Abstract
Small to medium enterprises and the general populations in developing nations are affected by infrastructural unavailabilities and isolated location restrictions, which hinder the adoption and growth of innovation. The effects of these restrictions have been noticeable in branchless banking services, where services such as loans and credit have inadequately adopted across rural areas. Also, financial services agents that facilitate branchless banking services is not as innovative as monopolies in the industry such as banks or financial technology institutions in developed nations. Branchless banking has seen significant growth in developing nations, however there is concern regarding trust for secure transactions and money management. Artificial intelligence has seen rapid growth in providing innovation in banking services across developed nations. Intelligent agents, a subdivision of artificial intelligence, provide autonomy and flexibility within the environments in which they operate. The dissertation examines the use of multi-agents in dialogue systems through text-based natural language processing. This provides branchless banking services that can be used in remote locations. The dissertation introduces a multi-agent mobile account model that offers secure transactions, online-offline scheduling for resources, service training and question and answer services. The model serves as a mobile solution to autonomous service offering and tests for usability, online-offline channel integration and secure transaction ability. Five different dialogue pipeline model configurations are used to train the natural language dialogue components of the agents. The F1-score evaluation metric performs significantly better on entity classification when compared to Q&A intent responses. Human-Computer interaction; agent beliefs, desires and intentions; natural language understanding and research goals were the categories for prototype testing. The human-computer interaction, agent beliefs, desires and intentions and research goals tests were quite successful as they performed tasks and conformed to researched outputs. The natural language understanding category illustrated that training and test datasets require more examples and consideration towards a human interaction with the model to train the system better.
M.Sc. (Computer Science)