Research On Netnews Event Extraction Based On Deep Learning And Embedded Feature Space

In the era of Internet with vast information,how to obtain valuable information that people need from a large amount of fragmented text data has become a huge challenge.Event extraction technology can effectively mine and refine text information so that users can get what they need quickly and accurately,then it has gradually become a research hotspot in the field of natural language processing.However,text data has high dimensionality and high ambiguity.Traditional methods based on feature engineering divide event extraction into two tasks: trigger words recognition and event arguments semantic role classification.In this process,a large number of features need to be analyzed and selected.The over-reliance on natural language processing tools leads to the failure of large-scale application of the model,the cascade propagation of errors,the complexity of feature engineering and other shortcomings.Therefore,this paper focuses on embedded feature representation and multi-event extraction,based on pattern recognition and deep learning.From the perspective of transfer learning,the research trend of event extraction is combined.In this paper,the embedded feature representation and multi-event extraction is studied.The high-level information hidden in text(semantic features,syntactic features,etc.)is deeply mined,and the performance of the event extraction model is improved ultimately.This paper mainly contains the following contents.For the complicated feature engineering problem and analyzing the reasons for the completion of the event extraction task in stages,the paper proposes a solution to introduce the trigger word embedding feature into the event extraction model.In order to obtain the embedded feature representation of the trigger word,the paper models the trigger word recognition task as the sequence labeling problem,and proposes a trigger word recognition method based on the Bi-directional long and short time memory network.Firstly,considering the different effects of various features on the trigger words recognition task,four types of features are introduced: word embedding representation,entity type embedding,dependency embedding and part of speech embedding.Then,under the framework of the Bi-LSTM,the semantic information representing the global context is effectively extracted.Finally,the event trigger words are classified according to local word information and context semantic information.Experiments show that the proposed method can combine local and global information to effectively improve the recognition effect of event trigger words.For problem of the error cascade communication,the regularization migration method of embedded features is studied.This paper proposes an event extraction model based on joint embedding features.By migrating the trigger word embedding representation and combining location features,the internal connection hidden between the event element and the event trigger word is deeply explored with the rich information of the joint embedding feature.For multi-event extraction,this paper maps the trigger word to the event description one by one.By classifying the event element semantic role classification task into the relationship classification problem of(trigger word,event element),the problem can be effectively improved.Experiments show that the proposed model can effectively migrate existing trigger word information and improve the effect of event extraction.Finally,this paper evaluates the proposed method in two typical event annotation datasets of ACE and CEC datasets,and compares it with popular methods to prove the effectiveness of the proposed method.More deeply,the benefits of multiple embedded features in the model comparison model indicate the superiority of the joint embedding feature.