| In the era of big data,various articles and reports emerge in an endless stream,and text information is increasing day by day.There are various events and relationships in the text,and causality,as one of them,is particularly important in logical reasoning.How to identify the causal relationship in text has also become a hot area of research.The matching algorithm can effectively identify the causal relationship with causal connectives,but it can’t do anything about the implicit causal relationship that does not contain connectives.The development of deep learning enables the algorithm model to effectively extract the semantic features of the text,which can help the model identify hidden causal relationships.causal relationship.However,in the study of causal relationship recognition,the extracted semantic features are insufficient,the boundary judgment of causality is not accurate enough,and there is also the problem of longdistance dependence on long text sequences.Aiming at the problem of causal relationship recognition,this thesis proposes a causal relationship sequence annotation model combined with graph convolutional network and a causal relationship classification model combined with multi-head attention,and conducts experiments on the self-built data set ECE.The main research content of this thesis is as follows:(1)Part of the corpus was selected from the public corpus for manual annotation,and the experimental data set ECE causal data set was obtained.Aiming at the problem of insufficient feature extraction and inaccurate judgment of causal relationship boundaries,this thesis proposes a BERT-GCN-BiLSTM-CRF model combined with a graph convolutional network for sequence labeling of causal relationship recognition,and uses syntactic analysis to obtain the syntactic relationship between words Features,combined with the word character-level representation formed by the convolutional network and the word vector representation of the BERT pre-trained model,feature extraction and fusion are performed by the bidirectional long-term short-term memory network and the graph convolutional network respectively,and the distance information of the central trigger word is integrated into the semantic features.The extracted semantic features are enriched,and finally the labels are predicted using conditional random fields.(2)Aiming at the problem of long-distance dependence in causal relationship recognition,this thesis proposes a BERT-Bi LSTM-CNN-Multi Att model combined with multi-head attention mechanism for classification and regression of causal relationship recognition,using hole convolution and residual block method Extract the deep N-gram features of the text from the output of the hidden layer,and then use the multi-head attention mechanism to calculate the weight distribution for multiple features,which effectively alleviates the problem of long-distance dependence.The output layer uses a softmax classifier to classify the results.For the above two improved models,this thesis conducts experiments on the ECE dataset.The results show that the model proposed in this thesis has better performance in causal relationship recognition than the baseline model,which verifies the effectiveness of the model. |
