News Event Evolution Analysis Based On Event Graph

With the vigorous development of network media and its technologies,online news has become an important part of Big Data on the Internet.At the same time,due to the timeliness and one-sidedness of news reports,the problem of news fragmentation has become increasingly prominent,and it is difficult for people to grasp the development of events from the complicated news feed.Event evolution analysis technologies can be applied to solve the problem,which aims to mine the evolution relationship between events and track the evolution process of events.However,existing event evolution analysis methods still have the following shortcomings: Firstly,the existing event feature modeling technologies do not model and utilize interactions between event elements.Secondly,traditional evolution graph construction strategies are prone to generate highly complex structure and do not model the multi-topic evolution relationships between events.Thirdly,existing methods generally assume that each event belongs to a single topic rather than multiple topics,which may make evolution processes incomplete and incoherent,thus affects the detection and tracking of new topics.To solve the problems above,this thesis proposes a news event evolution analysis method based on event graph.The main research work is as follows:(1)To enrich the feature representation of news events,a news event representation learning method based on semantic and syntactic information is proposed.Firstly,the sentence level event triplets are obtained from the news text based on the syntax rules,and the news keywords are extracted as the event core words.Then,the event embedding features that capture semantic and syntactic information are obtained by training the event core word prediction and the missing element prediction model.(2)To reduce the complexity of the evolution graph and model the multi-topic evolution relationship,a method for constructing Multi-Topic News Event Evolution Graph(MTNEEG)is proposed.Instead of performing topic clustering in advance,the method directly calculates the evolution strength between the target event and its related old events to retain the multi-topic evolution relationships.Furthermore,based on the rationality constraint of evolution pattern,non-redundant evolution relationships are constructed to simplify the structure of evolution graph.(3)To effectively detect and track the evolution process of multi-topic news events,a MTNEEG-based evolution process detection and tracking algorithm is proposed.Firstly,the detection and tracking tasks are transformed into the boundary event recognition problem in MTNEEG.Then,the multi-topic evolution relationships and coherence measurement mechanism are utilized to improve the integrity and coherence of evolution processes.Finally,MTNEEG is divided into multiple overlapping evolution processes.(4)Prototype system implementation and experimental analysis.Under the dual-structure network,we design and implement a prototype system of news event evolution analysis based on event graph.Then the experimental analysis of the proposed methods is carried out.Experiments show that the proposed news event representation learning method can effectively improve the performance of downstream tasks;the proposed evolution graph construction method can effectively reduce the complexity of event evolution graph and model multi-topic evolution relationship;the MTNEEG-based evolution process detection and tracking algorithm outperforms baseline methods significantly.