Research And Application Of Event Extraction Algorithm For Unstructured Text

The essence of event extraction is to extract complete events,including event trigger words and a series of event elements,from a large amount of unstructured and unprocessed information,and store and display them in a structured form.Event extraction technology provides technical support for situation awareness,social governance,decision assistance and emergency response in contemporary society,and has important research significance and value.At present,there are some problems in the field of event extraction such as poor extraction effect for complex semantic environment.Therefore,thesis focuses on the event extraction task and carries out the following main studies:1.To solve the problems of inaccurate extraction,positioning and classification of event-triggering words,existing methods mainly use sentence-level context information,ignoring document semantics and document subject information.Therefore,thesis proposes an event trigger word extraction method based on attention mechanism and BI-GRU.Context semantic information is mined by BI-GRU,word-level attention is used to pay attention to event trigger words in sentences,sentence-level attention is used to pay attention to sentences containing trigger words,and document semantics are obtained by integrating the results of attention.At the same time,the document topic information is obtained based on the encoding of the key structure sentences.Then,another BI-GRU is used to extract event trigger words combining the obtained semantic information and topic information.Experiments on ACE2005 data set verify the effectiveness of the proposed model in trigger word extraction.2.Due to the problems of role overlap and low classification accuracy in event element extraction,the existing methods are not effective in obtaining the deep semantic information of event element context.Therefore,thesis proposes an event element extraction method based on BERT and attention mechanism.This method combines the BERT pre-training model and multiple semantic information to obtain rich word representation,obtains event feature representation from extracted event trigger words and candidate event elements,and then obtains dynamic sentence vector representation based on attention mechanism,combines event feature and sentence representation to extract event elements.Experiments on the same data set demonstrate the effectiveness of the proposed model in element extraction.3.Completed the event extraction system for the news field.At present,there are new requirements for the intelligentization of social governance.In combination with the needs of national development strategy and based on the proposed event trigger word and event element extraction model,thesis designs and implements an event extraction system oriented to the news field,which provides theoretical and application support for social governance.