Chinese News Event Extraction For Military Aided Decision Making

Chinese military news has the characteristics of timeliness and large scale.It contains a large amount of valuable military information,which is of great significance to military decision-making.In particular,the information of five dimensions,such as the type of military events,combat objectives,combat tasks,time and space,has high application value for military decision-making.Compared with the text description in Chinese military news,the five-dimensional information of military events is highly abstract and interdependent,and the text description is often incomplete.In addition,there is a large amount of overlap of event information in military news reports,which brings challenges to the extraction of Chinese military news events.In view of the above challenges,this thesis studies the technical scheme of Chinese news event extraction for military aided decision-making,in order to support the development of an efficient military information extraction system,and to design the technical method of extracting the above five-dimensional information from Chinese military news data.The focus is to build a five-dimensional information extraction model of military events that can adapt to the overlap of event information in news sentences.The main work and achievements of this thesis include:1.A two-stage Chinese military news event extraction scheme is designed.Combined with the characteristics of Chinese military news event description and five-dimensional information,the scheme first extracts the event trigger words and five event arguments such as event subject,event object,weapons and equipment,event time and event area,so as to reduce the difficulty of directly extracting five-dimensional information and data annotation.Secondly,an information splicing method is designed on the event trigger words and five event arguments,It realizes the construction from event trigger words and five event arguments to five-dimensional information.2.Data set construction for Chinese military news event extraction model learning.The annotation specification of the data set is designed.On this basis,a Chinese military news data set is constructed by labeling the event trigger words and five event arguments in the collected data.3.In view of the large amount of overlapping event information faced in the stage of extracting event trigger words and five event arguments,this thesis proposes a joint learning model based on hierarchical decoder,which can be used for overlapping event extraction.It is composed of Bert based context semantic encoder,event type detection decoder,event trigger word extraction decoder and event argument extraction decoder,and adopts the event extraction method based on pipeline,Joint learning achieves global optimization.Aiming at the stage of transforming event trigger words and five event arguments into five-dimensional information,a rule template is designed to optimize the event extraction results,and realize the construction of event trigger words and arguments extracted from the event extraction model to five-dimensional information.The experimental tests on ace2005 and independently constructed Chinese military news data sets verify that the joint learning model based on hierarchical decoder proposed in this thesis for overlapping event extraction is efficient and robust in dealing with overlapping event extraction tasks,and the proposed technical scheme of Chinese news event extraction for military aided decision-making has application reference value.