Deep Learning Based Extraction Of Events And Temporal Relations

As important parts of information extraction,event extraction aims to extract events from unstructured natural language texts;temporal relation extraction is dedicated to identifying the relationships between extracted events and mining the logic between them.Event consists of trigger and argument.Trigger indicates the occurrence and the type of event,while argument is the main participant of the event.In general,event extraction methods first identify trigger and then identify the specific arguments.Accordingly,general event temporal relation extraction methods only consider the triggers and identify the temporal relation between them.For event extraction,this thesis proposes a Hierarchical Distillation Network(HDN)model to obtain the sentence representation with rich semantics,and then predicts the events by means of sequence labeling and pairwise category respectively.Specifically,HDN integrates a recurrent neural network(RNN)module to extract sequential information of sentences,and several graph convolutional network(GCN)modules to extract multi-order syntactic information.Moreover,a distillation module is designed based on bidirectional attention mechanism to reduce information redundancy between encodings.In the experiments,this thesis tests the performance of the model on two commonly used event extraction datasets:ACE 2005 and MLEE,the experimental results show that the proposed model has achieved significant improvement in the result of event extraction than the previous best.For event temporal relation extraction,a Logic-driven Deep Comparison Network(LDCN)model is proposed to solve the existing problems of(1)the feature construction of event pairs is too simple and(2)the integer linear programming(ILP)is separate from training process.LDCN not only employs the HDN as the encoder,but also enhances the features of events pair from the perspective of the commutative and the non-commutative features,and finally introduces the logic-driven differentiable training framework,enables the model to consider the logical consistency in training.In the experiments,this thesis tests the performance of the model on two commonly used event temporal relation extraction datasets: Time Bank-Dense and MATRES.The experimental results also show that the proposed model has achieved significant improvement in the extraction results of event temporal relation than the previous best.