Multi-Context Dependent Natural Text Generation for More Robust NPC Dialogue

Abstract

Generating natural text while capturing specific contexts has been a difficult problem within natural language processing. This work studies methods to generate video game non-player dialogue that encapsulates multiple categories of contextual data, including but not limited to personality, events, and actions. Three language model architectures were considered for the task: a character-level long-short term memory network (LSTM), a word-level LSTM, and a convolutional neural network (CNN)-LSTM encoder-decoder network. The models were trained on data from \textit{Animal Crossing: New Leaf} and was evaluated qualitatively via a subjective survey. It was found that the encoder-decoder network was able to capture the contexts in its generated texts well, almost comparable to that of the training data. The results affirm that spatial features are crucial for establishing long term dependencies to properly capture context within the generated content.