Contextual word embedding models such as ELMo and BERT have dramatically improved performance for many natural language processing (NLP) tasks in recent months. However, these models have been minimally explored on specialty corpora, such as clinical text; moreover, in the clinical domain, no publicly-available pre-trained BERT models yet exist. In this work, we address this need by exploring and releasing BERT models for clinical text: one for generic clinical text and another for discharge summaries specifically. We demonstrate that using a domain-specific model yields performance improvements on 3/5 clinical NLP tasks, establishing a new state-of-the-art on the MedNLI dataset. We find that these domain-specific models are not as performant on 2 clinical de-identification tasks, and argue that this is a natural consequence of the differences between de-identified source text and synthetically non de-identified task text.

In this work, we aim to build a unifying framework for relation extraction (RE), applying this on 3 highly used datasets with the ability to be extendable to new datasets. At the moment, the domain suffers from lack of reproducibility as well as a lack of consensus on generalizable techniques. Our framework will be open-sourced and will aid in performing systematic exploration on the effect of different modeling techniques, pre-processing, training methodologies and evaluation metrics on the 3 datasets to help establish a consensus.

SemEval 2018 Task 7 tasked participants to build a system to classify two entities within a sentence into one of the 6 possible relation types. We tested 3 classes of models: Linear classifiers, Long Short-Term Memory (LSTM) models, and Convolutional Neural Network (CNN) models. Ultimately, the CNN model class proved most performant, so we specialized to this model for our final submissions. We improved performance beyond a vanilla CNN by including a variant of negative sampling, using custom word embeddings learned over a corpus of ACL articles, training over corpora of both tasks 1.1 and 1.2, using reversed feature, using part of context words beyond the entity pairs and using ensemble methods to improve our final predictions. We also tested attention based pooling, up-sampling, and data augmentation, but none improved performance. Our model achieved rank 6 out of 28 (macro-averaged F1-score: 72.7) in subtask 1.1, and rank 4 out of 20 (macro F1: 80.6) in subtask 1.2.