| The serial verbs structure is a typical sentence structure that has received much attention in modern Chinese,and it is also a hot and difficult problem in modern Chinese research.Serial verbs sentences composed of serial verbs structures have become the focus of modern Chinese grammar research due to the complexity of their meaning and the particularity of their structure.The serial verb sentence with multiple predicates contains a wealth of knowledge.Serial verbs sentences represent multiple events,and these events are interdependent that produce interrelated events,and they present a semantic relationship such as way,succession,purpose,and cause–and-effect.Identifying serial verb sentences correctly and analyzing the semantic relationship of consecutive verb sentences are the basic tasks of the natural language processing,which can provide powerful support for upper-level applications such as dependency syntax analysis,Chinese abstract semantic representation analysis,and machine translation.Therefore,based on the artificially annotated corpus,this thesis uses neural network methods to carry out experimental research on the recognition of Chinese serial verbs and the recognition of the semantic relations of serial verbs.The research content are mainly in the following:The construction of the data set for the recognition of serial verbs and the recognition of semantic relations of serial verbs.The main work of this part is based on the theoretical knowledge and task requirements of the serial verbs to formulate the labeling specifications,manually label the content of the primary school Chinese textbooks from the first to the sixth grade and the content of the Tsinghua tree bank,and then construct the data set required by this thesis.This thesis proposes an automatic recognition method for serial verbs based on neural network.In order to alleviate the huge gap in the quantity between serial verbs and non-serial verbs,the real corpus is firstly preprocessed according to the simple rules,and then encoded using BERT(Bidirectional Encoder Representation from Transformers).This work uses multi-layer CNN(Convolutional Neural Network)and Bi LSTM The(Bidirectional Long Short-Term Memory)model jointly to extract features for classification,and then completes the task of recognizing serial verbs.The experimental results verify the effectiveness of the method.This thesis proposes a method of semantic recognition of serial verb sentences based on neural network.This method divides the semantic recognition of serial verbs into two steps.The first step is to recognize the serial verbs and their subjects in the serial verbs.Inspired by named entity recognition,this work uses the Whole Word Masking technology instead of BERT’s MASK method.Applies Bi LSTM-CRF to predict the next possible pos tag through the words that appear before,and finally divide the components of the serial verb according to the serialized tag obtained.The second step is to identify the semantic relationship between the serial verbs of the known serial verbs.This work uses BERT to encode the serial verbs and the serial verbs sentences,obtains the context information of the two through Bi LSTM,and obtains the semantics between the serial verbs and the serial verbs sentences through Attention Interaction,so as to achieve the purpose of semantic recognition.At the same time,the models mentioned above are connected in series as a pipeline model to realize the end-to-end solution to the recognition of serial verbs sentences and the recognition of semantic relations of serial verbs.This thesis carries out the joint research on the recognition of serial verbs sentences and the recognition of semantic relations of serial verbs to propose a joint model,and then verify its effectiveness through experiments.Compared with the pipeline model,it is proved that the approach this paper proposed that decomposing the problem and training different models for sub-problems is working. |
