Research And Implementation On Trending Event Extraction Method Based On Machine Reading Comprehension

In the information age,there is a lot of meaningless information in trending events.In order to avoid information overload,we can get event information efficiently through Event extraction task.Event extraction is divided into Event detection sub-task and Argument extraction sub-task,which can construct unstructured data into structured data by event extraction task to facilitate follow-up analysis of trending events.The traditional event extraction method relies on entity recognition technology as a pre-processing step or concurrent step,causing the well-known problem of error accumulation.At the same time,these traditional methods classify each argument role independently,ignoring the category similarity between the different argument roles,which leads to the loss of semantic information.In order to solve the above problems,this paper proposes a trending events extraction method based on machine reading comprehension.The main research contents are as follows:(1)In order to solve the problem of error accumulation,this paper transforms the two sub-tasks into machine reading comprehension tasks by building question templates,reducing the dependence of subtasks on entity information.In order to solve the problem of loss of semantic information of argument roles in event argument extraction,the argument role category described as natural language question,which makes the category information of argument role participate in the training.(2)In order to improve the accuracy of sub-tasks model,two techniques are used to introduce their own prior knowledge without changing the frame of machine reading comprehension model.In the event detection subtask,the paper constructs a priori knowledge base by using all the trigger words in the training set,and learning this priori feature by using Bi-GRU.While in the subtask of argument extraction,the regulation layer use conditional regulation layer to incorporate the semantics of trigger word.Finally,the experiment proves that both technical means improve the classification accuracy of the two sub-tasks.(3)Based on the above-mentioned model,a trending event extraction system is implemented with B/S architecture.The system uses Neo4 j database to store the structured data of the event,and D3.js to form a map of events for visual display.At the same time,users can use the system to retrieve and view the key information of events.In conclusion,the experiments on Du EE show that the improved strategy of machine reading comprehension and the introduction of prior knowledge can improve the accuracy of Event Extraction task.At the same time,the system shows the application of the event extraction technology in the trending events.The system can also be applied to the public opinion system to help collect public trending events.