Studies On The Approaches To Event Mining In News Domain

Nowadays,various online news media publish a large number of hot events in real time,from the daily life of the people to the national and international current affairs,which has become an important source of information for people.However,although it has become easier for people to get information,the information received is often fragmented,incomplete,partially redundant and noisy.This dilemma makes it difficult to have a clear understanding of the key events in a hot topic and their evolution patterns from the overall point of view.Hence,it is necessary to more effectively utilize and organize event-centric domain data.In the face of massive,redundant and mixed news streams,how to"perceive" the occurrence of events in time,effectively "analyze" and dig out the ins and outs of events and potential story development patterns,and further predict the future trend of events for "decision-making" has become a research problem of important practical significance in news domain.By solving this problem,it can help governments and individuals to fully understand the occurrence and development of events of interest,and provide an important basis for applications such as public opinion monitoring,emergency response,and auxiliary decision-making.In view of the above perspectives of "perception","analysis" and "decisionmaking" related to events,this thesis has studied and explored the event mining techniques in news domain.It focuses on the three tasks of event detection,event evolution and event prediction,and is aimed to solve their current challenges and limitations,and comprehensively improve their performance by proposing innovative and effective approaches.The research achievements are expected to provide original technical support for the actual business of important national departments and the daily life of people.The main contributions of this thesis are summarized as follows:(1)In response to the problem that the existing event detection methods in news domain cannot fully model the sequential-semantic correlation between temporal information and news content information,an event detection algorithm in news domain based on temporal characteristics(TCED)is proposed.First,Jaccard Similarity coefficient× Inverse Dimension Frequency with time order is utilized to represent the document embeddings,which effectively fuses the temporal information and content information in the feature selection stage,and alleviates the information loss of previous bag-of-word models and the neglect of temporal and semantic information.Second,BiK-means is employed to cluster all related news documents into news events,i.e.,the event detection task is completed.In addition,in response to the problem that the existing event detection methods in news domain cannot fully model the structural-semantic correlation information of documents,an event detection algorithm based on concept interaction graph and graph neural network(CGED)is proposed.A ConceptGraph of news document is first constructed by the co-occurrence relationship of keywords in the sentence,then followed by a Siamese Graph Convolutional Network for document representation learning.This algorithm fully takes structural and semantic information of a document into account,and achieves a more accurate document representation.Experimental results demonstrate the effectiveness of the proposed TCED and CGED algorithms.(2)In response to the problem that the existing event evolution methods in news domain often ignore the interdependence relationship between event detection and event evolution,which causes a large performance drop,a joint event detection and evolution algorithm in news domain based on multi-task joint learning(JLEE)is proposed.The attention mechanism is first introduced into the bidirectional GRU network to construct a shared semantic learning model of news documents and events.Then two continuous similarity metrics are learned on the basis of a Siamese network framework by using neural stacked strategy,so as to judge whether two news documents are related to the same event or two events are related to the same story.Experimental results show that the proposed JLEE algorithm can improve the accuracy of event detection and event evolution simultaneously.(3)In response to the lack of uniformly modeling multi-grained complex contextual event representations in event prediction,which results in an inaccurate prediction,an event prediction algorithm in news domain based on multilevel semantic fusion strategy(MSFEP)is proposed.The enhanced multilevel(event/chain/segment level)script learning is first developed to effectively model the both temporal and casual information as well as the rich structural relevance.Then a dual fusion strategy based on feature-level fusion and scorelevel fusion is utilized to fully integrate the complementary advantages of multilevel semantics by nonlinear feature composition and weighted score fusion.Experimental results show the proposed MSFEP algorithm can effectively predict the future events and present a good robustness.(4)In response to the problem that the global matching is difficult to achieve for event prediction in news domain,resulting in the prediction results being prone to locally optimal solutions,an event prediction algorithm in news domain based on hierarchical neural network(HNNEP)is proposed.A bottomup four-layer stacked neural network is built to effectively model the event encoding layer,event relation layer,event matching layer,and event prediction layer.The event encoding layer first employs stacked transformer networks to model fine-grained connections between event arguments.Then the event relation layer adopts a gated graph neural network to model the rich event evolution patterns.Through the effective learning of the event coding layer and the event relation layer,an event representation with rich context information is obtained.Given the candidate events,the event matching layer further integrates the feature-level local matching and global matching to aggregate comprehensive event contextual information to assist in prediction.Finally,the event prediction layer calculates the relatedness score between the context event chain and candidate events through a learnable classification network.The proposed algorithm fully models the multi-level event interactions,and makes decisions based on global contextual information,which effectively solves the problem of previous score-level matching-based methods being prone to locally optimal solutions.Experimental results show that the proposed HNNEP algorithm is significantly superior to other baseline methods.