Extracting disease-symptom relationships from health question and answer forum

In this paper, we address the problem of automatically extracting disease-symptom relationships from health question-answer forums due to its usefulness for medical question answering system. To cope with the problem, we divide our main task into two subtasks since they exhibit different challenges: (1) disease-symptom extraction across sentences, (2) disease-symptom extraction within a sentence. For both subtasks, we employed machine learning approach leveraging several hand-crafted features, such as syntactic features (i.e., information from part-of-speech tags) and pre-trained word vectors. Furthermore, we basically formulate our problem as a binary classification task, in which we classify the 'indicating' relation between a pair of Symptom and Disease entity. To evaluate the performance, we also collected and annotated corpus containing 463 pairs of question-answer threads from several Indonesian health consultation websites. Our experiment shows that, as our expected, the first subtask is relatively more difficult than the second subtask. For the first subtask, the extraction of disease-symptom relation only achieved 36% in terms of F1 measure, while the second one was 76%. To the best of our knowledge, this is the first work addressing such relation extraction task for both 'across' and 'within' sentence, especially in Indonesia.