Extracting Drug Names from Medical Reports

In this paper, we present new results on drug names and dosage extraction from text-format medical reports written in a low-resource language. We locate and collect drug dosage information in continuous texts to enable the statistical processing of this information. As a first step, we use rule-based text processing methods to gain a base model and generate learning data for language model creation. This rule-based annotation part component includes tokenization, tagging, and chunk parsing of tags, using domain-specific grammar to locate and connect related tokens. Then, to improve the extraction of the drug and dosage-related text elements, we introduce a semi-supervised learning process to train a BERT-based language model. This deep learning-based approach is able to handle the rigidity presented in the data to be processed. To evaluate our approach, we consider several standard performance metrics. Based on our experiments, which show a 3.5% improvement in extraction, we conclude that these combinations of the rule-based and deep learning-based natural language processing approaches are a valid approach for this task to improve the recognition of the drug and dosage entities in medical texts.