融合多种神经网络与多特征的答案排序方法

Answer sorting method combining multiple neural networks and multiple features

针对答案排序问题,提出并构建融合多种神经网络与多特征的答案排序模型。将问题和候选答案的词向量送入使用Leaky Relu激活函数的卷积神经网络进行学习,得到的学习结果与词汇特征、主题特征等相互拼接,输入到双向门控循环单元,其输出结果经由多层感知器进行处理后,通过softmax分类器得出最终答案排序的结果。实验结果表明,该模型在WikiQACorpus数据集上取得了较好实验结果,准确率略高于已有基线模型,达到74.43%。

Aiming at the problem of answer ordering,a solution ordering model combining multiple neural networks and multiple features was proposed and built.The word vectors of the questions and candidate answers were sent into the convolutional neural network for learning,in which the Leaky Relu activation function was used,and the obtained learning results were pieced toget-her with the vocabulary features and topic features,and they were inputted into the bidirectional gated recurrent units.The output results were processed by mutilayer perception,and the final answer ranking results were obtained through softmax classi-fier.Experimental results show that the proposed model achieves better performances on the WikiQACorpus dataset with slightly higher accuracy(74.43%)than the existing baseline model.