Improving the Fitness Function of an Evolutionary Suspense Generator Through Sentiment Analysis

Pablo Delatorre,Carlos Leon,Alberto Salguero Hidalgo

doi:10.1109/access.2021.3064242

Abstract

The perception of suspense in stories is affected not only by general literary aspects like narrative structure and linguistic features, but also by anticipation and evocation of feelings like aversion, disgust or empathy. As such, it is possible to alter the feeling of suspense by modifying components of a story that convey these feelings to the audience. Based on a previous straightforward model of suspense adaptation, this paper describes the design, implementation and evaluation of a computational system that adapts narrative scenes for conveying a specific user-defined amount of suspense. The system is designed to address the impact of different types of emotional components on the reader. The relative weighted suspense of these components is computed with a regression model based on a sentiment analysis tool, and used as a fitness function in an evolutionary algorithm. This new function is able to identify the different weights on the prediction of suspense in aspects like outcome, decorative elements, or threat’s appearance. The results indicate that this approach represents a significant improvement over the previous existing approach.

Highlights

One common approach to automatic generation of stories is to generate content and evaluate the results against a particular function
Half the stories were generation with the baseline fitness function and, the other half, with the new fitness function. They were randomly assigned to the participants so that they would receive one story per template (3 stories in total)
Whereas it can be argued that this might add a certain amount of bias in the research, the granularity of the research methodology forces a strict control on the content

Summary

Introduction

One common approach to automatic generation of stories is to generate content and evaluate the results against a particular function. This approach is of application in several generative schemes, namely state-space search or evolutionary approaches. Metrics for evaluating the effect of stories include coherence, causality, or character affinity [1]–[4]. The automatic storytelling system BRUTUS generates the sequence of events with a ruled-based system that produces a set of potential consequences for the characters [5]. TALE-SPIN [6] and Fabulist [7] tackle story generation as a planning process, taking a collection of characters with their corresponding intentions and goals, and generate a sequence of states and actions.

Methods

Results

Conclusion