Research And Application On Open Relation Indicated By Predicates Extraction

On account of the characteristics of open relation indicated by predicates extraction task,this thesis proposes two phrases method:Phrase One is recognition of open relation indicated by predicates,the goal of this phrase is judging whether a given sentence contains relations indicated by predicates;and Phrase Two is relation extraction,and the goal of this phrase is extracting relation triples from sentences that contain relations indicated by predicates.The detail is as follows.First,this thesis proposes employing dynamic pooling frame convolution neural network to recognize open relation indicated by predicates.The model uses word embedding as input,and can effectively alleviate the problem of sparsity;also the model employs dynamic pooling frame convolution,and can learn and excavate a rich set of structure information.The experimental results show that the proposed method gains a better performance than both classical shallow classifiers and tree kernel methods.Further on,this thesis discusses several types of features,and finds out that Part-Of-Speech Tagging and position features can improve the recognition performance obviously.Second,this thesis designs a rule generating method suitable for open relation indicated by predicates extraction task,and obtains totally 7 rules using this method.To settle the problem of ambiguity,which is caused by that one sentence may be matched by more than one rule,this thesis propose a priority based disambiguation strategy.The experimental results show that the proposed methods can get a satisfactory extraction performance.Finally,this thesis achieves a complete open relation indicated by predicates extraction module by cascading recognition and extraction methods referred before,and introduces this module into two applications:one is open relation indicated by predicates extraction system,the aim of this application is extracting all relations indicated by predicates from texts provided by users;and another is Trend Search system,the system takes a person name as input,search from web other persons and relations between other persons and the query person,and return the search relation results.