Research On Few-Shot Learning For Relation Extraction With Structure Information

There is a large quantity of valuable unstructured text data on the Internet,which needs to be converted into structured knowledge using information extraction technology.As one of the basic tasks of information extraction,relation extraction aims to transform unstructured natural language text into structured triples for efficient processing,storage,and retrieval.The existing relation extraction methods based on supervised learning require sufficient and fully labeled training data,but there is often a lack of abundant manually labeled data in real-world scenarios.Although the distant supervision model can automatically obtain a large amount of data by aligning relation triples to the knowledge base,it still cannot fundamentally solve the long-tail problem of instance distribution.In view of the above challenges,this thesis proposes two methods from two perspectives under the setting of few-shot scenarios:(1)A method of few-shot relation extraction with structure information.In order to solve the long-tail problem,this model uses non-parametric estimation for the characteristics of few-shot scenarios.The non-parametric estimation method considers that each class has its own class prototype,and calculates the similarity between the query instances and the class prototype through a pre-defined metric function.Because the general feature extraction method often ignores the structure information of the sentence,this method models the dependency tree of the sentence,uses the graph convolution network to extract the structure information from the dependency tree,and integrates the structure information into few-shot relation extraction.(2)Few-shot relation extraction method based on dynamic metrics.Since the metric function in the non-parametric estimation is artificially pre-defined,it cannot express the similarity very accurately,thus the relation between entity pairs cannot be predicted better.In order to solve the above problems,this model uses a deep learning method to learn a non-linear metric function,thereby performing a more comprehensive dynamic measurement of the similarity between query instances and support instances.Besides,in order to further verify the effectiveness of the structure information for few-shot relation extraction task,this thesis also integrates the structure information sentence into the dynamic metric-based model.In order to verify the performance of the above two algorithms,this thesis has conducted a lot of experiments and analysis on publicly used English datasets in relation extraction field.Experiment results show that both algorithms have improved the performance of few-shot relation extraction,and effectively alleviated the long-tail problem of instance distribution to a certain extent.In addition,the algorithm proposed in this thesis participated in the 2018 TAC Drug-Drug Interaction Relation Extraction Competition and achieved the first place.The algorithm was also applied to the knowledge computing engine of China Engineering Science and Technology Knowledge Center project.