Since machines have been created, humans have dreamt to communicate with them. For more than 60 years, research in Artificial Intelligence worked towards that goal. If there are already programming languages that allow some sort of human/machine exchanges, they tend to remain particularly artificial. Researchers want to make communications between humans and machines more natural.

Natural Language Processing is the research field that focuses on the understanding, the treatment and the generation of natural language by machines in order to create relevant and natural interactions with them.

Natural Language Understanding is a recent branch of NLP. As the name implies, NLU specifically focuses on the way the written language is understood by machines: they need to understand what the sentence “I killed the spider in my pyjamas” means even though it can be ambiguous: “I killed the spider while wearing pyjamas” or “I killed the spider that was inside my pyjamas”. NLU also focuses on tasks such as sentiment analysis, automatic text summarization, question answering systems or also conversational robots.

Nowadays, two kinds of chatbot are used: the first one functions with commands, that is to say predefined questions (textual interaction bots); the other kind tries to imitate natural language by analysing the speaker’s words. The second kind is generally called a conversational agent. Natural language is the feature that will make a difference between a regular question answering chatbot and a conversational agent.

In order to imitate human natural language, machines need to understand and learn how to use intention and emotion markers such a humour, pauses, discourse markers or even emoticons. They also need to share some sociolinguistic features of the human speaker in order to stay close to him: especially with language registers.

Unlike traditional chatbot, the strength of a conversational agent is the ability to contextualise dialogue to understand coreferences and avoid this kind of dialogue:

Human: “I ate pasta”

Machine: “Good appetite”

Human: “Do you want some?”

Machine: “I enjoy speaking with you”

It is obvious in this dialogue that “some” in the second human intervention is not understood by the machine because the context and the coreference were not taken into account. This is the main issue of conversational agents: they are not able to consider the conversation’s context or what has been said before. Thus, the longer the conversation is, the more difficult it will be for the chatbot to understand and discuss.

However, there are systems that steer up the conversation towards the subjects that can be handled by the bot. It can be a takeover: the conversational agent delegates the conversation to a human advisor or tells the user that it is limited in its answers. But it can also be a simple reminder of the original question. Thus, from the moment when the confidence index of the agent is not sufficient, it can ask for a clearer answer without overthrowing the user experience. In the case of an agent specialised in booking train tickets, for example, the request “Can you please reformulate your answer” is less engaging and natural than “I haven’t quite learned that yet. Do you still want to go to Roma?”