Research On Semantic Role Annotation Based On BiLSTM And Attention Mechanism

At this stage,the research on semantic role annotation has received much attention,and it is one of the most direct methods to achieve semantic analysis,which enables us to capture more information about the text and help us understand it better.Semantic role annotation of financial corpus can help decision makers understand the text more precisely and provide support to them.In this paper,we improve the existing LSTM model based on lexical features and use the Skip-Gram model for training to obtain vector expressions with semantic relationships between words as the input data of the model,and apply the model to the financial field.In the study it was found that the model with more input of text features from the financial corpus performed better during the experiment.However,the LSTM model based on lexical features cannot capture important information in sentences and cannot exploit the contextual relationships following the currently labeled words.To solve this problem,this paper applies the BiLSTM-CRF model to the financial field,and adds a BiLSTM layer to the model to extract lexical features of text,and introduces the attention mechanism into the BiLSTM-CRF model,and proposes the BiLSTMCRF model based on lexical features and attention mechanism.The data used in this paper are derived from finance-related datasets from the Chinese corpus of the University of Pennsylvania.In this paper,after obtaining the vector representation of words using One-Hot for processing,the Skip-Gram model is trained to obtain the vector representation with semantic relationships between words.Then an AttBiLSTM-CRF model based on lexical features is constructed to enhance the focus on keywords by introducing an attention mechanism and to extract the lexical features of the corpus.The extracted lexical features are fused with the word vectors of the corpus as the input of the model for training and testing,and then sequence annotation is performed by CRF using the IBOES strategy.In this paper,experiments on implicit layer nodes,experiments on the number of training times of the model,experiments on the number of implicit layers,and experiments on the activation function are designed and conducted,and compared with existing models.It is demonstrated that introducing the attention mechanism into the BiLSTMCRF model and using the lexical feature extracted feature vectors as input data can significantly improve the performance of the semantic role annotation model.During the processing of the financial corpus,several evaluation metrics of the model were improved.