Many people think that an AI "chatbot" is something that the more you use it, the more automatic and smarter the AI itself becomes.
Chatbots in general are tools that combine "chat" (= conversation) and "bot" (= automation, robot), and are programs that automatically respond to a person's voice or input text.
Chatbots can indeed, in theory, automatically become more accurate through conversations with users and improve the rate of correct responses.
Chatbots are broadly classified into two types: AI-type and scenario-type. Here, we omit the differences between AI-type and scenario-type programs, the merits of each, the purpose of introducing them, and the man-hours required for development.
AI chatbots don't get smarter automatically.
Chatbots can be roughly defined as systems that can interact like people. A typical implementation is one that performs natural language processing on user questions and messages, determines the intent of the user who sent the message, and processes and responds according to that intent and context.
However, the reality is that AI chatbots do not become automatically smarter in the short term, and there are two reasons for this.
The first reason is that most user feedback results are not useful data for learning. In order for AI to learn in the first place, someone needs to string together the correct data.
Specifically, this means that someone needs to tie the question "my PC has stopped booting" to the question "my computer won't turn on" as meaning the same thing.
Feedback on it is actually given via a button that asks if it was helpful after the chatbot's response. In other words, even AI that appears to be automatically trained is fundamentally driven by human manual feedback.
This means that it is unclear whether the user will necessarily give the correct answer to a question, and the user's use of the chatbot alone will not lead to a deeper understanding of the AI.
The second reason is that AI chatbots with poor accuracy will not be used enough to accumulate data.
The amount of data, or "example sentences," that needs to be accumulated before the AI's correct response rate exceeds about 70% is expected to be substantial. For example, assuming there are 300 answers to a question, at least 100 similar questions need to be associated with each answer.
To take the previous example, for the question "My PC stopped booting up," it is necessary to link multiple similar questions as shown below to learn example sentences.
1. My computer does not turn on.
2. My laptop is not working properly.
3. My computer stays frozen with a black screen.
4. Blue English is displayed on the desktop screen and I cannot proceed.
Moreover, since this is done for each of the 300 FAQs, at least about 30,000 question sentences must be prepared and manually tied together in this example. Thus, it takes a lot of hard work by the developer to improve the AI's understanding by having it learn the example sentences.
In view of the above, in many cases, a sufficient amount of data is not collected in the first place, as the AI is no longer used before it has progressed in learning, which requires a sufficient amount of data, or when it returns a misguided answer.
There is a limit to the pattern of chat responses that can be completed by AI alone, which is why the introduction of AI chatbots requires special ingenuity and human intervention.
What kind of AI is used in chatbots and how is it used? We provide information and opportunities to learn what kind of AI is used in chatbots and what kind of conversational databases are available.
If you are a company that is thinking about implementing a chatbot, or considering changing your current chatbot, this may be of interest to you.