针对机器问答中多跳问题的深度学习网络模型

Deep learning network model for multi-hop problemsin machine question answering

多跳问答(multi-hop question answering,multi-hop QA)是文本问答的一项重要且具有挑战性的任务。针对现有方法在解决多跳问题时答案推理能力弱、答案寻找的准确率低等问题提出一种多跳问题的深度学习网络模型AGTNet(albert graph attention network,轻量双向编码图注意力网络)。首先在神经网络隐藏层使用参数共享和矩阵分解技术,然后使用点积计算方式进行答案预测,最后使用已标注的数据集对AGTNet模型进行训练验证。试验结果表明,本模型经过训练后在测试集上的F_(1)值达到70.4;与现有的多跳问答推理模型相比,本模型拥有较优的实体级推理能力,能够有效提高多跳问答推理能力,从而提升了问答系统的响应速度和准确率。本研究结果为问答系统和多轮对话机器人的研发提供了理论依据。

Multi-hop question answering(multi-hop QA)is an important and challenging task in text question answering.Aiming at the problems that the existing methods are afflicted by weak reasoning ability and low accuracy in answer finding when solving multi-hop problems,a deep learning network model was proposed in the name of AGTNet(albert graph attention network)for multi-hop problems in the field of question answering.Firstly,parameter sharing and matrix factorization techniques were employed in the neural network hidden layer,and then the dot product calculation method was used to predict the answer.Finally,the labeled data sets were applied to train and verify the AGTNet model.The experimental results show that the F_(1) value of the model on the test set reaches 70.4 after training.Compared with the existing multi-hop question answering reasoning model,the proposed model boasts better entity-level reasoning ability,which can effectively improve the multi-hop question answering reasoning ability,so as to improve the response speed and accuracy of the question answering system.The results of this study provide a theoretical basis for research and development of question answering system and multi-round dialogue robot.