Research On Chinese Semantic Role Labeling Based On LSTM

With the rapid development and popularization of computer network technology and communication technology,the application demand of natural language processing technology has increased sharply.People urgently need practical natural language processing technology to help breaking the language barrierand providing convenient,effective,humanized service for exchanging information between human and machine.As one of the branches of natural language processing,Chinese information processing has been developed rapidly in recent years.In terms of the basic theory research or in the development of technology and industrialization,it has made remarkable achievements.Semantic role labeling is a kind of implementation of shallow semantic analysis,researchers have been concerned about it in recent years.Deep learning is a kind of technology that allows computer to learn featureautomatically.As it has achieved great success in image recognition,speech recognition and so on,the researchers gradually began to apply this technology in the field of natural language processing,which became a hotspot in this field.In the current multiple deep learning model,LSTM（Long Short-term Memory）is considered to be particularly suitable for processing of text sequences data,because it can efficiently utilize long-distance dependency information in sequence data.Therefore,this paper proposes a Chinese semantic role labeling model based on LSTM,it avoids complicated feature extraction and selection,and gets rid of the dependence of syntactic parsing,and the best annotation result is 70.34%.The main works of this article are as follows:（1）Determinethe experimental corpus and marker set: on the basis of the CPB annotation corpus,we determined 19 classesof semantic role.Considering the characteristic of the model,we choose IOBES to annotate sequences and form 77 kinds of labels.The experiment uses documents chtb₀001.onf – chtb₀399.onf of OntoNote 5.0by 3:1rate for training and test corpus.（2）Construct and train the semantic role labeling model based on LSTM: This paper takes the word as the basic annotation unit,utilizes the word embedding obtained by the Word2 Vecas input,constructs the network layer to learn the semantic role-related characteristicswith LSTM standard unit,and obtains the word corresponding semantic role label with the characteristic vectorsby the Softmax function computation and the post-processing.The model training is carried out by back propagation algorithm,and tests the parameters of the model.（3）Use the LSTM model to obtain the part-of-speech（POS）vectors and combine with the word embedding for semantic role labeling: Firstly,we constructeda LSTM network to learn the POS vectors.Then we combined the POS vectors with the word embedding as input to construct and traina LSTM network to get the semantic role labels of each word.Finally,it experiment with the model parameters and compare the results with the preceding model.The result shows that the POS can help the semantic role recognition and classification.The model paper proposed is effective in automatic semantic role labeling.Although the model in this paper could not comparable with the best results based on the feature of artificial extraction,it has been achieved good results,and shows the strong ability of LSTM in semantic role labeling tasks.