Event Extraction and Temporal Ordering towards Narrative Model Generation

Narrative generation is the process of automatically generating narrative stories (fictional or not) through a narrative model. With the advances in machine learning and natural language processing techniques, narrative generators have been able to create more creative and interesting text in terms of narrative contents and story telling techniques. These systems have applications in a range of domains including education and patient recovery. Narrative generation models are traditionally planning problems and uses a schema to represent the narrative world. This schema contains states and actions with preconditions and postconditions and the sequence of actions that lead to one or more goal states (dependent on the author’s goal) is considered the skeleton for the narrative story. Many modern systems then employ heuristics to measure and improve storytelling aspects such as narrative suspense. However, these schemas are usually defined manually, requiring significant user input, or otherwise can only represent simplistic events and generate simple narratives. Here we explore the building of a system that can automatically generate these schemas to be used within a narrative model. We present the initial stages of such a system, pertaining steps of lexical processing, event extraction, and temporal event ordering. Our event extraction is performed by a rule-based system that can obtain informative triples accurately to our task without the need of training data. Determining the temporal ordering of extracted events is then done through multiple multilayer perceptron networks using features obtained through statistical and rule-based techniques. Our temporal ordering achieves accuracies of above 70 per cent for both event-event pairs and time-event pairs on Timebank and Opinion corpera. This pipeline presents the initial steps of a system to be able to automatically create states and actions with preconditions and postconditions which will form the narrative schema. We discuss several limitations of the techniques used and present potential improvements and future work towards the complete narrative generator system.