Research On Event Extraction Methods Based On The Machine Reading Comprehension Framework

With the vigorous development of the Internet,a large amount of unstructured information has been stored in cyberspace.In order to retrieve and utilize the growing unstructured text information more efficiently,event extraction in the field of natural language processing has become a hot research direction.Event extraction is mainly used to automatically extract the elements that can describe events from unstructured text and display them in a concise and refined structured way.In recent years,deep learning methods has become the mainstream method of event extraction because of its strong learning ability and high portability.Most of the deep learning methods use the form of classification task and sequence labeling task to extract events.Because there are nesting problems and error propagation problems,there is more works to use the form of machine reading comprehension task.The method based on machine reading comprehension not only introduces a priori knowledge to strengthen the text representation,but also can solve the nesting problem and error propagation problem.However,the current methods based on reading comprehension need to enumerate tag space to construct samples,and do not make use of the characteristics of the task field in the current application.Specifically,in the event detection task,the method based on machine reading comprehension independently determines whether the event type is triggered or not,ignoring the correlation between events.In the element extraction task,the method based on machine reading comprehension does not make use of the entity characteristics of the event elements,which is not conducive to the boundary judgment when the model extracts the element fragments.Therefore,this thesis mainly carries out the following research to solve the problems:Firstly,this thesis proposes an event detection method based on fusing event relationship,which can detect all event types at one time while fully contacting the event relationship.Firstly,the model integrates all label information through bi-directional attention mechanism.Then the relationship between different events is combined for event detection by introducing self attention mechanism and a priori knowledge related to event types.When introducing the self attention mechanism to model the event relationship,we use the additive attention mechanism and matrix interaction mode of element product to obtain the globally perceived event global vector.When introducing a priori knowledge to model the event relationship,we integrate the conditional probability of two events into the model as a priori knowledge.Finally,the comparative experiment on ace2005 data set shows that the event detection method of fuse event relationship can improve the performance of the model.Secondly,this thesis proposes a two-stage event subject extraction method.In the first stage,the event type is judged through the multi label classification task,and the useless event types are filtered for the event subject extraction in the next stage.In the second stage,the method based on reading comprehension is used to extract the event subject spans combined with the event type information.In the first stage,multi label classification enhances text representation by fusing label information through attention mechanism.In the second stage,considering the entity characteristics of the event subject,the boundary of the event subject spans is better distinguished by introducing additional entity information.Finally,experiments on ccks2020 event subject extraction data set show the effectiveness of this method.