Background Semantic textual similarity (STS) is a natural language processing (NLP) task that involves assigning a similarity score to 2 snippets of text based on their meaning. This task is particularly difficult in the domain of clinical text, which often features specialized language and the frequent use of abbreviations. Objective We created an NLP system to predict similarity scores for sentence pairs as part of the Clinical Semantic Textual Similarity track in the 2019 n2c2/OHNLP Shared Task on Challenges in Natural Language Processing for Clinical Data. We subsequently sought to analyze the intermediary token vectors extracted from our models while processing a pair of clinical sentences to identify where and how representations of semantic similarity are built in transformer models. Methods Given a clinical sentence pair, we take the average predicted similarity score across several independently fine-tuned transformers. In our model analysis we investigated the relationship between the final model’s loss and surface features of the sentence pairs and assessed the decodability and representational similarity of the token vectors generated by each model. Results Our model achieved a correlation of 0.87 with the ground-truth similarity score, reaching 6th place out of 33 teams (with a first-place score of 0.90). In detailed qualitative and quantitative analyses of the model’s loss, we identified the system’s failure to correctly model semantic similarity when both sentence pairs contain details of medical prescriptions, as well as its general tendency to overpredict semantic similarity given significant token overlap. The token vector analysis revealed divergent representational strategies for predicting textual similarity between bidirectional encoder representations from transformers (BERT)–style models and XLNet. We also found that a large amount information relevant to predicting STS can be captured using a combination of a classification token and the cosine distance between sentence-pair representations in the first layer of a transformer model that did not produce the best predictions on the test set. Conclusions We designed and trained a system that uses state-of-the-art NLP models to achieve very competitive results on a new clinical STS data set. As our approach uses no hand-crafted rules, it serves as a strong deep learning baseline for this task. Our key contribution is a detailed analysis of the model’s outputs and an investigation of the heuristic biases learned by transformer models. We suggest future improvements based on these findings. In our representational analysis we explore how different transformer models converge or diverge in their representation of semantic signals as the tokens of the sentences are augmented by successive layers. This analysis sheds light on how these “black box” models integrate semantic similarity information in intermediate layers, and points to new research directions in model distillation and sentence embedding extraction for applications in clinical NLP.

Despite recent advances in the application of deep neural networks to various kinds of medical data, extracting information from unstructured textual sources remains a challenging task. The challenges of training and interpreting document classification models are amplified when dealing with small and highly technical datasets, as are common in the clinical domain. Using a dataset of de-identified clinical letters gathered at a memory clinic, we construct several recurrent neural network models for letter classification, and evaluate them on their ability to build meaningful representations of the documents and predict patients' diagnoses. Additionally, we probe sentence embedding models in order to build a humaninterpretable representation of the neural network's features, using a simple and intuitive technique based on perturbative approaches to sentence importance. In addition to showing which sentences in a document are most informative about the patient's condition, this method reveals the types of sentences that lead the model to make incorrect diagnoses. Furthermore, we identify clusters of sentences in the embedding space that correlate strongly with importance scores for each clinical diagnosis class.