基于GAN的跨语言文本匹配算法研究

Research on Cross-language Text Matching Algorithm Based on GAN

针对跨语言文本匹配问题,提出一种基于GAN+AT-CNN的文本匹配算法。具体则通过监督式GAN文本特征映射模型和AT-CNN文本匹配模型相结合,增加样本丰富性,简化跨语言特征映射过程,从而提高映射速度和文本匹配准确度。分别将监督式GAN特征映射模型与传统的全连接神经网络(NN)、经典机器翻译模型ConvSeq2Seq,AT-CNN文本匹配模型与Bi-LSTM、ABCNN模型进行比较。结果表明,在特征映射模型的实验中,监督式GAN特征映射模型精确度平均值较其他模型高0.12%-8.46%,较无监督式GAN映射模型精确度高30.89%;训练时间则较NN长0.2 h,较ConvSeq2Seq模型短2.2 h。而在跨语言文本匹配实验中,AT-CNN文本匹配模型精确度平均值则较其他模型高1.78-7.1,但训练时间也较其他模型高127 s~1176 s。实验证明,无论是在训练时间还是精确度上,本文使用的模型综合上都优于其他对比模型,值得应用于未来的跨语言文本匹配工作中。

Aiming at the cross-language text matching problem,this paper proposes a text matching algorithm based on GAN+AT-CNN.Specifically,the GAN text feature mapping model and AT-CNN text matching model are combined to increase the sample richness,simplify the cross-language feature mapping process,and thus improve the mapping speed and text matching accuracy.The supervised GAN feature mapping model is compared with the traditional fully connected neural network(NN),the classical machine translation model ConvSeq2Seq,and the AT-CNN text matching model with Bi-LSTM and ABCNN models.The results show that the average accuracy of supervised GAN feature mapping model is 0.12%-8.46%higher than that of other models,and 30.89%higher than that of unsupervised GAN.The training time is 0.2 h longer than that of NN and 2.2 h shorter than that of ConvSeq2Seq model.In the cross-language text matching experiment,the average accuracy of AT-CNN text matching model is 1.78-7.1 higher than that of other models,but the training time is 127 s-1176 s higher than that of other models.Experimental results show that the proposed model outperforms other comparison models in terms of both training time and accuracy,and is worthy of being applied in future cross-language text matching work.