Word embeddings as knowledge base for flavour text generation
Judith van Stegeren


Game development involves writing a large amount of game text. Game writers prefer to spend most of their time writing the “core” game texts: the text that defines the game, such as the story or NPC dialogue. However, writing the so-called flavour text, i.e. the text that determines the look and feel of the game, can take up the majority of their time. In this research we look at automatically generating flavour text, given a short human-written input text. Specifically, we try to generate fictional news headlines given a set of seed words. The idea is to generate headlines that feel coherent with the input text (for example, core game text), by employing content words that are related to the input text.
We have implemented a pipeline for generating fictional headlines: we extract seed words from a human-written input text, scrape headlines from news outlets and substitute parts of the headlines with words that are related to the input. We pick the to-be-inserted words from a knowledge base of related words.
Indeed, using a knowledge base in the generation process is common practice in both game content generation and NLG systems. However, instead of manually building the knowledge base, we have chosen a knowledge base in the form of a vector space of word embeddings. Using word embeddings has the advantage that game developers can train the knowledge base instead of writing it by hand, or even load a pre-trained model that can be used immediately.