| Event is one of the fundamental concepts for describing the principles of human activity and changes of the objective world.Events occur throughout a person’s life.Massive events happen around the world every day.Hence,understanding the events,along with patterns that describing the consecutively evolutionary between events in time and space,is of utmost importance for comprehending the patterns of human behaviors and social development,and can also be helpful for a variety of artificial intelligent applications,such as recommendation system,and dialogue system.However,understand the events and the relationships between events still remains challenging.As first,the scope of events is extremely broad.Second,events cannot exist in a vacuum of context which leads to the conditionality of event logic.The same event may correspond to quite distinct goals and purposes when it occurs under different background,and lead to entirely different results and effects.Moreover,the pattern of event relationships could be quite complex.On the one hand,there are various kinds of relationships between events,such as temporal relationship,causal relationship,upper and lower position relationship.On the other hand,relationships with different topological complexity may also be formed between events.Last but not least,a thorough comprehension of event relationship patterns entails not only “knowing what”,but also “knowing why”,or in other words,the ability to go beyond simply concluding the phenomena and,but to abstract and summarize the ”theoretical”explanations that describe why these relationships between events exist in some extent.This serves as both a test of the model’s ability to generalize and a barometer for how well it comprehends the connections between events.All of the aforementioned problems’ remedies must be grounded in enough commonsense knowledge.Hence,this thesis systematically investigates how to integrate a variety of external knowledge,such as event background knowledge,event graph knowledge describing the relationship pattern between events,and knowledge graph knowledge describing the relationship between entities and concepts,to support the understanding of events and their relationships and then complete reasoning tasks.The research of this paper mainly includes the following aspects:(1)An event background knowledge enhanced event attribute knowledge inference method.Events are inextricably linked to the particular context in which they took place.The precise context in which the event may be located must therefore be taken into account while comprehending an event and extrapolating event-related features(such as the emotions,intents,and personality traits of event participants).To address this issue,this paper proposes a context-aware variational autoencoder,which introduces an additional latent variable to capture event context information from corpus with abundant event background knowledge,and then guide the subsequent inference process with the captured contextual information.Experimental results on two event attribute knowledge inference task show that the context-aware variational autoencoder can improve the accuracy and rationality of event attribute knowledge inference.(2)An event graph guided complex event relationships reasoning method.The main challenge in event reasoning is that there can be complex patterns of relationships among events,which can forms into a heterograph between events.To address this issue,this paper proposes to enhance the deductive event reasoning process with an event graph that describes the event evolution pattern.However,the flexibility and diversity of event expressions makes event graphs inherently sparse.This limits the application of the traditional retrieval-based knowledge acquisition method to the event graph.On the basis of the previous method,this paper proposes a predictive event deductive reasoning model for integrating the event graph knowledge.Specifically,we extend the BERT model by introducing a structured variable,which captures the connection strengths between events.During training,both context events and external event graph information are used to train the structured variable.During testing,the structured variable which describes connection strengths between events is obtained using the context event only,which is used for finding the next event.Subsequently,we encode the predicted link strength for making a prediction.(3)An event graph enhanced explainable event reasoning method.On the basis of understanding event relationships,this paper hopes to further improve the prediction process of the model to an interpretable space,so as to promote the application of related models in fields with high requirements for the reliability of results.To this end,this paper explores to introduce two levels of explanatory information into the causal reasoning process,i.e.,procedural explanations that describes the mechanism of causal transmission,and conceptual explanations that describe the necessary preconditions for why the causality can exist.To this end,this paper proposes an interpretable causal inference framework based on an eventual graph to obtain procedural explanations from a pre-built eventual graph to reveal the causal transmission mechanism behind causal facts.Experimental results show that introducing procedural explanations into the causal inference process can enhance the accuracy and reliability of the model.Considering the difficulty of automatically inducing conceptual explanations,this paper first constructs an interpretable causal inference dataset providing conceptual explanations.The causal facts and conceptual explanations within the dataset constitute into an important part of the event graph.This dataset also constitutes the largest causal inference dataset to date.Based on this dataset,this paper trains a conceptual explanation-enhanced causal inference model and investigates the potential use of conceptual explanations in causal inference.Experiments demonstrate that extra conceptual justifications can enhance the stability of the findings and foster a deeper comprehension of the causal facts.(4)A multi-category knowledge graph enhanced explainable event reasoning method.To achieve deep udserstanding of the event relationships,abductive reasoning calls for tracing the explanations theories from the observable phenomena events..In order to complete this process,it is necessary to simultaneously synthesize the relationships between events,entities,and concepts.After that,the explanation’s plausibility should be assessed using the broadest possible set of empirical knowledge.To address this issue,on the basis of the deductive event reasoning method proposed in this article,this paper proposes to further integrate the event graph with the knowledge graph that describes the relationship between entities and concepts,and construct a multi-knowledge graph to support the tracing to the explainative theory.Moreover,a multi-knowledge graph enhanced variational language model enhanced is further proposed to model the probability distribution for the plausibility of empirical knowledge,so as to find the best explanation according to the most plausible empirical knowledge.Experiments show that the method proposed can effectively utilize all kinds of common sense knowledge contained in the multi-knowledge graph to improve the performance on reasoning process.In conclusion,considering the challenges within the event reasoning task,this paper starts from the task’s characteristics and systematically investigates ways to enhance the existing approaches from the perspectives of models,knowledge representation,and model training,in order to integrate a variety of external knowledge into the reasoning process,to enhance the understanding of events and the relationships between events.This thesis also aspires to promote the development of downstream related applications,and ultimately fostering the growth of the overall field. |
PDF Full Text Request