoding semantic predications.
• To compare the performance of ESP against PSI across various biomedical knowledge tasks.

Main Methods:

• Adapted the Skipgram with Negative Sampling (SGNS) algorithm to create the Embedding of Semantic Predications (ESP) model.
• Encoded biomedical concepts as subject-relation-object triplets (semantic predications).
• Evaluated ESP and PSI on information recovery, semantic similarity/relatedness estimation, and therapeutic/harmful relationship identification.

Main Results:

• ESP demonstrated advantages over PSI in specific tasks, indicating improved vector representation quality.
• Performance differences between ESP and PSI varied depending on the task.
• The study identified contexts where the computational cost of ESP is justified by performance gains.

Conclusions:

• Embedding of Semantic Predications (ESP) offers a promising neural-probabilistic approach for generating biomedical concept vectors.
• The choice between ESP and PSI depends on the specific downstream application and required performance.
• This research highlights the potential of integrating neural-probabilistic methods with structured biomedical knowledge.