P0-MS4: Part 285: Final Testing, Debugging
by PhoenixGames on Dec.10, 2017, under Astral Realms, P-0, Software and Games, Spiritus Astrum
I have spent a few days working on a concept for a conversational AI system for this project.
I am currently using a simple AIML-powered bot, and I will probably still use AIML for actual data storage for the knowledge base, but the existing system has serious limitations.
The main ones are that it is based on queries and responses only, IE, the AI can’t direct the conversation, only respond to the user. Also, other than some ability to set variables, the AI can’t really “remember” what has been said. It is not possible, for example, for a conversation to flow naturally from one statement to the next, since each query and response is, essentially, self-contained, and has little effect on other queries and responses.
The system that I have come up with is very simple, but it should be able to address at least some of these issues.
It uses templates to store information about similar topics, or similar blocks of information. Within each template, there are several responses, each of which can match with one or several input patterns, and can produce one or several outputs. Not only that, but each response can have an effect, such as setting a variable, or calling a function. This allows easy storing of information, each response can have any number of complex effects carried out, involving multiple instance variables, function calls, etc.
I can also nest templates within individual responses, so I can create a branching structure. IE, a template might contain two different choices, with each choice having another, unique template linked to it. Only the template linked to the choice that the user has made will be loaded. This is something that, as far as I am aware, AIML can’t do, AIML essentially just searches through its entire knowledge base for a match for each individual input.
The AI also have a limited ability to speak out of turn, and this could possibly be extended to allow the AI to change the subject, in certain situations.
This system is a very simple proof of concept for now, but it seems to have worked out easier than I thought.
