| With the development of science and technology and the improvement of people’s living standards,private cars are rapidly pouring into our lives,while the number of traffic accidents is increasing.At the same time,the causes of traffic accidents also show the characteristics of diversification and complexity,the analysis of accident causality has become an increasingly important topic.However,the current analysis of traffic accidents generally focuses on the direct influence of several factors such as people,vehicles,roads and environment,and lacks the detailed analysis of the evolution of causality in the process of accidents.In view of the above problems,we use deep learning technology to construct a traffic accident mechanism map to analyze the causality of traffic accidents,among which causal event extraction and causal relationship extraction are indispensable parts in the process of constructing the mechanism map.The main work of this paper is as follows:(1)In view of the current situation of missing text data set of traffic accident causality,we collected the traffic-related texts on the official government website from the Internet and processed them,the processed texts were manually annotated according to the requirements,then constructed the traffic accident dataset.(2)In view of the problems such as the difficulty in finding event boundaries and insufficient text representation during the extraction of traffic accident causal events,the extraction task is transformed into sequence labeling task and a model based on attentional mechanism is proposed which name multihead self-attention convolution Bi-LSTM network with mixed char-word and BERT embeddings(MACL).The model uses a mixture of word vectors,word vectors,and position vectors to encode the representation of text,and also introduces a vector matrix generated by bidirectional encoder representations from transformers(BERT).Convolutional neural networks(CNN)and bidirectional long-and short-term memory networks(Bi-LSTM)were then used to extract contextual features and long-range features respectively.Multiheaded attention mechanism(MHSA)is used to assign weights so that event boundary locations receive more attention.Finally,a conditional random field classifier is used to classify each word to obtain the semantic role of the word and thus extract the causal event.(3)In order to extract the causality successfully,we adopt the way of question answering tasks to extract the causality.First,we use the BERT model to get vector matrix and encodes the words to enhanced presentation capabilities of traffic accident texts,then we extract the cause events from the input sentence by the strategy of point net,combine the extracted cause with the incomplete question template to obtain complete question sentence,then transform the causality extraction task into a question answering task,finally we get the cause-effect pair.(4)We built the application system of traffic accident causality event evolution graph.Based on the traffic accident causal events extraction and the causality pair,we get the causality event evolution graph what is generated by cosine similarity calculation method,and be applied to the event query and accident process evolution.In this paper,we use the constructed traffic accident causality dataset to verify the above two key task models and compare with other mainstream models.The experimental results verify the validity of the proposed model.At the same time,the generated traffic accident causality event evolution graph shows the evolution process of accidents more intuitively,which provides support for preventing and avoiding traffic accidents from happening again. |
