Research On Multi-domain Text Classification Algorithms In Low Resource Scenarios

Text classification usually refers to the automatic category analysis of text according to certain standards by using a computer.In the past,some researchers used feature engineering,which required high costs,and more experts were required to define features.In recent years,the research on deep learning has been very in-depth,and the computing power has been continuously improved.The deep learning-based model combined with large-scale data has greatly improved the effect of text classification,and sometimes even exceeds the accuracy of manual tasks in traditional tasks.Even so,the relevant research on text classification is far from over.In real scenarios,there are often situations such as lack of data or inconsistent distribution of upstream and downstream data,which is very different from the training requirements of traditional models,resulting in poor model results.The low-resource problem and multi-domain problem have become urgent problems to be studied and solved in text classification.Text classification involves a wide range of tasks and fields.Some problems,such as sentiment analysis,public opinion analysis,topic labeling,etc.,are essentially text classification problems.When different tasks face low-resource and multi-domain scenarios,the problems presented are not consistent,and the solutions are naturally different.This article hopes that when facing different tasks,it can combine tasks to give a general solution as much as possible,and also provide reference and reference for solving other different tasks.This paper conducts in-depth research on multi-domain sentiment classification tasks,few-shot multidomain entity classification tasks,and span extraction classification tasks,points out the problems existing in each task,and gives a general solution.Overall,the contributions of this paper are as follows:First,this paper systematically studies some past work on text classification in the first chapter,and finds some problems in the traditional algorithm in the task of few-shot multi-domain text classification,including the poor robustness of the system(Chapter 3),poor generalization and easy overfitting(Chapter 4),and low efficiency of the algorithm framework(Chapter 5),and the existence of these problems was confirmed in the experiments in the corresponding chapters;Second,for the problem of poor system robustness,this paper proposes a robust algorithm based on domain information decoupling for multi-domain sentiment classification tasks to alleviate the impact of data distribution inconsistencies caused by multiple domains on the model.The algorithm decouples multi-domain text features through domain contrastive learning and category contrastive learning,and enhances the robustness of the system through adversarial training.Experiments prove the advantages of the proposed model in terms of robustness and algorithmic efficiency;Third,for the problem of poor generalization of the model and easy overfitting in the source domain,this paper proposes a generalized domain transfer algorithm for few-shot entity classification tasks,which uses category prototype learning to alleviate category representation confusion in low-resource scenarios problem,and use label smoothing technology to solve the problem of unrecognizable unknown categories in multi-domain scenarios.Some subsequent experiments proved the effect of the proposed algorithm on generalization and migration ability;Fourth,in view of the inefficiency of some previous algorithm frameworks,this paper proposes a QA-based generative algorithm framework based on hint learning for the few-shot multi-domain slot filling task,which uses explicit hint learning and implicit hint learning to Take full advantage of the knowledge inherent in pre-trained models.Different from previous QA-based taxonomic span extraction classification algorithms,the proposed algorithm changes the task form,presents the results uniformly in the form of text generation,and strengthens the relationship between slot categories and slot entities through upstream pretraining tasks.This approach increases the theoretical upper limit of the model in an open way,and outputs results in a way closer to human expression,which has significantly improved generalization and training efficiency.