一种简化门控结构的增强序列文本语义匹配模型研究

Research on a Text Semantic Matching Model of ESIM with Simplified Gating Structure

在自然语言处理的文本相似度匹配方面,针对长短期记忆网络拥有多个控制门层,导致其在训练过程中需要一定的硬件计算能力和计算时间成本,提出一种基于Bi-GRU的改进ESIM文本相似度匹配模型。该模型在双向LSTM(BiLSTM)的ESIM模型的基础上,通过Bi-GRU神经网络进行数据训练,提高模型的训练性能。实验表明,在公开数据集QA_corpus和LCQMC上分别进行测试,改进后的ESIM模型较之原先模型,在结果数据对比图中,绝大部分组的损失函数数值均小于原先模型,准确率数值均大于原先模型。

In terms of text similarity matching in natural language processing,the long and short-term memory network has multiple control gate layers,which requires a certain amount of hardware computing power and computing time cost during the training process.Aiming at these problems,this paper proposes an improved ESIM(Enhanced Sequential Inference Model)text similarity matching model based on Bi-GRU.Based on the ESIM model of bidirectional LSTM(Long Short-Term Memory),the proposed model is trained by Bi-GRU neural network to improve the training performance of the model.The improved ESIM model is tested on QA_corpus and LCQMC(Large-scale Chinese Question Matching Corpus)respectively.Test results show that compared with the original model,the loss function values of most groups are lower than the original model,and the accuracy values are higher than the original model.