Research And Application Of Event Extraction Technology Based On Deep Learning

Event extraction(EE)is a relatively complex subject in information extraction.In an actual industrial management software,event extraction has an extremely high value in handling customer complaints and can help companies achieve rapid product update iterations and maintain customer relationship.Event extraction is not only a current and challenging research hotspot in natural language processing but also plays an essential role in knowledge graph-related tasks.Event extraction aims to process unstructured natural language texts in a structured manner and automatically extract and store the event information of customer requests.At present,the pre-training model represented by BERT can handle multiple NLP tasks well,but the effect in event extraction-related tasks needs to be further improved.A significant challenge for researchers is how to improve the performance of the event extraction model,break the limitations of independent text processing,and improve the effect of open domain event extraction.This thesis studies the event extraction task to improve the pre-training model and apply it to the actual scene.The main work is as follows:(1)This thesis uses Ro BERTa-wwm as the baseline model and improves the trigger extraction task and argument extraction task.In the trigger extraction model,annotated data is used to build a trigger knowledge base,and in the trigger extraction task,the words in the sentences are matched with the trigger words in the knowledge base.The matching results are used as additional text features for model training to improve the effect of trigger extraction.Conditional Layer Normalization is used in argument extraction to control the generation behavior of the transformer.The relative distance from each word to the trigger is used as a feature,and the trigger is used as the conditional input of the Layer Normalization so that the entire sentence is integrated into the semantic information of the trigger,thereby improving the effect of argument extraction.(2)In view of the lack of data sets in the field of event extraction,this thesis uses two public open-domain data sets in the experiment.The self-made data set from corporate complaints is used in an actual application scenario.Besides,we adopt the method of synonym replacement for data augmentation.In this thesis,we verify and evaluate the improvement of the model and the effectiveness of data augmentation through comparative experiments.(3)The improved model adopts the PGD adversarial training algorithm and takes the stochastic weight averaging(SWA)optimization method in the training smoothing stage to improve the robustness and generalization performance of the model,which is confirmed by comparative experiments.(4)We migrate the improved model from the open domain to a specific domain,conduct comparative experiments before and after model migration,and analyze the effect of model migration.(5)In this thesis,the event extraction technology is applied to a vehicle accessories enterprise's customer relationship management platform.The structured event information is extracted from the unstructured customer complaint data.In this way,event information forms can be filled out automatically,and events information can be displayed in the form of a knowledge graph.Through this application,the service department can efficiently handle the customer complaints.The R?D department can make timely technical improvements and product updates based on the feedback to realize the application value of this thesis.