Using Pseudo-Synonyms to Generate Embeddings for Clinical Terms

Santosh Purja Pun, Oliver Obst, Jim Basilakis, Jeewani Anupama Ginige

Abstract: Existing approaches attempt to explicitly learn clinical term embedding from clinical datasets by training a model, such as word2vec and recurrent neural network or fine-tuning a pre-trained large language model (LLM). While the corpus-based methods require exposure to a rich vocabulary in the training corpus, insufficient contextual information, in clinical terms, makes LLMs prone to failure to generate meaningful embeddings. In this regard, we propose a novel method to generate embeddings for clinical terms using pseudo-synonyms - terms that might be associated with a clinical term but not the exact synonyms. The proposed method uses an LLM as a black-box tool and requires no training or fine-tuning. To demonstrate the effectiveness of the learned embeddings, we compared our approach with existing corpus-based embedding approaches on semantic textual similarity (STS) tasks on five benchmark datasets. Our proposed method outperformed all existing approaches (https://github.com/Xujan24/pseudo-synonyms-for-clinical-term-embedding).

External IDs:doi:10.1007/978-981-96-8298-0_17