The Enhancing Moral Machine
Moral Machines tend to be defined by a static set of rules. The research in ethically evolving ones is still young. This project researches and presents two possible ways of achieving changing morality in a chatbot through user interaction.
Aleksa Petakovic, 2019
Bachelor Thesis, Institute for Information Systems, School of Business FHNW
Betreuende Dozierende: Bradley Richards
Keywords: Moral Machine; Chatbot; Machine Learning; Markov Chains; Changing Morality
Many of the currently available methods have very limited support for autonomous changes in a chatbot following its deployment. Models built using mainly AIML or Machine Learning can provide an output based on data available only prior to the deployment. While AIML is more adept at implementing a rule-based ethical code for a chatbot, it would require manual modifications each time it is desired to change its behavior. A Machine Learning model is capable of building one on its own, based on the dataset that is given to it. However, once it is trained and deployed, it does not change its behavior to account for post-deployment interactions.
In order to satisfy the requirement of a dynamic chatbot, two main approaches are proposed: Machine Learning and Markov Chains. Although Machine Learning is not well suited for changes post-deployment, it is proposed to follow a cycle of training, deployment, collecting data, re-training, and re-deployment. This approach makes it more dynamic. The second approach is built around Markov Chains, which are more suited for a dynamic chatbot as it does not need to be trained in order to consider newly acquired information in its output. To increase the quality of the chatbot, a neural model for sentence analysis was included.
Both of the models developed are functional and capable of evolving their morality. The Machine Learning (ML) one is very slow at changing as it requires a considerable dataset and many hours of training to product quality results. With a large dataset, the acquired information has a smaller impact on the output.
The model built with Markov Chains and sentence analysis is capable of using new information almost immediately and requires a fraction of the dataset necessary to build a quality ML model. New interactions can have a high impact on the chatbot, depending on the size of the dataset used. Similarly to the Tay chatbot it changes its morality through user interactions. The quality of the output was considerably increased by adding sentence analysis to identify input's object of a preposition (pobj) and the subject (nsubj), with pobj having a higher priority. This inclusion helps the chatbot deliver an output that is more likely to be related to the input.
Studiengang: Business Information Technology (Bachelor)
Fachbereich der Arbeit: Wirtschaftsinformatik & IT-Management