Research On Person Relationship Analysis Method Based On Network Media

Character relationship analysis mainly studies how to effectively and automatically extract structured character relationships from massive data,which plays an important role in building character knowledge maps,character search engines and other systems.Compared with structured text,more unstructured natural language text is stored in network media.Therefore,how to extract relevant attributes of people from unstructured text,and then analyze the relationship between people has become a hot research issue.At present,the analysis methods of person-based relationships based on network media face the following challenges: extracting person attributes requires a large amount of labeled data as a support,and the lack of person-related labeled data is currently a major problem;on the other hand,the use of attention mechanisms in deep learning can enable the performance of relation extraction has been further improved,but existing attention mechanisms usually focus on low-level semantic features such as sentence level,and lack attention to the global semantic information of the entire relation.Aiming at the above challenges,this thesis proposes research on unstructured texts related to characters in network media based on distant supervised dataset augmentation and denoising methods and multi-level attention mechanism methods.Main tasks as follows:(1)Aiming at the problem of lack of labeling data of entity relationship in the person domain,this thesis proposes a method for automatically augmenting a large number of labeling data sets by using Web tables to expand the entity pairs of the existing knowledge base and using distant supervision.This method uses the tables in Wikipedia to expand the data set related to the characters in the NYT training set,so that the expression of the corresponding relationship becomes richer and more diverse.Aiming at the noise problem that traditional distant supervision methods may bring,this thesis proposes a data set denoising method based on relational semantics.The semantic information expressed by the sentence is used to determine whether the relationship label corresponding to the sentence is correct or not,which reduces the problem of incorrect labeling that the distant supervision may introduce into the data set.Finally,it is experimentally shown that the data set denoising method proposed in this thesis can effectively filter the noise sentences of the distant supervised data set,improve the overall quantity and quality of the data set,and improve the performance of the relation extraction algorithm.(2)In order to avoid complex feature construction in traditional relation extraction methods,this thesis uses deep learning methods for relation extraction.Aiming at the problems that existing deep learning relation extraction models cannot deal with the noise in sentences and the attention mechanism only pays attention to the features of sentence level,this thesis proposes a relation extraction model based on multi-level attention mechanism.The model is first improved on the input layer,and a subtree analysis method of sentences is proposed so that the input only retains the words that describe the key relationships.At the same time,a bag-level attention mechanism is proposed,and a sentence-level attention mechanism is integrated to form a multi-level attention network for relationship prediction.Experiments show that the model can achieve good results in relation extraction.Through comparative experiments,it is found that using subtree analysis to remove intra-sentence noise and multi-level attention mechanism to focus on global semantic information can improve the accuracy of relation extraction.