On The Research Of Event Extraction Algorithms Based On Few Samples

With the development of the network,it is difficult for people to acquire essential knowledge from a large amount of data,rich data types,and more complex data structures.It is crucial to use Event Extraction methods to extract structured event information from unstructured text in this context.Currently,Event Extraction models based on deep learning usually require a large amount of labeled data to optimize model parameters.In the Event Extraction task,the annotation of each sentence or chapter requires each annotator to understand the semantics of the sentence and annotate according to specific rules.Due to the high cost of labeling,the total number of samples is small,and it is not easy to extract new events.Aiming at the above problems,this topic mainly studies the Event Extraction algorithm from the following three aspects:1.Aiming at the problem that the data distribution formed by a tiny number of samples is quite different from the actual data distribution,which leads to the low generalization ability of the Few-shot Event Detection model,a Few-shot Event Detection model based on Label Attention Prototypical Network(LAPN)is designed.The model leverages the prior knowledge provided by the event labels to represent each word in the event sequence better.By calculating the embedded representation of instances given different weights under different event label distributions,the feature encoder is continuously updated to obtain a more generalized event prototype representation.Finally,the classification results of the model synthesize the similarity of instance and event prototype representation as well as the weight values of instance and different event labels.Extensive experiments on the Few Event dataset show that the method can effectively improve the generalization ability of new event types.2.Aiming at the lack of the current Chinese Few-shot Event Detection algorithm,and the problem of word mismatch when the English Few-shot Event Detection model is applied to Chinese corpus,a Chinese Few-shot Event Extraction model based on Syntactic Enhanced Projection Network(SEPN)is proposed.In order to integrate the semantic information between characters and words in Chinese,the model uses the character as the granularity to encode the representation and incorporates the syntactic dependencies between words learned by the Graph Convolutional Network into the character representation with contextual information.Finally,the event prototype representation and feature representation are projected into a new embedded space through the mapping matrix to help better classification of the model.Numerical experiments were carried out on the ACE2005 Chinese dataset and compared with a series of existing models to verify the effectiveness of SEPN in the Few-shot Chinese Event Extraction task.3.Aiming at the problem that the general performance of the Event Extraction model is poor and cannot be applied to zero-shot scenarios,a unified Event Extraction algorithm based on the question answering model is explored.The model represents both event types and event arguments as semantically rich natural language query statements.Then,according to the answer to the question,the corresponding label span is obtained.A multihead attention mechanism is designed to capture the interaction information between the query question and the input text.Experimental results on ACE and ERE data sets show that excellent results can be obtained in traditional Event Extraction tasks and Zero-shot learning scenarios without changing the existing model structure,which improves the model’s generality.