Research On Temporal Relation Identification Between Events Via Neural Network

Temporal relation identification between events is a task to identify the pairs of events that have a temporal link and classify the temporal relations between them,which is crucial to any system that attempts to understand natural language in depth,such as question an-swering(QA),information extraction(IE),text summarization,etc.Early studies on event temporal relation identification usually focus on extracting linguistic features.These meth-ods rely heavily on manually-annotated features and external knowledge bases,while the application of neural networks with more flexible forms and more generalization ability is relatively rare.This dissertation focuses on the application of neural methods to temporal relation identification between events,and the main contents are as follows:(1)Identifying temporal relations between events by deep Bi-LSTMNeural networks illustrate their advantages in comparison with traditional feature-en-gineering based methods for event temporal relation identification.However,existing neural networks are usually of shallow architectures(e.g.,one-layer RNN or CNN),so they may be unable to explore the potential semantic representation space in different abstraction lev-els.To address this issue,this paper proposes to use deep bidirectional long short-term memory networks to identify temporal relations.By concatenating the outputs of all prior layers together as the input for the subsequent layer in the network,the information can flow more sufficiently.The experimental results indicate that the proposed method can improve the identification performance effectively.(2)Temporal relation identification between events on combination of self-ttention and neural networkIt remains a major challenge for conventional RNNs or CNNs to handle structural in-formation and capture long distance dependence relations.To address this issue,this paper proposes a neural architecture for event temporal relation identification based on self-atten-tion mechanism which can directly capture the relationships between two arbitrary tokens.The identification performance is improved significantly through combing self-attention mechanism with nonlinear layers.The contrast experiments prove that the proposed method outperforms most of the existing neural methods.(3)Joint learning for event temporal and casual relation identification via neural net-workTemporal and causal relations between events are closely related and one relation often dictates the other.However,joint study on these two relations has been extremely limited.This paper proposes a joint learning framework based on neural networks to integrate the two identification tasks,in which the interaction between the two relations is realized by sharing the information between the network layers of the two identification models.The experimental results show that the joint learning framework significantly improves the per-formance of temporal relation identification by leveraging causal information between events.In this paper,three effective methods of event temporal relation identification are pro-posed.Aiming at the problems existing in event resolution,such as shallow architectures,hard to handle structural information and capture long distance dependence relations,and ignoring the correlation between temporal relation and other event relations,different solu-tions are proposed,and the performance of event temporal relation identification is greatly improved,which demonstrates great value of academic research and practical application.