Research Of Event Extraction Technology In The Field Of Accident And Disaster

In recent years,environmental issues have become increasingly serious,and accidents and disasters have posed a huge threat to people’s lives and property worldwide.We need to obtain first-hand accident and disaster information in a timely manner in the event of an accident or disaster,and make corresponding response strategies to reduce losses.Event extraction is a type of information extraction that is of great significance for the mining and extraction of event information.In the event extraction task,there is no natural pause between Chinese short text information and there is also a phenomenon of polysemy in Chinese words.Therefore,during the extraction process,there are issues of mismatch between trigger words and event types,and multiple event types corresponding to one trigger word,which can affect the accuracy of downstream tasks.On the other hand,in the task of text level event extraction,it is more important to pay attention to the long-term dependence on neural network models.This article has optimized previous event extraction models.This article is mainly divided into the following two parts for research:(1)In order to solve the problem that trigger words and event types do not match in the event extraction task,and that one trigger word corresponds to multiple event types,characters are taken as the minimum unit in the event type extraction part,and vector representation and learning of cleaned data are performed by using Skip Gram model.The processed vector is input into the Lattice LSTM model for feature extraction,while extracting character level and word level information.By adding a self attention mechanism layer to learn the correlation between the characters at each position and the remaining positions,the learned trigger words are more accurate and semantically relevant.Finally,add a sequence marker that connects the cell states between each starting and ending word.These information are recorded in semantic relationships,which can be combined to optimize the problem of event type mismatch corresponding to polysemous trigger words.(2)Due to the fact that most of the information we obtain from the internet appears in the form of short text,unlike event extraction in individual sentences,in discourse level event extraction tasks,the event arguments of an event may be contained in several different sentences.If we want to extract all the arguments of the same event scattered in several sentences,Attention should be paid to addressing the long range dependency on neural network models in event extraction tasks.The extraction of event roles proposes an end-to-end neural sequence model that integrates mixed features for discourse level event extraction.We transform the discourse level event extraction task into a sequence labeling task for a set of consecutive sentences in the article,and the model enriches the contextual features in the Embedding layer by using word embedding and context embedding.By using DMCNN and Bi LSTM to dynamically learn sentence level and discourse level features,and using a gated fusion mechanism to dynamically fuse features at different levels.Finally,the self attention mechanism and CRF layer are used to aggregate feature information,capture the internal and long-range dependencies of features,and jointly model labels to improve the performance of the model.