Monte Carlo Beam Search Decoder For Improving Quality of Automated Generated Stories

Abstract

Automated story generation is among the many challenges under the natural language processing field. In automated story generation, the machine will try to generate a sequence of sentences that form a coherent story. This paper will focus on methods to improve the quality of each generated sentence. Each sentence is turned into an event which will be used to generate the next event. In this study, translating events to sentence using a Monte Carlo beam decoder will be explored. In Monte Carlo beam search, the beams which contain sentences will be reweighed based on if they contain certain words from the input event. This will allow a sentence with high matching that might have low score to become the top candidate.