Chatbot implementation and testing

Damjan Košir (2013) Chatbot implementation and testing. EngD thesis.

Preview

PDF Download (544Kb)

Abstract

In my thesis I did a research on different chatbot technologies. I only looked at technologies that enable the use of chatbots on a local computer. There are two approaches to determine the chatbot's content, including knowledge and personality, the scripted and the learned approach. If we choose the scripted approach, we determine the content ourselves, if we choose the learned approach, we get it with machine learning from a conversation database. Since I didn't have access to a conversation database, I had to take the scripted approach to determine the content. I identified AIML (Artificial Intelligence Markup Language) and Chatscript as the technologies with the most potential. The main advantages of AIML are a large community, a large collection of accessible chatbots available for free, which can serve as a base for a new chatbot, and many interpreters. The main advantages of Chatscript are a strong pattern language, enabling faster and easier content writing, active development and the possibility of the use of databases, such as Wordnet. After researching the performance of both chosen technologies, I decided to use Chatscript to make my own chatbot. Next came the description of the chatbot's personality and writing the content divided into different topics. For proper changing of topics and a good conversation flow with the chatbot I had to adjust the control script. After the first development phase there were several testing and improvement iterations. Due to a lack of objective chatbot quality metrics this was the best way of development. With the newly acquired knowledge and experience I concluded that the technologies available at the moment are good enough to be used for entertainment or as virtual assistants, but they still haven't reached the level to pass the Turing test and thus make people unable to distinguish a chatbot from a human conversation partner.

Actions (login required)