This title is the summary of what is happening in the market today, mostly encouraged by Facebook’s move for Messenger bots.
The ChatbotConf 2017 revealed this sad truth. There are 200,000 Messenger bots today, most likely none of them have a real AI backbone. A recent article summarizing the conference draws a similar conclusion.
End users of chatbots would not really care whether there is AI backbone or not if the chatbot they are using solves their problem. In a small fraction of cases, chatbots without AI can be helpful, especially in e-commerce transactions where buying and selling options are rudimentary, and the conversations can be buttonized. However, the AI issue surfaces when chatbots try to service higher complexity tasks. The way chatbots can be used in real life, this corresponds to, maybe, 90% of the cases. So, what is the AI backbone that is required?
The AI Backbone
Chatbots that represent AI must have some (if not all) of the capabilities listed below:
- NLP: Capability to understand users’ responses in their most variant form.
- Answering Questions: Ability to communicate with the user about a subject matter by absorbing knowledge and answering questions about it.
- Asking Questions: Ability to ask questions to navigate the user to solve a problem.
- Dialogue Behavior: Ability to engage users in certain behavior in concert with the chatbot’s objective (sales, transactions, advice, training, story telling, idea sharing, etc.)
- Learning from Conversations: Ability to ask users for answers and to learn from them. This should be optional since social input may not be desirable for certain objectives.
- Short-term Memory: Ability to remember the topic of conversation and interpret pronouns correctly. This requires chatbot to take into account what was said 2, 3, or 4 steps earlier.
- Long-term Memory: Remembering previous chat sessions and starting conversation from where it was left of.
- Emotions and Attitude: Ability to detect unproductive conversations, change strategy, or abort not to waste resources.
- Awareness: Ability to self-assess its performance, produce reports about its performance, and suggest bot builders the weaknesses encountered.
- Infinite Speech: Not to be restricted by a pre-defined steps of conversation.
Canning Responses Instance-by-Instance is not AI
Most chatbot platforms today are requiring instance-by-instance input from its builder to develop every step of the intended conversation in a rigid sequence. This approach is feasible for banking transactions, travel bookings, or other similar interactions where dialogue is restricted to solid options. Obviously, there is no AI backbone needed for such chatbots.
Chatbot science is at its infancy while most developers are expecting adult behavior.
Deep Learning is not a Silver Bullet
One of the latest misconceptions emerged in the market is that if there is enough data thrown at deep learning system, all the requirements listed above as AI backbone can be satisfied. Deep learning can only handle some parts of the required list, and the rest must be called the “chatbot science”. The only way to produce a chatbot development platform in the scope of AI backbone is to offer data-driven tools and/or knowledge-driven tools with certain level of built-in functions, where those functions define the secrets of the chatbot science.
——- FOLLOW OUR LAB ———-
Talk to Vera, exclone’s company representitive.
For exClone’s Chatbot Platform, click here for free trial via LinkedIn access.
Join our CHATBOTS linkedin group
#chatbot #chatbots #AI #artificialintelligence #ConversationalAI #Virtualassistants #bots #machinelearning #NLP #DL #deeplearning