Research On Chinese Framenet Semantic Role Identification Based On BiLSTM

Semantic analysis is a hot and difficult issue in the field of natural language processing for many years.It is also the main technical bottleneck of machine translation,information retrieval,search engine and other application systems.Chinese frame semantic role identification is an important part in Chinese-oriented frame semantic analysis.At present,The neural network based on Bi LSTM is widely used in semantic role identification and labeling tasks and has achieved good results.However,this model also has some problems.For example,the prediction results of the model are not stable and the reproducibility is poor,training the model on a large corpus is computationally expensive,the performance of the model is also closely related to the setting of input features.In this paper,we study the modeling method of Chinese frame semantic role identification task based on Bi LSTM neural network.We Focus on the research on hyperparameter tuning and the improvement of representation learning in modeling.We select four candidate features(the current word,the POS of the current word,the target word,and the position information of the current word relative to the target word)and two network design choices(the number of Bi LSTM layers and whether a CRF classifier is added at the top of the model)and consider them as hyperparameters of the model.The traditional hyperparameter tuning method generally splits the corpus into training set,validation set,and test set according to the ratio of8:1:1,then uses the greedy strategy of adding hyperparameters to the model one by one to select the optimal combination of hyperparameters.However,the traditional method is computationally expensive and results in poor stability.Therefore,this paper presents a hyperparameter tuning method based on the robust design.This method only needs to perform experiments with 3 × 2 cross-validation on a small corpus to obtain the mean and variance of the model performance estimation,adopts the idea of robust design and uses the signal-to-noise ratio of model performance metric as the optimization target to select the optimal combination of hyperparameters of the model.Experimental results show that the tuning method in this paper is better than the traditional method based on greedy strategy.It can also use ANOVA to quantitatively analyze which hyperparameters have a significant impact on model performance,making the optimal model selected in this paper has better interpretability.In this paper,we proposed three improved methods for the representation learning of the initial vector of the input features.One is an improvement on the representation learning method(Glo Ve)of current word feature,and proposes a calibrated word representation learning method.The method calibrates the co-occurrence frequency based on the Zip'f distribution to learn higher precision word vectors.The second is to give a word representation learning method that can distinguish left and right context information,which effectively integrates the position information of the current word relative to the target word.The third is to propose a frame representation learning method based on frame disambiguation.The frame representation learned by the method can maximize the distinction between different frames evoked by the same word,and effectively integrate the frame representation information of the target word into the model.Experimental results show that these three improved methods all improve the performance of the semantic role identification model,to varying degrees.The methods in this paper provide new ideas for solving the problems of hyperparameter tuning and feature representation learning in the modeling of other natural language processing tasks based on deep neural network.