基于Tri-Training的事件关系分类方法研究

事件关系分类是一项研究事件之间存在何种逻辑关系的自然语言处理技术。针对事件关系分类任务中训练语料不足的问题,提出了基于Tri-Training的事件关系分类方法。该方法首先根据已标注的语料训练三个不同的分类器,以多数投票的方式从未标注集中抽取置信度较高的样本对训练集进行扩充,然后利用新的训练集重新训练分类器,反复迭代,不断完善分类模型,最终达到提升事件关系分类性能的目的。实验结果表明,以F1值为评价标准,基于Tri-Training的事件关系分类方法在四大类事件关系上的分类性能为64.36%。

Event relation classification based on Tri-Training

As one of natural language processing techniques, event relation detection aims at exploring logical relationship between pairwise events. To solve the problem of lacking enough training data in event relation detection tasks, we propose a novel approach based on Tri-Training to augment the training corpus. We firstly use labeled training data to learn three different classifiers, and then exploit majority voting method to expand training corpus with higher confidence samples, iteratively optimize the model and eventually improve the performance of event relation classification. Experimental results show that compared to other methods, the Tri Training based method achieves 64.3% F1-score over four general semantic relations.