Research On Event Storyline Generation And Evolution Inference Based On Event Evolutionary Graph

In recent years,with the rapid development of information technology,there emerge more and more kinds of text information on the Internet.When searching for an event,people are often overwhelmed by massive texts,and unable to obtain effective information from them to grasp the context of event evolving.In general,readers want to rapidly understand the development process of the event and the subsequent trend.This requires mining valuable information from the text,and then using these information to inference the event evolving trend.Therefore,how to extract information from large number of news texts and analyze the development trend of events has become a key issue in the field of natural language processing.To solve the aforementioned problems,this thesis proposed some approaches for information extraction and event inference tasks.First,although the traditional storyline generation method can capture the main information of the event,most of them adopt the abstract extracted from the news text as the node of the storyline.These kind of storylines are less readable and are not accurate enough.The storyline construction method based on knowledge graph proposed in this thesis takes the triplet of events as the node of the storyline.Compared with the summary-based storyline,taking the triplet as the node is more readable,during the information extraction process.The location entity is extracted and used to guide the construction of local story lines,and its accuracy is higher.At the same time,with the development of various pre-trained language models,some disadvantages of traditional static word vectors have been improved,such as the problem of polysemy.In order to obtain dynamic word vectors with semantic information,the pre-trained model BERT is used to generate better word vectors.In the event inference model,in order to get a better event representation,this thesis proposes an event representation layer model based on the self-attention mechanism,which can not only get the representation of the event itself,but also obtain event segments information.The combination of individual event and event segment makes the model can capture more event information.Finally,to test our proposed storyline model based on knowledge graph and event inference model based on event evolutionary graph,this thesis conducted several experiments on typhoon news data,and the experimental results demonstrate the effectiveness of the proposed models.At the same time,the comparison experiments of different modules also verify the effectiveness of each module in the proposed models.