A Study on Training Story Generation Models Based on Event Representations

For story generation tasks, the event representation, a semantic abstraction of sentences, improves neural network based approaches. That is, to expect better generated results, we can train generative models with the extracted event representations as input, instead of the original sentences. However, the process for training such event-based generative models has yet been adequately studied. In this work, we found that proposed n-gram based approaches outperformed the prior all-to-all approach; among n-gram based approaches, 2-gram obtained best training and validation perplexity of generated results. Moreover, we found that an event reversing strategy with a single directional recurrent neural network (RNN) outperformed more advanced architectures such as bidirectional RNNs (biRNNs) and transformers, in terms of training perplexity.