The different levels of sophistication of chatbots or conversational assistants

by Tristan Thillaye Du Boullay - Senior Digital Consultant | minutes read

The first and last rule in business is to know your customer. And the most effective way of achieving that primary business goal is, even in our digital age, by having a real-life conversation. Today’s chatbot technology can help you initiate and maintain that all-important customer conversation in an online environment too. However, as with any promising new technology, there's a real risk of overselling what chatbots can actually do and deliver to your business. Time to set the record straight!

Purely functionally speaking, a chatbot or conversational assistant is a type of user interface (UI) designed to deliver a richer user experience (UX) when interacting with a computer, so that it feels more like a human-to-human than a human-to-computer interface for the end-user. Technically speaking, a chatbot is a computer programme that tries to understand human conversation via Artificial Intelligence (AI) technologies (or not necessarily) such as Natural Language Processing & Understanding (NLP & NLU) and then simulates natural human answers via text or text-to-speech technology. Because chatbots are by their very nature conversational, they are often referred to as conversational agents.

Non-AI versus AI-powered

While all chatbots are conversational in their nature, they nevertheless come in different shapes and sizes. Or rather: they come with different levels of sophistication, ranging from the static, so-called symbolic or rule-based to the more dynamic AI-powered chatbot. Those are in fact the two main types of chatbots we at Sopra Steria currently distinguish and offer to our customers.

The first type is a symbolic robot, in that it automates simple, repetitive conversations that require a deterministic answer, like a yes or no, with little to no interpretation involved. Although a rule-based chatbot uses Natural Language Processing (NLP), enabling it to understand the meaning of human language, interactions with this type tend to be quite specific and highly structured, like a flowchart or decision tree. Its capabilities are fairly basic and limited, to the extent that it effectively resembles an interactive FAQ. Since chatbots based on this purely deterministic model require pre-set rules and scripted responses, inducing a strong labour-intensive approach, they can be rather rigid and slow to develop. Some Robots that do not use Artificial Intelligence technologies work pretty well when it is about interacting with users that always use the same vocabulary and do not need to evolve a lot. The limitation becomes quickly visible when the robot has to learn from its users and adjust his answers accordingly.

Chatbots powered by artificial intelligence (AI) technologies or, more particularly, by machine learning (ML) are more complex than their counterparts that don’t. They can manage complex tree conversations, interconnect different conversation scenarios and interact with other information systems. They also tend to be more conversational, data-driven and predictive. Not only can they understand what users are saying, but they can also grasp their preferences or intentions and ask them for more clarification. Moreover, supervised AI technologies allows the bot to evolve and learn from the way users interact with him. The more users chat with the virtual assistant the more accurate it will become and the better it will understand them. Context management is strengthened too, enabling bots to respond within the context of the conversation.

Only supervised machine learning (for now)

At Sopra Steria, we have been successfully building symbolic chatbots for our customers for several years now. Our customers can also count on us for developing AI-powered chatbots, but –at the moment- only in a supervised learning mode, where the machine already knows the output of the algorithm before it starts working with it or learning from it. You can read more about all the different machine learning techniques here. But basically the difference between a supervised and an unsupervised approach boils down to whether or not human interaction is required. Unsupervised machine learning allows systems to learn automatically and improve on their own, without human intervention. But since that method by definition allows a certain degree of probability or uncertainty, it is bound to produce errors. We’ve had the perfect example with Tay chatbot developed by Microsoft in 2016 that had to be retired after less than a day because its unsupervised machine learning algorithm became racist. Tay used to learn with no supervision from the conversation she had with humans on twitter and humans quickly used this capacity to alter its knowledge not in a very ethic way.  

Therefore, until the technology becomes more mature and reliable, we only use it for R&D purposes or in very specific use cases. At Sopra Steria we want to deliver to our clients the best of AI technologies without its drawbacks.

In my next blog post I will present an innovative new chatbot concept lurking on the not-so-distant horizon: the master conversational agent or Master Bot.