The system can be divided into two parts. The first part is the knowledge graph question answering and the second part is if the knowledge graph cannot find the answer of the user inputted question, the second part will help to retrieve the top k most related answers by computing the medical question similarity from a medical question answer pair dataset.
Here are some useful reference links.
https://github.com/liuhuanyong/QASystemOnMedicalKG
The pre-train model will be used to load word embedding before training the BiLSTM+Attention model and HBDA model.
https://drive.google.com/file/d/0B7XkCwpI5KDYNlNUTTlSS21pQmM/edit?usp=sharing
https://github.com/google-research/bert
https://zhuanlan.zhihu.com/p/31638132
https://github.com/likejazz/Siamese-LSTM
https://github.com/LuJunru/Sentences_Pair_Similarity_Calculation_Siamese_LSTM
https://github.com/uhauha2929/examples/blob/master/Hierarchical%20Attention%20Networks%20.ipynb
https://www.medicinenet.com/medterms-medical-dictionary/article.htm
https://www.nhsinform.scot/illnesses-and-conditions/a-to-z
Model train_dev_test dataset are filter out the medical related questions from Quora question pair dataset.
https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs
https://github.com/LasseRegin/medical-question-answer-data
The total number of medical related data from Quora dataset is nearly 70000, but we randomly pick the 10000 as the (train/dev/test) dataset.
The number distribution of train: dev: test = 6:2:2
Model | Average Eval_accuracy by three times | Range of change |
---|---|---|
BERT baseline model | 0.7686 | (-0.0073, +0.0057) |
HDBA model | 0.8146 | (-0.0082, +0.0098) |
Bi-LSTM + Attention model | 0.8043 | (-0.0103, +0.0062) |
The scale of knowledge graph about 700 diseases. For each disease, there exists symptom, accompany_disease, prevent_way, cure_way and totally 6 entities.
Hierarchical attention networks for document classification (https://www.aclweb.org/anthology/N16-1174)
Siamese Recurrent Architectures for Learning Sentence Similarity (https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewPaper/12195)
This research was supported by summer scholarship funding from the Precision Driven Health research partnership.