【摘要】 针对小样本关系抽取问题，提出一种基于注意力机制的归纳网络.首先，利用归纳网络中的动态路由算法学习类别表示；其次，提出实例级别的注意力机制，用于调整支持集，并获取支持集与查询集样本之间的高级信息，进而获得与查询实例更相关的支持集样本.该模型很好地解决了训练数据不足时如何进行关系抽取的问题.在小样本关系抽取数据集FewRel上进行实验，得到的实验结果为：5-way 5-shot情形下准确率为(88.38±0.27)%,5-way 10-shot情形下准确率为(89.91±0.33)%, 10-way 5-shot情形下准确率为(77.92±0.44)%, 10-way 10-shot情形下准确率为(81.21±0.39)%.实验结果表明，该模型能适应任务并且优于其他对比模型，在小样本关系抽取中取得了优于对比模型的结果.更多还原

【Abstract】 Aiming at the problem of few-shot relation extraction, we proposed an induction network based on attention mechanism. Firstly, we used dynamic routing algorithm in induction network to learn the class representation. Secondly, we proposed instance-level attention mechanism to adjust support set and obtain high-level information between support set and query set samples, thereby obtaining the support set samples that were more relevant to the query instances. The proposed model effectively solved the problem of how to extract relationships when the training data was insufficient. The experiment was conducted on the few-shot relation extraction FewRel dataset, and the experimental results showed an accuracy rate of(88.38±0.27)% in the 5-way 5-shot case,(89.91±0.33)% in the 5-way 10-shot case,(77.92±0.44)% in the 10-way 5-shot case,(81.21±0.39)% in the 10-way 10-shot case. The experimental results show that the model can adapt to tasks and outperforms other comparative models, achieving better results than comparative models in few-shot relation extraction.更多还原