Multi-task Character-Level Attentional Networks for Medical Concept Normalization

Recognizing standard medical concepts in the colloquial text is significant for kinds of applications such as the medical question answering system. Recently, word-level neural network methods, which can learn complex informal expression features, achieved remarkable performance on this task. However, they have two main limitations: (1) Existing word-level methods cannot learn character structure features inside words and suffer from "Out-of-vocabulary" (OOV) words, which are common in noisy colloquial text. (2) Since these methods handle the normalization task as a classification issue, concept phrases are represented by category labels. Hence the word morphological information inside the concept is lost. In this work, we present a multi-task character-level attentional network model for medical concept normalization. Specifically, the character-level encoding scheme of our model can alleviate the OOV word problem. The attention mechanism can effectively exploit the word morphological information through multi-task training. It generates higher attention weights on domain-related positions in the text sequence, helping the downstream convolution focus on the characters that are related to medical concepts. To test our model, we first introduce a labeled Chinese dataset (overall 314,991 records) for this task. Other two real-world English datasets are also used. Our model outperforms state-of-the-art methods on all three datasets. Besides, by adding four types noises to the datasets, we validate the robustness of our model against common noises in the colloquial text.