基于半监督学习的蛋白质关系抽取研究

Research of protein-protein interaction extraction based on semi-supervised learning

使用半监督学习方法中的自训练、协同训练方法,利用少量已标注样本和大量未标注样本来完成蛋白质关系抽取的任务.首先使用基于词特征的SVM(support vector machine)模型进行自训练,然后使用基于词特征的SVM模型和基于依存树特征的SVM模型进行协同训练.通过对4个语料的实验,验证了自训练及协同训练方法在蛋白质关系抽取领域中的应用效果.相比于自训练,协同训练可以通过两个相对独立的视图相互补充、相互学习,进而可以有效利用未标注数据.

Semi-supervised learning methods including self-training and co-training were shown in the task of PPI on how to alleviate the tag burden as much as possible. On self training a word feature based SVM model was applied; In co-training word feature and dependency tree based SVM models were used. Experiments on four copra showed that the two methods were effective in reducing the amount of labeling PPI（protein-protein interaction）. Comparing self training to co-training can be more effective by using two separated views.