Research On Event Reasoning And Prediction Of Scientific And Technological Literature Based On Hypergraph Neural Network

With the increase of the amount of scientific and technological literature,the research work of scientific research hotspot mining becomes more and more difficult.Existing scientific research hotspot mining methods are mainly based on topic model or keywords,but these methods can not express the deep meaning of scientific research hotspot,it is difficult to achieve deep semantic coreference resolution for the events between different scientific and technological literature,and loss of higher-order correlation in event relationship prediction,taking scientific and technological literature as the object,research on cross-document event coreference resolution and event relationship prediction.The main work of this thesis is as follows:(1)Cross-document event coreference resolution based on hypergraph convolutional neural networkMost of the existing methods are serial processing to analyze event data,which do not take into account the complex relationship between events,and the deep semantics of events are difficult to be mined.To address this problem,this thesis proposes a cross-document event coreference resolution method based on hypergraph convolutional neural network.First,events extracted from several scientific and technological literatures were labeled,to determine the trigger words of the events,and the multi-document event hypergraph was constructed around the trigger words.Then,hypergraph convolutional neural network is used to learn the higher-order semantic information in the multi-document event hypergraph,while a multi-headed attention mechanism is introduced to learn the hidden features of different event relationship types.Finally,feedforward neural network and average-linkage clustering method are used to calculate the coreference score of events and complete coreference event clustering.The experiment shows show that the cross-document event coreference resolution method in this thesis is better than the contrast model.(2)Event relationship prediction based on time-series hypergraph convolutional neural networkIn order to solve the problem of losing original event higher-order interaction and relationship information caused by modelling complex multi-relational time-series events using traditional graph structures,this thesis proposes an event relationship prediction method based on time-series hypergraph convolutional neural network.The method extends the relation-aware graph convolutional neural network to the hypergraph and hyperbolic space for capturing structural dependencies in the hyperbolic hypergraph of time-series events at each moment,and introduces hyperbolic gate recurrent unit to learn the time-series features in the hyperbolic hypergraph of time-series events.In addition,the static attributes hyperbolic hypergraph constraint layer to merge the static attributes of entities in events with entity representations.Finally,the event relationship prediction is completed based on the entity and relationship representations of the events.The experiment shows that the proposed method has good effect on event relationship prediction task.(3)Scientific research hotspot mining in scientific and technological literatureThis thesis abstracts the scientific research hotspot into events,to realize scientific research hotspot mining of scientific and technological literature,display the development trend of scientific research hotspot,help researchers gain insight into scientific research trends,track emerging fields,and provide reference for scientific research topic.