Natural Language Content Generation for Computer Games

Abstract: The demand for engaging and fresh game content is steadily increasing in the modern computer game industry. However, the manual process of game content creation is expensive and unscalable which has led to an increased use of Procedural Content Generation methods.In this thesis we explore the possibilities of combining Natural Language Generation with Procedural Content Generation methods to produce natural language game content for the computer game Crusader Kings II. We present several theoretical proposals for Natural Language Generation techniques suitable for Crusader Kings II. Furthermore, we propose, implement and evaluate a new method, MergeTree, that is a variation on the traditional NLG pipeline. The new method combines the existing softwares CoreNLP and SimpleNLG into a new pipeline architecture in order to rewrite existing surface texts. The MergeTree parser is an intermediate tool which creates a Text Specification for a given surface text. This allows for additional microplanning in order to generate linguistic variations of the existing surface texts.Fifty sentences from existing event descriptions in Crusader Kings II were given as input to the MergeTree parser, which could correctly handle 30% of the inputs. Additionally, 26% of the inputs caused minor errors which could be resolved with future development of the MergeTree method.