Deep Learning Based Medical Named Entity Recognition

The combination of deep learning and medical domain attracted attentions from researchers.People trying to retrieve valuable information from a massive amount of medical text data,for example,construct knowledge graph from the texts and then visualize the knowledge of medical domain.We can also design the automatic dialogue and question answering system based on the knowledge graph,which can provide assistance for patients and doctors to quickly diagnose according to the existing knowledge.And the named entity recognition and relation extraction are the pre-task of knowledge graph.Because the propagation of the error when using word segmentation-> NER pipeline,basically the character-based models outperform the word-based models.In general domain,the performance of model can be improve by the lexicon.While most of the lexicon-based NER models people proposed now are hard to be calculate in batch,which make the training and inference speed quite slow.Not only but also,during the task of Named entity recognition in Chinese medical text,we found that the existence of amounts of professional oriented words,which makes the general domain lexicon based models are unable to get a great result as on other general datasets.Image semantic segmentation and named entity recognition can both be treated as sequence labeling problem.In semantic segmentation task,we already have some basic common solution based on deep learning which can be used for reference.Meanwhile,most of the deep learning based NER methods,take the conditional random field(CRF)as decoder in the decode layer,which make the inference speed getting slow with long text.In this paper,our work is as follows:(1)Compare the similarity between image semantic segmentation task and named entity recognition task,propose to draw lessons from the methods for semantic segmentation,trying to apply FCN model on NER with the hints from image semantic segmentation.FCN model are only composed of several convolution layers and transpose convolution layers,and is able to combine high-level features with low-level features of the text.At the same time,it achieves fast computing speed,and has competitive performance compared with the benchmark model commonly used in the industry.The feasibility of replacing CRF layer is tested.The label based attention mechanism proposed by predecessors is used as decoding layer to improve and add self attention calculation,so that the attention mechanism can be combined with global information.The attention mechanism of multi tap is added to further improve the performance of decoding layer.At the same time,the performance is similar to that of conditional random field decoding.