Recognition And Extraction Of Text Causality In The Financial Field

Causal relationship recognition and extraction is an important and challenging task in natural language processing,which aims to identify causal texts and extract causal entity pairs.In view of the lack of feature information and low accuracy in the process of causal relationship identification and extraction in the current financial field,this thesis proposes to identify causal relationship based on Bi-LSTM and two-way CNN network model BTCNN(BI-LSTM and Two-way CNN).A causal extraction model of Graph convolutional network(T-GCN)based on attention mechanism(Transformer)is proposed to make it more suitable for the task of causal extraction.The specific research contents are as follows:A causal relationship recognition model(BTCNN)combining BI-LSTM and dualpath CNN is proposed to fully mine the causal semantic information implied in the text and solve the problem of missing feature information in the recognition process.Firstly,bi-LSTM network model is used to generate text feature matrix.Then,dual-path CNN is used to set different convolution to check the causal features in the text feature matrix for further extraction,and the feature vectors obtained by two different pooling methods(average pooling and maximum pooling)are used for splicing.Finally,the spliced feature vectors are input to the full connection layer for output.Compared with the known methods,experimental results show that BTCNN model improves the accuracy of causal relationship identification.On the basis of BTCNN model for causal relationship identification,extract causal entity pairs from texts with causal relationship.Bi-LSTM +CRF was used as the benchmark model,and T-GCN was introduced to enhance the semantic features of causality and reduce the loss of feature information.Firstly,the text is input to BI-LSTM after word vector encoding,and semantic information between contexts is mined and causal feature vectors are initially generated.At the same time,input the text into Transformer model to get the degree of association between individual words.Then,according to the degree of association between words,the weights of the causality feature vectors are assigned by T-GCN to further strengthen the semantic features of causality and reduce the loss of feature information.Finally,the enhanced causal feature vector is input into CRF layer and the causal label is obtained.The experimental results show that the proposed BTCNN model achieves the maximum precision,recall and F-measure of 82.3%,80.04% and 81.15%,respectively,compared with other models on the OIECAA dataset.In the process of reason extraction,the model in this thesis increases the accuracy of the benchmark model Bi LSTM+CRF,which is 3.4 and 3.6 points respectively.Table [9] Figure [19] Reference [81]...