Inferencing over common-sense knowledge for story generation
Abstract/Summary
Story generation systems rely heavily on their knowledge base in order to come up with stories. Most of these systems manually build their knowledge base from scratch. As a result, the information contained in the knowledge base is often very specific for the intended stories. This greatly affects the quality and quantity of the stories to be generated. This research made use of existing sources of knowledge, primarily ConceptNet, together with domain-specific knowledge for the automatic generation of childrens stories. Information from other sources of knowledge, WordNet and VerbNet, have been extracted to supplement ConceptNet. Based on the results of the evaluations, ConceptNet has been found to be able to generate an ample amount of stories when it knows a lot about the concepts needed in order to tell the stories. Otherwise, additional knowledge have to be supplied. Furthermore, due to the nature of common-sense knowledge, the quality of the stories produced will increase as more domain-specific knowledge is added. Setting a threshold value on the minimum confidence score of an assertion before it can be queried that balances both the correctness and amount of information retrieved has also been found to produce better quality stories. Keywords: storytelling, story generation, knowledge representation.
