Effective Use of Bidirectional Language Modeling for Medical Named Entity Recognition

Biomedical named entity recognition (NER) is a fundamental task in text mining of medical documents and has a lot of applications. Existing approaches for NER require manual feature engineering in order to represent words and its corresponding contextual information. Deep learning based approaches have been gaining increasing attention in recent years as their weight parameters can be learned end-to-end without the need for hand-engineered features. These approaches rely on high-quality labeled data which is expensive to obtain. To address this issue, we investigate how to use widely available unlabeled text data to improve the performance of NER models. Specifically, we train a bidirectional language model (Bi-LM) on unlabeled data and transfer its weights to a NER model with the same architecture as the Bi-LM, which results in a better parameter initialization of the NER model. We evaluate our approach on three datasets for disease NER and show that it leads to a remarkable improvement in F1 score as compared to the model with random parameter initialization. We also show that Bi-LM weight transfer leads to faster model training. In addition, our model requires fewer training examples to achieve a particular F1 score.