Deep Contextualized Medical Concept Normalization in Social Media Text

Abstract

Abstract In recent times with more internet users, social media evolved as a rich source of information related to various domains including health. Mapping colloquial expressions in social media text to standard medical concepts is useful in applications like pharma-covigilance, medical question and answering etc. However, it is challenging owing to the noisy nature of social media text. To map colloquial expressions to standard medical concepts, we use BiLSTM with deep contextualized and traditional word vectors. Em-beddings from Language Models (ELMo) being a) char based, offer quality vector representations for rare and misspelled words which occur often in social media text and b) contextual and deep (i.e., contextual, as it models context into word vector representation and deep, as it is based on three layer representations), encode rich information. Evaluation on CADEC-MCN dataset shows that our model achieves 5.07% improvement in accuracy over the state-of-the-art method. Further, we show that clinical ELMo embeddings can better model medical terms compared to general and biomedical ELMo embeddings.