Memory-based approach to learning commonsense causal relations from tex

In this research, causation is viewed as pragmatic knowledge about occurrence of events, their interrelations, and the mechanisms by which events come into existence. As such, causal relations can be learned from events described in natural language text.;A memory-based organization of episodic events and causal relations is presented. Memory is organized into episodes each of which contains a semantic representation of a coherent set of events. Causal reasoning is carried out by causal heuristics that can rule out causal relations or induce new ones in the given context. Causal learning is defined as finding inter-episodal and intra-episodal causal connections between events. Learning involves acquiring causal knowledge in the absence of any precedents, and modifying the current causal hypotheses in light of new episodes. A set of causal connections forms a hypothesis of the causal phenomena at hand, which is represented as an scAND/ scOR causal graph. Modification of causal graphs is accomplished by recalling from memory the episodes for which the hypothesis was derived. The knowledge structure is capable of representing different facets of causation in a unified framework.;The utility of the proposed mechanism is shown in the design and implementation of the causal learning program called NEXUS. The program processes episodes expressed in natural language one at a time and reflects its understanding of the causal phenomenon in its long term memory structures. The test results have shown the feasibility of such a system.