| Accidents and disasters are one of the key point of public emergencies,which have an important impact on public security risk identification and emergency response.The popularity of the Internet and we-media makes it possible for us to obtain valuable information about public emergencies in all kinds of short texts.Event extraction plays an extremely important role in the mining and extraction of accident and disaster event information,for example,obtaining information on the situation in time,establishing a structured case library,etc.Because the text data feature determines the upper limit of the effect of event extraction method,it can make the model obtain more higher dimension and richer semantic information by raising the semantic feature of character granularity to word granularity or entity granularity.Based on this,the dissertation studies the task of event extraction from the perspective of text semantic feature enhancement,based on the short text information related to accident and disaster events.The research content includes the following two aspects:(1)Research on event extraction method based on named entity recognition task feedback enhancement.In order to enable the event extraction model to learn semantic features of higher-dimensions in the data,this paper improves the classical event extraction model Bi LSTM-CRF,and proposes an event extraction method FB-LatiiceBi LSTM-CRF based on named entity recognition task feedback enhancement.Firstly,the Lattice mechanism is integrated with the bidirectional long-term short-term memory network Bi LSTM as the shared layer of the model to obtain the semantic features of the words in the sentence;secondly,the named entity recognition auxiliary task is added to jointly learn and mine the entity semantic information.At the same time,the output of the named entity recognition task is fed back to the input,and the word segmentation result corresponding to the entity is extracted as the external input of lattice mechanism,so as to reduce the load of large number of self-organizing words in the mechanism,and further strengthen the extraction of entity semantic features.Finally,the maximum Gaussian likelihood estimation method is used to maximize the uncertainty of the same variance,calculate the total loss of the model,and solve the problem of loss imbalance caused by multi-task joint learning.The experimental results show that the improved method based on the feedback enhancement of named entity recognition task effectively improves the effect of event extraction.(2)Research on text preprocessing method based on knowledge graph.By adjusting the model structure,the learning ability of the model to the high-dimensional semantic features of the text can be enhanced.Meanwhile,it leads to the problems of high structural requirements and poor generalization.Accordingly,based on knowledge graph technology,this dissertation proposes a text preprocessing method that can increase the explicit semantic features of text corpus,which can be learned by model easily.Firstly,it define the entity type and extract the entity from the accident disaster event related corpus,establish the relationship between entities through entity relationship expansion,and construct the knowledge graph of the target event "accident disaster" based on neo4 j graph database.Secondly,the sliding window scanning method and entity linking technology are used to recognize the entity and segment words in the text corpus.Finally,through the embedding of text corpus word granularity as well as entities and their attributes,the text embedding results are obtained,containing the semantic information of entity attributes and entity segmentation words.Experimental results show that this method can be widely used in all kinds of event extraction models,and can significantly improve the effectiveness of the model. |
