The Research On Improved Multi-model And Its Fusion For Few-shot Text Classification

Intelligent dialogue system has been very common in people's life,they need it to correctly identify their own problems or requests,and with increasing demand of people,the understanding of intention in intelligent dialogue system has become a greater challenge.There are some problems in the text classification of few-shot learning of intention recognition in intelligent dialogue system,such as insufficient data,insufficient feature representation and poor distance measurement effect.Based on the idea of few-shot learning,this paper proposes new schemes in distance measurement and feature representation,as well as the scheme of model fusion.The main work is as follows:1.In the scenarios of text classification of few-shot learning in intelligent system,the few-shot learning model has the problem of insufficient feature representation due to the limitation of data,resulting in poor model effect.To solve the problem of insufficient feature representation for few-shot learning,an a Bi-LSTM model based on Wide&Deep(WDAB-LSTM)is proposed to optimize the text feature representation in the scenario of intention recognition.The model consists of two parts: shallow neural network and deep neural network.The shallow neural network is used to extract shallow information such as syntactic structure,and the deep neural network is used to extract deep information such as semantic information.At the same time,WDAB-LSTM model for text feature representation includes the attention of words in Bi-LSTM model towards LSTM,so that the output vector representation contains relevant important information in the context.2.In the scenarios of text classification of few-shot learning in intelligent system,aiming at the problem that the effect of few-shot text classification is not good enough,this paper optimizes the existing few-shot learning network.We propose WDAB-LSTM siamese network,WDAB-LSTM prototype network,WDAB-LSTM re-lation network,and WDAB-LSTM induction network.In order to solve the problem that the distance measurement module is not good enough,the distance measurement module is optimized in the above few-shot text classification model proposed in this paper.In the induction module of the proposed Wide&Deep a Bi-LSTM induction network,the original ”routing softmax” normalization is optimized into ratio,making the influence of each coefficient fairer.According to the average accuracy adopted by predecessors,the WDAB-LSTM models proposed in this paper are improved by 8.78%,4.18%,1.42%,and 1.02% respectively compared with the original model.The experimental results show that the WDAB-LSTM models proposed in this paper are relatively effective.3.This paper presents a few-shot text classification model fusion scheme: stacking model fusion based on dynamic kernel ridge regression.In order to solve the problem that single model can easily fall into local optimal solution in complex calculation,and the risk of poor effect of single model,the proposed a Bi-LSTM siamese network based on Wide&Deep,a Bi-LSTM prototype network based on Wide&Deep,a Bi-LSTM relation network based on Wide&Deep,and a Bi-LSTM induction network based on Wide&Deep are integrated.Compared with single model and existing model,the experimental results of the fusion model are improved.