Generalized Zero-shot Text Classification Research And Implement

Classification task are often faced with "long tail problem",and some categories even may have not any corresponding labeled samples,which is called zero-shot.On the other hand,the deep learning model is limited to recognizing samples seen during training,which fails to classify the new emerging categories.The application value is limited due to the generalized zero-shot scenario when it is deploying.This paper studies the generalized zero-shot text classification(GZSTC)problem.We design and implement a model with the ability of GZSTC,which can deal with new emerging categories in real time,and will not forget the previous classification ability.The main contributions are summarized as follows:(1)We propose a meta learning framework for generalized zero-shot text classification.Through problem and derivation,the proposed method is capable of real-time adapting to new GZSTC task.(2)We propose a generalized zero-shot text classification algorithm LTA-G.A generative network is used to generate the parameters of the new parameters based on the corresponding task,which is combined with the original parameters to form a new adapted classifier to deal with the new generalized zero-shot classification task.(3)We propose a gradient prediction generalized zero-shot text classification algorithm LTA-G.The gradient estimator network is used to output the gradient of the classifier parameters corresponding to the task,which implements the adaptation process of the classifier parameters.(4)We implement a FAQ application with the ability of generalized zero-shot text classification.Based on the LTA algorithm,the prototype system verifies the effectiveness of the algorithm model and provides a complete practical application tool for merchants.