Research On Few-Shot Text Classification Based On Meta-Learning

With the booming development of deep learning in recent years,text classification algorithms based on deep learning have achieved good results in the field of natural language processing.At present,deep learning-based text classification algorithms often require a large amount of labeled data to get good accuracy.However,in some special fields,insufficient training data is an inevitable problem,and data collection is very complicated and expensive,which makes it very difficult to build a large-scale,high-quality data set with labeling.Therefore,the problem of few-shot text classification has become a hot topic of current research.Firstly,this paper investigate a important problem of few-shot text classification that the small size of annotated dataset.And this paper propose a new metalearning framework for few-shot text classification model,which can identify important lexical features to quickly improve the classification accuracy on few-shot dataset.Then,to address the problems of noise and poor labeling quality in some few-shot datasets,this paper introduces an attention mechanism based on the prototype network to reduce the impact of noise and improve the classification accuracy under the noisy environment as much as possible.The effectiveness of the two few-shot text classification models proposed in this paper is also verified by scientific methods such as comparison experiments and ablation experiments.A prototype system for few-shot text classification is also developed,which is based on a microservice architecture to ensure the robustness,usability and ease of expansion of the system.The specific research contents and contributions are as follows:(1)Propose MLADA,a domain adaptive network in a meta-learning framework.:The model utilizes two competing neural networks that play the roles of domain discriminator and meta-knowledge generator,respectively,to enhance the adaptability of the meta-learning framework to new datasets through adversarial networks.In addition,the model combines transferable features generated by the meta-knowledge generator with sentence-specific features to rapidly generate high-quality sentence representation vectors not seen in the training data.Finally,the model uses ridge regression to obtain the final classification results.(2)Propose DAPROTO,a model based on dynamic prototype network.:The model is mainly used for few-shot text classification tasks in the presence of noise.The model is based on a metric-based meta-learning model prototype network,with the addition of coarse-grained attention to highlight words related to the classification category and fine-grained attention to alleviate the feature sparsity problem,so that the model can better reduce the impact of noise.(3)Design and implement a prototype system for few-shot text classification based on microservice architecture.:The system basically implements most of the functions related to the few-shot text classification problem,such as data collection,data preprocessing,model training,model management,and provides functions such as dataset generation and online few-shot classification.The highlight of the system is the use of microservice architecture,each microservice is relatively independent from each other,and can run on different servers when deployed,but the business can keep the data and information interaction,so as to ensure the high scalability and stability of the system.In summary,this paper studies the few-shot text classification problem and proposes two classification models based on the meta-learning framework.The experimental results show that these two models achieve good text classification results on small-scale and noisy few-shot text datasets.And in this paper,a prototype system for few-shot text classification is developed by combining the microservice architecture.