University of Twente Student Theses


Affective dialogue generation for video games

Kalbiyev, A. (2022) Affective dialogue generation for video games.

[img] PDF
Abstract:Affective text generation has been a topic of interest within the Natural Language Processing community and has been left scarcely explored within the context of the gaming industry. With this project, we aimed to bring the paradigm of affective text generation into video games. This was done by developing a generative language model that generates affective dialogue for the video game Fallout 4, comparing the generated text to human-written one and measuring how accurate it is at exhibiting the correct emotion. To do this we have selected, extracted and pre-processed the Fallout 4 dialogue dataset, then fine-tuned the Generative Pre-trained Transformer (GPT) 2 language model on the prepared data. Afterward, we have implemented the affective extension which incorporates affect into dialogue generation. Using the fine-tuned GPT-2 model together with the Affective Extension and Top-K sampling method, we developed a generation pipeline that generates affective dialogue for Fallout 4 given any in-game prompt string. Then, a human evaluation was performed to compare the responses generated by the model to human-written responses on metrics - Coherence, Relevance, Fittingness and Human-likeness - and to measure how accurate the responses are in exhibiting a given target affect. The results of the survey suggest that the model-generated responses compares poorly to human-written ones on all of the metrics and the responses generated by the model exhibit affects unsuccessfully.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page