Promobot, a resident of the Information Technologies Cluster of the Skolkovo Foundation, has developed a dialogue system based on neural networks for its service robots. However, she does not need an Internet connection.
Prior to this, Promobot robots communicated using a linguistic base - that is, a question-answer system. Now they will be able to freely recognize requests, generalize them on a given topic, and generate responses using neural networks.
The new dialogue system allows you to compare the voice request of the interlocutor with the location of the robot, the time of the request, as well as with the context of the previous dialogue. Promobot will be able to conduct long dialogues, and not only respond to single requests, as existing systems do.
Andrey Nosov, head of the dialogue system development at Promobot: “A robot, unlike voice assistants, will always focus on the time and place of the conversation in answering a question. This is a critical task for service robots, since they provide a service at a specific place, say, an airport, and must always be “aware” of the context in which they are. ”
10 key topics for business: for example, transport (How to get there? How to buy a ticket? Where is it?) Or building (What address? How to get a pass? In which office?)
To train one intent (user need), you need at least two thousand relevant samples (specific examples of a request in context). To train 10 topics, data engineers had to analyze about 6.5 million query lines. This took 1000 business days.
To improve accuracy, the dialog is divided into five modules.
• The first relies on simple communication that does not require an attachment to any goals. Its purpose is to soften or maintain dialogue.
• The second module responds to business inquiries, processing them according to the required context.
• The third is used as a search engine - to constantly update the robot's database.
• The fourth module of the system is directly related to the third. He plans to compare the facts taken from the search engine in cases where you need to get some kind of comparison or unloading that is important for decision making. For example, to get answers to the questions: Where is it cheaper?, Who is higher?, Where is it better to go? etc. In this case, the system does not need to have access to the Internet at the time of the request: the databases are constantly updated while the robot is online, but the robot does not lose access to them if the Internet is disconnected.
• The last, fifth, module connects to the Promobot linguistic base, which the robot's creators have been collecting for five years.
Oleg Kivokurtsev, Development Director of Promobot: “All the voice assistants that exist now are, for us, essentially just one of the system modules, which is responsible only for chit-chat. Any voice assistant needs the Internet to work properly. Our solution can also function offline, as it focuses on specific business tasks.”
The implementation of the dialogue system on the existing Promobot robots will begin in December 2020. There are about four hundred of them now working in our country: in the Museum of Contemporary History of Russia, branches of the MFC, Sberbank, shopping and business centers, clinics and airports.
Pavel Krivozubov, Head of Robotics and Artificial Intelligence, IT Cluster, Skolkovo Foundation: “Russian solutions in the field of service robotics are in steady demand on the world market. This happens, in part, because our companies in some aspects are the legislators and founders of new technological and business trends in this area. Promobot is the leader in the field of service robots in Russia, introducing new algorithms and technological solutions, which will further strengthen the position of our country and our technologies in this competitive market.”