Research On Transition-based Kazakh Parsing With Neural Network

Parsing as an important part of natural language processing,is necessary to carry out research work on Kazakh parsing.Parsing is generally used graph-based method and transition-based method.This paper uses transition-based method to study the Kazakh sentence analysis.The transition-based method needs to use the buffer to store sentence information,and use the stack area to store the analyzed sentences.The information is then generated according to the transfer method and stored in the historical action area.The transition-based method is generally contain bottom-up and top-down method.The essence of the method is to input the syntactic tree to be analyzed and convert it into a corresponding action sequence according to certain rules.Then the parsing problem it also became a serialization problem how to predict the best action sequence.The top-down approach begins with the analysis of the top of the sentence.First,the non-terminal character is first opened.In the continuous analysis of the sentence,the word is moved into the stack,and after satisfying the conditions for forming the subtree,the current non-terminal character is closed and formed.Subtree,if the open nonterminal is wrong,not only the current subtree will be wrong,but also the formation of other subtrees.In the bottom-up method analysis process,the terminal is first moved to the stack.When the condition of the subtree is satisfied,the specification is made.At this time,the non-terminal is predicted,but the bottom-up method still has insufficient in capturing the whole sentence..In response to these shortcomings,this paper has done the following three aspects in the analysis of Kazakh parsing:1.In the current Kazakh parsing,the transition-based method generally adopts a bottom-up or top-down approach.This paper adopts a in order transition-based method on the traversal syntax tree,which can compensate for the degree to some extent.The lack of bottom-up and top-down methods.2.In this paper,the transfer-based method is combined with the neural network,and a syntactic parser composed of three LSTMs is used for parsing.The three LSTMs correspond to the stack area,the buffer,and the historical action area used in the transfer method.When performing the REDUCE action,the Bi-LSTM is used to extract the time-characteristics of the subtree,and the three LSTMs are passed to a softmax layer to obtain the next one.The probability of action.3.In this paper,a reordering method combining discriminative model and generative model is adopted.Firstly,a certain number of candidate trees are generated by the discriminant model through the sampling algorithm,and then the candidate tree is re-scored by the generative model,and the highest score is selected from the new scores.The tree is output as the best result.