| Text event extraction is one of the main tasks in knowledge graph construction,which has a wide range of applications in information retrieval and question answering systems.Text event extraction technology tries to extract accurate structured information from unstructured natural language,and provides a more convenient way for human text processing.Document-level text event extraction is an important technology for text event extraction,but the commonly used text event extraction models such as DCFEE cannot solve the problems of argument dispersion and multi-event in text.In view of this,this paper improves the encoder and decoder,In this paper,roberta-DEE(Ro BERTa Document Event Extraction)model based on hybrid neural network and roberta-SL-DEE(Ro BERTa)model based on self-attention mechanism and improved decoder are proposed Self-Attention Event Extraction),which improves the performance of document-level text event extraction.The main research work is as follows:(1)Aiming at the problem of argument dispersion and multi-event in document-level event extraction events,that is,an argument of an event has multiple events in multiple sentences and a sentence has multiple events,a document-level text event extraction model Ro BERTa-DEE based on hybrid neural network is proposed.Ro BERTa-WWM and Bi LSTM(Bi-directional Long Short Term Memery)are used to encode text.Experiments on datasets in the financial field show that the average F1 value reaches 76.62%,which proves that the effect of document-level text event extraction based on this model is better than that of comparison models DCFEE-O,DCFEE-M,Greddy Dec and Doc2 EDAG.(2)The conditional random field as a decoder has weak ability to capture the context information of long text,and the effect of decoder to extract feature information is not ideal.In order to solve the problem,a text event extraction model Ro BERTa-SL-DEE based on self-attention mechanism and improved decoder is proposed.The model uses the self-attention mechanism and LSTM-CRF as the decoder,and captures the long-distance dependence within the sequence by weighting the representation of each position information in the input sequence to better understand the structure of the sequence.Through the design of the gated unit,the processing ability,memory ability and stability of long sequence data are enhanced.Experimental results show that the Ro BERTa-SL-DEE model is superior to the comparison models in most F1 values on single-event and multi-event datasets,which proves that the extraction effect of the proposed model is improved. |
