| Text plays a very important role in the data generated by human beings,and how to automatically discover knowledge from text is a classic problem with a long history.With the development of artificial intelligence and data science,machine learning and data mining have become the mainstream of text knowledge mining.However,the existing technologies can only represent a part of the structural features of the data and lack the ability to capture complex text features.Therefore,there is still a gap between the current text mining models and the complicated real-world tasks.This thesis focuses on text mining tasks in complicated scenarios such as heterogeneous,multi-source,dynamic,and multi-level.Compared with the traditional text mining tasks that only focus on the text-word structure,there are complex and hidden interaction structures in the internal data or the external data in these problems.The fusion of various structural features can express the potential relationship in the text more comprehensively,and while can find the knowledge structure that the text does not exhibit significantly.This thesis proposes a novel idea for feature fusion:representing the features as a linear combination of its neighboring features.Based on this idea,this thesis studies the problems such as citation influence modeling,multi-source comparison text mining,multi-layer topic evolution,and topic taxonomy construction.The main contributions of this thesis are as follows:(1)A hierarchical topic evolution model is designed,which extends the traditional singlelayer topic evolution analysis to a multi-layer structure evolution analysis.This research considers the hierarchical topic structure as a tree-shaped feature relationship.The model synthesizes the lower topics into the upper topics by learning the feature fusion factor between the upper and lower layers and generates a multi-level topic structure from bottom to top.Compared with the traditional hierarchical topic models,the model can determine the weight between the child topic and the parent topic,which greatly promotes the accuracy of the structural evolution analysis.The model can effectively observe the details of the evolution,and can discover the interaction between the topics and the evolution mode of a discipline.(2)The learning method for topic taxonomy based on a deep structure is proposed,and the topic tree with a cross structure is realized,that is,a subtopic can belong to multiple parent topics.The cross structure is more in line with the organization of actual knowledge.This method discovers topics with different abstract levels and the cross-subordination relationship between topics based on the transforming of topic feature spaces with different sizes.With the help of a deep learning structure,the model selects appropriate lower-level topics to form upperlevel topics using two regular terms and induces the fusion between child topics under the same parent topics.This algorithm architecture can automatically determine the width of the topic tree and improve the robustness of the model.(3)In order to solve the problem of multi-source text comparative mining,a comparative text mining model based on matrix factorization is developed.This research is based on a twostage algorithm scheme that can distinguish the same and different topics from two text datasets,and automatically determine the number of the same and different topics.The model recognizes the feature relationship in different datasets as related features and irrelevant features.For related features,the model forces them to close to each other,and for irrelevant features,Hilbert-Schmidt independence criterion is introduced to maintain their independence.This model can not only discover the similarities and differences of knowledge from different datasets,but also effectively eliminate the interference of incompatible features and improve the complementarity of the fusion of compatible features.(4)This research introduces the citation influence into the learning model for the citation network,and dynamically encodes the citation relationship,which replaces the binary structure of the original citation network.The citation influence improves the performance of the model in tasks such as text clustering and influence prediction.This research integrates the citation influence matrix into the matrix factorization framework and learns the weight of citation influence while learning word features and document features.The model combines document features and citation features with the weight provided by citations,and represents the document as a combination of citation features and its own features.At the same time,the model introduces the concept of citation sparsity,finding high-value core citations,and suppressing or deleting low-value citations.The above four models dynamically deconstruct the different forms of text structure into the different feature relationships,which not only improves the accuracy and rationality of feature fusion,but also discovers the rich and valuable knowledge hidden in the text.The method proposed in this paper has shown good performance in the experiments,which proves that the fusion of multi-modal text features is an effective way to solve complicated text mining problems.The research in this paper is helpful to the theoretical development of feature fusion and text mining and provides a new way for text data modeling in complex scenes. |
