Kernel-based Entity Relation Extraction

Entity relation extraction is an important topic in the field of Information extraction,which purpose is to take out the binary relation from unstructured free text defined in advance for further analysis on the basis of named entity extraction.In recent years it has become an essential part of natural language understanding.Kernel-based entity relation extraction makes full use of planar features and structured features such as parsing tree.The optimal solution can be achieved by convex optimization.Kernel-based entity relation extraction using a supervised learning method,which uses kernel trick mapping the original input space to the dual feature space,in order to solve the nonlinear separable problem.In high dimensional feature space we only need to compute samples' inner product(namely kernel function,referred to as ”kernel”),instead of working out the specific function value.This makes the implicit feature mapping to replace the dominant feature mapping,providing a new solution for those machine learning methods based on feature functions.At the same time,kernel-based problem is a quadratic convex optimization problem,avoiding the local optimal solutions,and will not lead to over-fitting.Kernel-based entity relation extraction methods include singular kernel learning,composite kernel learning and multiple kernel learning with respect to kernel type.Meanwhile,because the ideal space is dual to kernel space,VCA method based on vanishing ideal analysis is put forward.However,up till now,the kernel functions for texts and their composition have not been studied thoroughly,multiple kernel method has not been applied to entity relation extraction,and VCA has not been studied deeply.This thesis are summarized as:(1)With respect to kernel functions and their composition methods,entity kernel function and syntactic structure string kernel function are put forward.They are combined with shallow tree kernel function in linear and polynomial,and better performance is achieved for entity relation extraction.(2)With respect to multiple kernel learning method,a multiple kernel classification method based on clustering is suggested for the first time.This method uses target function in clustering,and a new viewpoint for classification is put forward,which can reach optimal solution in linear time.Better performance is achieved for entity relation extraction.(3)With respect to VCA,a new decision function is suggested,which is based on the ranking of vanishing components instead of their values.A grouping strategy for training set is put forward for the first time,and the vanishing component analysis is done on individual groups instead of on the whole training set.Then,the vanishing components from different groups are combined together for further testing.The experimental results on various data set indicate that the new decision function and grouping strategy can improve the performance of GVCA greatly.Meanwhile,the experimental results show that GVCA can converge rapidly compared to other classification methods and it can be used in entity relation extraction.