Creativity Evaluation Method for Procedural Content Generated Game Items via Machine Learning

Abstract

Procedural Content Generation via Machine Learning (PCGML) refers to methods that apply machine learning algorithms to generate game content. In particular, the generation of game item descriptions requires techniques to evaluate the similarity between items, and consequently their creativity. This paper improves the BLEU2vec text similarity evaluation technique by integrating it with Byte Pair Encoding (BPE) to capture the relevance of compound words in generated game item descriptions. This novel technique, called Split BLEU2vec, splits compound words into sub-words enabling their similarity evaluation. Our results show that when compared to BLEU2vec baseline, Split BLEu2vec is able to account for semantic embedding of compound words in item descriptions of the game Legend of Zelda.