Em Tsifrova Assistant

Bank bot of Em — the digital assistant whose cornerstone the algorithms which are responsible for selection of the answers to the message of the client which are most suitable on sense are.

2018

The basic version of a virtual assistant of Em on the basis of a neuronet

On October 10, 2018 Modulbank reported that together with "Consolidation of the Cognitive Associative Systems (CCAS)" started in June, 2018 the basic version of a virtual assistant of Em on the basis of a neural network dialogue system. Em communicates with clients in chats, she is capable to give answers to questions of clients with a high accuracy, to execute their instructions formulated in a free form in Russian.

For October, 2018 Em solved about 17,000 requests in a support service. She accepts every fourth question from clients, solves a half of them itself, in other cases calls to the aid colleagues.

Em works at ensemble of the neuronets of different type which are carrying out different tasks. At first the expressions arriving from clients undergo preprocessing on the basis of neural networks and the systems of machine learning. They filter the expressions which are not relating to target subject, isolate a slang, correct the allowed typos. The basis of a bot is formed by ensemble of a large number konvolyutsionny, recurrent, LSTM and other neural networks which process different signs of the expression and jointly define the correct final answer.

Em has two brains. The second neural network ensemble undergoes daily training at a basis of again arriving expressions of users. If it shows the best results after training, than the existing neural network ensemble, then it is allowed before work with clients. For training of neuronets and submission of the semantic contents of the text the multidimensional vector spaces defining ratios between characters, morphemes, words and phrases were constructed previously.

All these difficult algorithms live on the graphic supercomputer with a performance of 42 TFLOPS so our machine can communicate along with thousands of clients through their personal accounts in Modulbank.

For October, 2018 the machine can answer with a high accuracy 261 types of frequently asked questions, give answers to questions of personal character and solve private problems. For example, she will quickly find out a situation with the sent or expected payment, will ask again the necessary information if the client forgot it to specify. She remembers the customer interaction history, considers a dialog context. It is steady against typos, in communication with it it is possible to use different words and grammatical constructions. It eliminates trolling, any nonsense and so far eliminates an offtopic.

The OKAS commands in the field of neural networks provided to Razrabotoki the accuracy of answers of a system with the quantity of errors which is not exceeding 0.9% of total number of the asked questions already at the initial stage of system implementation and permanent improvement of these results during permanent after-training of neuronets. If a system knows the answer to the corresponding type of questions, kind of they were formulated, then a system will answer correctly in 97.2% of cases. If the network has low confidence in correctness of the answer, then it will ask to answer the specialist of bank.

After implementation of the provided version of a system the percent of a scope of answers to questions of clients of Modulbank rose from 9% to 40% of all customer appeals in bank. OKAS and Modulbank continue to work on implementation of cognitive functions of a dialogue system.

Announcement

Modulbank announced on March 1, 2018 development of the digital assistant Em who not just reacts to a key word, and understands sense of phrases and can conduct dialogue, without losing conversation thread. Bot knows answers to the most popular questions on work of bank, at the same time is capable to study as Siri or Alice.

It is expected that the bank bot is designed to lower load of business assistants and to get rid of need to constantly increase their staff because of growth of number of clients (only for the last year their quantity in bank increased by 3 times). For its creation using neural networks dialogs with clients for all history of bank were grouped. At the same time developers had to face a number of difficulties — typos because of what they had to be sorted manually occurred in messages of clients from old dialogs. Besides, in four years answers of support changed that too created difficulties when grouping, told in bank.

We tried different approaches, including, trained Em at dialogs from series — Andrey Varikov, the director of development center of Modulbank shared. — In a case with series a problem that bank language is very specific. Though we also communicate with clients a simple language very few people in a daily conversation discuss deposit interests, changes in rates or the currency transfers commission.

As of March 1, Em already began to communicate with clients of Modulbank. They can estimate her answer, having delivered "like" or "dislike". If Em is mistaken and receives dislike, she transfers the client to the real person from the necessary division at once. In addition, Em prompts answers to specialists of support in a chat, and, thus, saves time for consideration of a question of the client.

According to Modulbank, algorithms of training of a bot are constantly optimized. All dislikes which are received by Em are analyzed to improve her work. Initially, if the answer received many dislikes, it was disconnected. However, if the client has a serious problem, for example, blocking of the account, most often it puts dislike even at the correct answer. Therefore the training algorithm as a result was changed, explained in the company.

While Em accepts about 10% of all addresses and in half of cases solves the client's problem independently, without involving the specialist. The bank predicts that Em will be able to undertake 30% of client addresses already soon and will help to lower load of business assistants. She will independently solve a number of problems, for example, to open accounts or deposits to clients. In the long term, according to plans of Modulbank, Em will be able to accept up to 70% of all requests that will allow to optimize work of specialists and will accelerate the answer to the client.