Knowledge Discovery From Biomedical Literature Based On Semantic Resources And Semi-supervised Learning

Nowadays, with the rapid growth of biomedical literatures, it is difficult for biological researchers to find their related literatures from amounts of biomedical literatures. Meanwhile, there are few communications between different science disciplines as different disciplines have their own focus. It is easy to ignore the useful and potential information in the interdisciplinary field. Swanson(1986) first found a connection between Fish Oil and Raynaud’s syndrome in disjoint literature areas. The purpose of Swanson’s pioneering work is to find links that are not previously founded by researchers, and he defined the process as Literature Based Discovery (LBD). After him, many researchers put a lot of efforts to discover hypothesis links in disjoint literature areas.However, using traditional methods based on co-occurrence is difficulty to find useful and valuable information, because those methods produce too many target concepts. This paper presents a method based on semantic resources, using SemRep system to extract relationships between entities within the sentence. By employing a combination of the semantic type, concept information amount and association rules, it is effective to filter the linking and target concepts and sort the target concepts with the statistical information. The experimental results demonstrate that our method works well on the classic cases found by Swanson.On the other hand, the semantic relations extracted by SemRep are not comprehensive, as the recall rate is only55%, it may lose a lot of useful information. This paper presents a method which is based on bag of words and kernel of graph to extract the relationship between the two concepts in a sentence. The co-training is used to expand the training set for getting a good model. Compared to SemRep, this paper’s method perform well. By using the SVM classifier, we develop two models which are the model of AB and BC respectively. Compared with SemRep, the two models perform better on the classic cases found by Swanson.