Research On Few-shot Learning Algorithm And Its Application Of Relation Extraction

With the development of deep neural networks,deep learning technology is widely used in various fields such as computer vision,speech,and natural language processing.The success of this technology benefits from the acquisition of a large amount of labeled data.How to learn from few numbers of samples is a key issue to be solved in the field of deep learning.Therefore,few-shot learning arises at the historic moment.Few-shot learning is also widespread in the field of relation extraction.The relation extraction task aims to classify the relationship between two given entities according to their relevant context,which is a very important task in natural language processing.Relation extraction has received great attention in recent years.However,relation extraction often suffers from the long-tail problem,that is,there are very few instances of a large of relationships.How to learn from a few examples is also a very important issue in relation extraction.Therefore,few-shot learning has become a new challenge in the field of deep learning and relation extraction.Many scholars have proposed a series of methods to solve the few-shot learning problem of deep learning,and conducted research on the problem of few-shot relation extraction.However,the current few-shot learning algorithms still cannot solve the following problems:（1）Overfitting caused by too few training samples.（2）Compared with traditional deep learning,the performance of these algorithms needs to be further improved.（3）Since text cannot be directly calculated,traditional few-shot learning algorithms cannot be directly applied to relation extraction tasks.The topic of the thesis is from a project of the National Natural Science Foundation of China.Regarding the issues above,a new few-shot learning algorithm and a new few-shot relation extraction algorithm are proposed.（1）Aiming at the problem of low accuracy of few-shot learning algorithms,a two-way dilated convolutional network with residual attention mechanism is proposed in this thesis.It more fully extracts features of different sizes in the data and improves the network’s attention to important features.The accuracy of the algorithm is effectively improved.In view of the overfitting of current few-shot learning algorithms,a loss function that adds a l₂ regularization term is proposed,which alleviates the problem of over-fitting by limiting the complexity of the model.Aiming at the problem of lack of supervision information for few-shot learning,a self-supervised auxiliary task is added to pre-train the model,which augments the supervision information.Based on the above ideas,a self-supervised based meta transfer few-shot learning algorithm SS-MTL is proposed in this thesis.（2）The model of SS-MTL few-shot learning algorithm proposed in this thesis is too complicated and low in efficiency.In order to ensure the efficiency and accuracy of applying few-shot learning algorithms to the relation extraction tasks,the simple meta-learning algorithm MAML is applied to the relation extraction tasks.Aiming at the problem that the semantic information of text data cannot be used effectively in the few-shot learning algorithms,an instance re-representation method based on matching degree is proposed,which highlights the training samples that are beneficial to the test samples.A task filtering strategy is added to effectively improve the accuracy of the algorithm.Aiming at the problem of low parameter adaptation efficiency caused by the shared learning rate of MAML,a rapid adaptation method of model parameters based on the fusion of class name and instance information is proposed in this thesis.Different learning rates for each instance are calculated to make the model parameters be better quickly adapted.Based on the above ideas,an information fusion meta-learning few-shot relation extraction algorithm IFML-FSRE is proposed.（3）The two algorithms proposed in this thesis are compared respectively with the latest related algorithms based on authoritative public data sets.The few-shot learning algorithm SS-MTL proposed in this thesis is used for bird recognition and character recognition.Experimental results shows that the algorithms proposed in this thesis have achieved the latest performance.Since the SS-MTL algorithm proposed in this thesis and most of the few-shot learning algorithms are aimed at the field of computer vision,and text cannot be directly calculated,these algorithms cannot be directly applied to the field of relation extraction.The author’s future efforts are to further improve the few-shot learning algorithm proposed in this article so that it can be applied to both computer vision and text fields.And we will verify the performance of the few-shot relationship extraction algorithm proposed in this thesis on the real long-tailed data set.