Incorporating Domain Knowledge Into Language Models Using Graph Convolutional Networks for Clinical Semantic Textual Similarity (Preprint)
BACKGROUND While electronic health record systems have facilitated clinical documentation in healthcare, they also introduce new challenges such as the proliferation of redundant information through copy-and-paste commands or templates. One approach to trim down bloated clinical documentation and improve clinical summarization is to identify highly similar text snippets for the goal of removing such text. OBJECTIVE We develop a natural language processing system for the task of clinical semantic textual similarity that assigns scores to pairs of clinical text snippets based on their clinical semantic similarity. METHODS We leverage recent advances in natural language processing and graph representation learning to create a model that combines linguistic and domain knowledge information from the MedSTS dataset to assess clinical semantic textual similarity. We use Bidirectional Encoder Representation from Transformers (BERT)¬–based models as text encoders for the sentence pairs in the dataset and graph convolutional networks (GCNs) as graph encoders for corresponding concept graphs constructed based on the sentences. We also explore techniques including data augmentation, ensembling, and knowledge distillation to improve the performance as measured by Pearson correlation. RESULTS Fine–tuning BERT-base and ClinicalBERT on the MedSTS dataset provided a strong baseline (0.842 and 0.848 Pearson correlation, respectively) compared to the previous year’s submissions. Our data augmentation techniques yielded moderate gains in performance, and adding a GCN–based graph encoder to incorporate the concept graphs also boosted performance, especially when the node features were initialized with pretrained knowledge graph embeddings of the concepts (0.868). As expected, ensembling improved performance, and multi–source ensembling using different language model variants, conducting knowledge distillation on the multi–source ensemble model, and taking a final ensemble of the distilled models further improved the system’s performance (0.875, 0.878, and 0.882, respectively). CONCLUSIONS We develop a system for the MedSTS clinical semantic textual similarity benchmark task by combining BERT–based text encoders and GCN–based graph encoders in order to incorporate domain knowledge into the natural language processing pipeline. We also experiment with other techniques involving data augmentation, pretrained concept embeddings, ensembling, and knowledge distillation to further increase our performance.