由於缺乏人工標註高品質的疾病之間關聯語料庫，在本篇論文中我們建構了一個疾病之間關聯語料庫，並用於建構與評估我們的系統。最後我們建一個末端對末端（End-to-end）的最大化邊界上下文感知神經網絡。在我們的實驗結果顯示相較於單純的卷積類神經網路而言，支持向量機達到 77.82% F1度量，高於CNN模型 2.47% F1度量。接著我們將卷積類神經網路的結果作為特徵值加入支持向量機分類元件中，檢查是否可以提升分類效果，而最好的實驗結果為 77.32% F1度量，比只使用該特徵值的支持向量機低 0.5% F1 度量，主要原因是在訓練支持向量機的同時無法同步更新類神經網路，導致分類效果沒有提升。因此我們建構一末端對末端最大化邊界上下文感知神經網絡來分類疾病關聯，達到最高的 84.34% F1度量，精確度80.65%和召回率88.39%。;In our study, we constructed a disease-association corpus then use it to build and evaluate the disease-association extraction system. Finally, we propose a max-margin context-aware neural network. The results show that the support vector machine(SVM) achieves the highest F1-measure of 77.82%. The SVM-based approach is higher than the convolutional neural networks(CNN) by F1-measure of 2.47%. Then we merge the softmax layer of CNN as feature to the SVM then check whether the performance was improved or not. However, the best result is an F1-measure of 77.32%, which is 0.5% lower than the original SVM which using only its feature. The possible reason may be the NN can’t be updated synchronously while training the SVM. Therefore, we constructed a max-margin context-aware neural network to classify disease associations and achieve the highest F1-measure of 84.34%.
