A Computational Analysis Of Event Causation: Theory And Applications

Causal relations of various kinds are a pervasive feature of human language and theorising about the world. The specification of a satisfactory general analysis of causal relations has long proved difficult. The research described in this thesis is an attempt to develop a computational theory of causal relations between temporally ordered events,and more precisely causal relations among partial states and actions,from the causal reasoning point of view.More specifically,the research provides an appropriate framework of entities among which causal relations are to hold;it also develops a theoretical framework of event causation,under which the structures and elements of causal relations holding among these ontological entities can be described;it gives a general representation tool for event causation supported by the ontological and theoretical frameworks,under which causal relations can be formalized as causal rules for practical reasoning,e.g.,predictive reasoning,in which nonmonotonicity,as well as the other general properties and the nature of elements involved,can be captured;it constructs computational frameworks for abstract causal reasoning models,such as causal prediction,causal explanation,and causal diagnosis;and it finally extends and utilizes these abstract reasoning models to formalize causal knowledge in specific domains to develop practical causal reasoning systems for AI research,e.g.,story understanding and legal reasoning. The research is original from several aspects as follows:(1) The analysis of the internal structure of events provides a fundamental ontology for causal relations. The new ontology is based on entities of partial states and actions,where such traditional different entities as objects,properties,events,states and actions are naturally related to each other.(2) The structural analysis of event causation exhausts possible types of eventcausation. A broad distinction is made between potential causation and actual causation,and in particular,for potential causation,a set of standing conditions in the causal field are distinguished from cause and effect. A analytical theory is established by putting causal elements into partial states and actions,which deepens our understanding of event causation at the level of partial states and actions.(3) A causal rule representation is mapped into default logic formalism,based on the examination of general properties of causation. The default rule representation provides a concise syntactic and semantic formalism for potential causal relations to be used in causal reasoning models such as predicting,explaining and diagnosing. It captures general properties and the nature of elements involved. In particular,it captures intuition that the cause,the set of standing conditions in the causal field play different roles,and particularly that the cause is more important than the others,for inferring the effect from a causal relation. It also captures nonmonotonicity in causal reasoning.(4) Specific causal reasoning systems are constructed. In the domain of story understanding,where default causal chains are used for filling in the gaps between isolated sentences of a simple story to make the story causally well-connected,in turn provides evidence for the computational approach. And in the computer-aided law agency,causal reasoning of legal rules is the core of the system and results can help lawyers have more efficient law documentation.The use of causation is one of basic intelligence for human beings. The systematic work conducted here may be helpful for fundamental research of artificial intelligence.