Research On Dynamic Topic Modeling And Evolution Prediction Based On Graph Structure

Text analysis is crucial for extracting valuable information from large volumes of textual data,finding wide applications in decision-making,knowledge discovery,and social media analysis,among others.Topic modeling serves as a central technology,especially vital in revealing emerging themes to support decision-makers.However,despite static topic modeling’s ability to discover latent themes in textual data,it often struggles to capture temporal characteristics of topics when dealing with dynamic text data.Moreover,existing dynamic topic modeling techniques,such as word-topic probability distribution methods based on LDA-like models and document-topic similarity methods based on document clustering,have failed to fully consider high-order semantic relationships among keywords within a topic,making it challenging to accurately grasp the thread of topic evolution in complex scenarios.Likewise,current graph-based studies often overlook the semantic and temporal properties of topics.Therefore,thesis mainly addresses these issues,focusing on graph-structured dynamic topic modeling and its evolution prediction,making the following major contributions:(1)In response to the fact that existing topic modeling methods do not effectively model the relationship between topics and keywords,and lack consideration of dynamic evolution prediction,we propose a new method for dynamic topic modeling and evolution prediction from the perspective of graph structure.The core idea is to model a text corpus as a dynamic keyword-keyword network.G-DTM aims to capture deep theme information by learning high-order keyword association information through node representation learning and time series neural network learning.Furthermore,we have collected and collated a brand-new dataset to aid scientific research.Experiments show that G-DTM not only can extract evolving themes that correspond to the temporal development of content disseminated during a large-scale epidemic,achieving significant improvement in topic consistency,but also demonstrates a strong ability to predict emerging themes in the future.(2)In view of the issue that the node representation capability in existing graph structure research is inadequate,resulting in weak topic consistency,we propose a dynamic topic modeling method based on the Attribute Graph Neural Network(Attr GTM).This model enriches the feature representation of nodes and strengthens the connection intensity and association in the graph structure context by considering semantic information like word vectors as node attributes.The model also applies Laplacian smoothing to the attributes to combat noise.Finally,we use a variable-length time series neural network to fully utilize temporal information and topological structure information,thereby improving the quality and effectiveness of topic modeling.Experiments demonstrate that Attr GTM has better topic extraction capabilities and obtains a higher consistency score.