To date, the number of existing chatbot is large enough and continues to grow actively. They are used in various fields as entertainment helpers. programs, tools for searching and processing information, as well as managing various processes. One of the prospects for their use are the processing modules of the natural language in artificial intelligence systems and speech robot subsystems. However, at the moment most of the chatbots available are based on templates, i.e. limited in the set of processed replicas. A fairly small number of similar programs to date. able to process a variety of phrases and generate various unique replicas in response. Therefore, this development sector is still promising and open to new ideas. Based on the foregoing the relevance of the master's work is to create a chatbot that will be able to come close to the level of a human interlocutor in the breadth of coverage of processed and generated replicas for further use this system in the fields of artificial intelligence and robotics.

The purpose of this study is to select the most appropriate architecture for the subsequent implementation of an ontological chat bot, which will be able to imitate a human interlocutor in a natural way.

To achieve this goal it is necessary to perform the following tasks:

• Analyze the functional features of the chatbots.
• Investigate the basic structural elements and architectures of neural networks.
• Investigate existing methods of teaching neural networks.
• Analyze the information received and determine the most appropriate architecture and method of training the neural network to achieve its goal.

Thus, as a result of the work, the neural network architecture will be chosen, on the basis of which the ontological chatbot will be implemented. It will allow you to fully realize the desired result. and to achieve the most complete breadth of replicas and words processed by the system. It will also determine the most appropriate method of training the neural network, which will provide high quality replica perception and correctness of the generated responses.

Chatbot is a program that can communicate with users, imitating the behavior of a person, to achieve any goal or entertainment. This is usually done through any exchange platform. messaging, such as Facebook Messenger, Slack, Telegram or Viber. Chatbots can be relatively simple rules-based programs, or they can use artificial intelligence, which makes them much more sophisticated, but suitable for a wide range of tasks. Such systems perfectly understand the user, remember the information and allow you to perform a number of actions: pay bills, learn about interesting events, order a taxi, book a room, forward funds, etc.

There are various approaches to classifying chatbots. The most popular of them are the classification according to the principle of operation and scope. According to the principle of operation, there are two types of chatbot.

• Template based chatbot.

Such systems when communicating with the user select keywords and respond to them. The developer at the same time pre-registered commands for each word or phrase. If the user does not use during correspondence keywords, the bot just does not understand it. Naturally, the functionality of such bots is very limited. However, in some cases they are quite effective.

• Trained chatbots based on neural networks.

Such systems are developed on the basis of modern solutions in the field of artificial intelligence. They do not use templates for answers, and are able to process natural speech, correctly answering the questions posed. In addition, the chatbot analyzes the dialogue to further improve their communication skills. In other words, he is able to predict what the user expects before he asks the question.

Currently chatbots are used in a large number of spheres of human life. Let's consider their main types according to the scope.

• Chatbots—helpers.

These systems allow you to automate some simple human actions in his daily life. For example, most of these chatbot allow you to set a schedule according to which the user will provide certain information by sending it to any instant messengers or displaying on a smartphone screen. Such information may be a weather forecast, exchange rates, news, etc.

• Entertainment chatbots.

Such systems are directed, respectively, to the user's entertainment. These can be chatbots based on templates, such as text games, or complex systems created on the basis of neural networks. and artificial intelligence technologies that allow the user to communicate.

• Business chatbots.

They can be integrated into various software for organizations designed to automate strategies for interacting with customers (clients). Also, business chatbots can be part of systems that are able to perform transactions, such as remittances.

In addition to the above classifications, among chatbot, which are based on neural networks and are aimed at maintaining a dialogue with users, there are also several additional types. Consider them in more detail.

• Highly specialized chatbots.

Such systems have a small knowledge base and are aimed at working within the framework of a single knowledge base of the industry.

• Widely specialized chat bots.

These systems have an average knowledge base and can be used in several industries and areas of activity.

• Ontological chatbots.

These systems have an extensive knowledge base and ample opportunities to learn and supplement the database with new knowledge gained through conversations with users. Ontological systems are able to maintain a conversation with the user on various themes, imitating the person of the interlocutor in the best way.

Despite the fact that there are bots for a long time, they gained popularity relatively recently. Not a small role in this was played by the successful use of bots in marketing. Multinational companies for a long time uses chatbots to simplify the purchase process. With their help, each client receives a timely response from the company and, which is very pleasing, forgets about the existence of queues forever.

The positive quality of chatbots is their availability. No need to be a huge company to have your chatbot. Even a small business can afford to buy such an assistant.

One of the main ways to use a rule-based chatbot is to work with clients and support. In other words, they are used to ask customers a series of questions, thereby leading to resolve their problems.

Chatbots created using artificial intelligence offer more options, including another additional way to interact with customers. In fact, they can be used. to create direct conversions and act as a digital version of the seller. An example would be a consumer who wants to buy a new pair of shoes. This consumer can go to the website of the retailer, view the range offered goods and then decide. If the seller has a suitable chatbot, the customer can send him a message and say something like: «e;I would like to buy a pair of shoes.»e; Then the chatbot starts a conversation with the client to find out what type of shoe he would like before presenting him with suitable options.

Basically, the popularity of chatbots is due to the growing use of instant messengers. People use them all over the world, often as the primary means of communication. At the same time, people are less likely to download applications on their devices, especially if they are not necessary. For example, in the US, the user installs the application once every two months. Remember yourself when was the last time you installed something on a smartphone or tablet? Development Mobile Application — The process is long and expensive. For example, a private cosmetologist or massage therapist does not make sense to develop it in order to automate the process of communicating with clients. But chatbots give to individuals, companies and brands are an effective way to communicate with users on platforms that they like and want to use, including instant messengers.

The ongoing process of technology development gives the right to say that in the near future bot will appear in all branches of human activity assistants endowed with artificial intelligence and opportunities that will become real thanks to neural networks.

Neural network — mathematical model, as well as its software or hardware implementation, built on the principle of organization and functioning of biological neural networks — neural networks cells of a living organism. This concept arose when studying the processes occurring in the brain, and when trying to simulate these processes. A neural network is a system of simple processors interconnected and interacting with each other. (artificial neurons). Such processors are usually quite simple (especially in comparison with processors used in personal computers). Each processor of this network deals only with signals that it periodically receives, and signals that it periodically sends to other processors. And, nevertheless, being connected to a fairly large network with controlled interaction, such separately simple processors together are capable of performing rather complex tasks [2].

Neural networks are not programmed in the usual sense of the word, they are trained. Learning Opportunity — One of the main advantages of neural networks over traditional algorithms. Technically training is in finding the coefficients of connections between neurons. In the process of learning, a neural network is able to detect complex dependencies between input data and output data, as well as perform a generalization. This means that if successful, the network will be able to return correct result based on data that was missing from the training set, as well as incomplete and / or «noisy», partially distorted data.

Neural networks are used to solve complex problems that require analytical calculations similar to those of the human brain.

There are various approaches to the classification of neural networks. Let us consider in more detail the main ones. One of the existing classifications of neural networks — on the direction of communication.

• Networks without feedback.

The advantages of networks without feedbacks are the simplicity of their implementation and the guaranteed receipt of a response after passing the data across the layers. The disadvantage of this type of network is minimization of network size — neurons are involved in data processing. Smaller network facilitates learning [2].

• Networks with feedback.

The advantages of networks with feedback is the complexity of learning, caused by a large number of neurons for algorithms of the same level of complexity. Disadvantages of this type of network — require special conditions guaranteeing the convergence of calculations [2].

Another classification of neural networks was determined based on the mechanisms of their functioning and the nature of connections. It includes four main types.

• Direct distribution networks.

All connections are directed strictly from the input neurons to the weekend. Examples of such networks are the Rosenblatt perceptron, the multilayer perceptron, the Vord networks [5].

• Recurrent networks.

The signal from the output neurons or neurons of the hidden layer is partially transmitted back to the inputs of the neurons of the input layer (feedback). Hopfield recurrent network «filters» input data, returning to sustainable state and, thus, allows to solve problems of data compression and construction of associative memory. A special case of recurrent networks are bidirectional networks. In such networks, there are connections between layers as in the direction from input layer to the output, and in reverse. A classic example is the Kosco neural network [12].

• Radial basis functions.

Artificial neural networks have been developed that use radial–base functions (also called RBF–networks) as activation functions. General view of the radial–base function is shown in the figure 4.1.



Figure 4.1 — General view of the radial–base function

In the presented function, x — vector of input signals of the neuron, σ — width of function window, φ — decreasing function (most often, equal to zero outside a certain segment). The radial–basic network is characterized by several features. It has a single hidden layer and only its neurons have a non-linear activation function. Also, the synaptic weights of the connections of the input and hidden layers are equal to one. [3].

• Self-organizing maps.

Such networks are a non-teacher, competitive neural network that performs the task of visualization and clustering. Is a method of projecting multidimensional space into a space with a lower dimension (most often, two-dimensional), it is also used for solving problems of modeling, forecasting, etc. It is one of the versions of Kohonen's neural networks. Kohonen's self-organizing maps are primarily used for visualization and initial (»reconnaissance») data analysis.

The signal in the Kohonen's network goes immediately to all neurons, the weights of the corresponding synapses are interpreted as coordinates of the node, and the output signal is formed according to the principle of «the winner takes everything» — that is, a nonzero output signal has the neuron nearest (in the sense of weights of synapses) to the object fed to the input. In the process of learning, the synapse weights are adjusted so that the lattice nodes «are located» in places of local data thickening, that is, they described the cluster structure of a data cloud, on the other hand, the connections between neurons correspond to the neighborhood relations between the corresponding clusters in the feature space [3].

Also, neural networks can be classified by the nature of training into two main types.

• Networks trained with the teacher.

When training with a teacher, each training input example requires knowledge of the correct answer or the function of assessing the quality of the answer. Such training is called managed. Neural network values are presented input and output signals, and it is on a certain the algorithm adjusts the weights of synaptic connections. In the course of training, the weights of the network are corrected by comparing the actual output values with the input ones known in advance. [11].

• Networks trained without a teacher.

When learning without a teacher, the internal structure of the data or the correlation between the samples in the data set is revealed. The outputs of the neural network are formed independently, and the weights vary according to an algorithm that takes into account only the input and their derived signals. it learning is also called unmanageable. As a result of such training, objects or examples are divided into categories, the categories themselves and their number may not be known in advance [11].

Consider the basic structural elements that are typical for all types of neural networks. Figure 4.2 presents a general schematic representation of a neural network.

Neuron is a computational unit that receives information, performs simple calculations on it, and passes it on. They are divided into three main types: input (blue), hidden (red) and output (green) (see. Fig. 4.2). In that case, when a neural network consists of a large number of neurons, the term layer is introduced. Accordingly, there is an input layer that receives information, n hidden layers (usually no more than 3 of them) that process it and an output layer that displays the result. Each of the neurons has 2 main parameters: input data (input data) and output data (output data). In the case of an input neuron: input = output. In the others, the total information of all neurons from the previous layer gets into the input field, after which it is normalized using the activation function and enters the output field. It is important to remember that neurons operate with numbers in the range [0,1] or [-1,1]. Therefore, for example, for processing numbers outside these ranges, the simplest solution is to divide 1 by this number. This process is called normalization, and it is very often used in neural networks [3].

A synapse is the connection between two neurons. Synapses have a weight. Thanks to him, the input information changes when transmitted from one neuron to another. Suppose there are 3 neurons that transmit information to the following. Then we have 3 weights, corresponding to each of these neurons. The neuron that has more weight will have the information that will dominate the next neuron (example — color mixing, shown in Figure 4.3). In fact, a set of weights of a neural network or a matrix of weights it is peculiar the brain of the whole system. Thanks to these weights, input information is processed and turned into a result. It is important to remember that during the initialization of the neural network, the weights are placed randomly [3].

Consider also a number of basic concepts related to the operation and preparation of neural networks to perform the tasks.

Activation feature is a way to normalize the input data. That is, if at the input we have a large number, by passing it through the activation function, we will get an output in the desired range. There are a lot of activation functions: linear, sigmoid, hyperbolic tangent and so on. Their main difference is a range of values. The linear function is almost never used, unless it is necessary to test the neural network or transmit the value without transformations. Sigmoid is the most common activation function, its range of values varies from 0 to 1. Most of the examples are implemented on it, it is also sometimes called the logistic function. Accordingly, if there are negative values ​​among the source data (for example, stocks can go not only up but also down), then a function that captures and negative values. Use the hyperbolic tangent should be when the source data can be both negative and positive, since the range of the function varies from -1 to 1 [3].

Training set is a sequence of data with which the neural network operates during the execution of its learning procedure.

Iteration is a kind of counter that increases each time a neural network passes one training set. In other words, this is the total number of training sets traversed by the neural network.

Epoch is the value that is set to 0 when the neural network is initialized and has a manual ceiling. The greater the era, the better the network is trained and, accordingly, its result. The age increases each time we go through the whole set training sets. It is important not to confuse the iteration with the epoch and understand the sequence of their increment. First, the iteration increases n times, and then the epoch and not vice versa. In other words, you cannot first train a neural network only on one set, then on another, etc. Need to train every Set once per era. So you can avoid computation errors.

Error is a percentage value that reflects the discrepancy between the expected and received responses. A mistake is formed every epoch and should go into decline. If this does not happen, then the training goes wrong. Error can be calculated in different ways, but there are Three main ways: Mean Squared Error (hereinafter MSE), Root MSE and Arctan. The principle of error counting is the same in all cases. For each set is considered an error subtracting from the ideal answer received. Further, either it is squared, or the square tangent from this difference is calculated, after which the resulting the number is divided by the number of sets [3].

Neural networks require training, otherwise the correct result is unlikely to be obtained. There are many different methods of teaching neural networks. Among them there are three most interesting ways, these are the methods of reverse and elastic propagation, as well as genetic analysis. Consider them in detail.

• Back Propagation Method (MPA)..

This method uses a gradient descent algorithm. This algorithm is a method for finding a local minimum or maximum of a function using motion along a gradient. Construct a graph, where the x-axis will be the weight values of the neuron (w) and the y-axis — the error corresponds to this weight(e) (fig. 5.1).



Figure 5.1 — Gradient descent graph

Based on the graph shown in figure 5.1, the graph of the ƒ(w) function is the dependence of the error on the selected weight. Note the global minimum on the chart point (w2, e2) or, in other words, the place where the graph comes closest to the x-axis. This point will mean that choosing the weight w2 we get the smallest error — e2 and as a result, the best result of all. The gradient descent method allows finding this point (fig. 5.1, the ƒ’ function). Accordingly, each weight in the neural network will have its own graph and gradient and everyone needs to find a global minimum [4].

Thus, the gradient itself is a vector value that determines the direction and steepness of the slope. The gradient is found using the derivative of the function at the desired point. Caught at a certain point with a weight value that is distributed in a random order, the gradient is calculated and the direction of the descent is determined, and so on at each of the following points, until a local minimum is reached that prevents further descent. To cope with this trouble, you need to set the desired value of the moment, which will allow to overcome part graphics and reach the desired point. If this value is insufficient, then the bulge cannot be overcome, in the case of setting a higher value than you need to have a chance to slip the global minimum. In addition to the moment of acceleration, there is also a concept that defines the overall speed of network learning. This value, like the previous one, is a hyperparameter and is selected by trial and error. The best option is never known in advance; you can recognize it only by spending a few trainings and adjusting the value each time in the right direction. Now you can go directly to the method learning and determine what he is. In the process of receiving information, the neural network sequentially transmits it from one neuron to another through synapses, until such time as the information is on the output layer and is not issued as a result. This method is called transfer. forward. After the result is obtained, the error is calculated and on its basis we perform the postback. The essence of which is consistently change the weight of the synapses starting with the output and moving towards the input layer. At the same time, the weight value changes to the best result To use this method Learning fit only those activation functions that can be differentiated [4]. The training scheme using the method of reverse propagation is presented in figure 5.2.



Figure 5.2 — Diagram of the method of back propagation (animation: 19 frames, 8 cycles of repetition, 194 kilobytes)

In order to successfully complete the training, it is necessary to extend the resulting error to the entire weight of the network. By calculating the error at the output level, and there you can calculate the delta, which will be sequentially transmitted between neurons. Then you need to calculate gradient for each outgoing connection. Then, having all the necessary data, it is necessary to update the scales and calculate the value due to the MPA function, which will become the magnitude of the change. You should not forget about the time and speed of learning. One iteration of the MPA gives a small percentage of error reduction, therefore, they need to be repeated again and again until the error rate is close to 0 [11].

• Method of elastic distribution.

The following network training method is called Resilient propagation or Rprop. The method of reverse propagation described above has the disadvantage of spending time on the learning process inappropriate if it is necessary to get a quick result. To speed up process was proposed many additional algorithms that accelerate the process. One of which is the current method. The present algorithm uses epoch-based training as a basis and uses only the signs of the derivatives of a particular case to adjust the weights. Used a certain the rule by which the calculation of the weighting factor is calculated. If at this stage of calculations the derivative changes its sign, it means that the change was too large and the local minimum was missed and you need to roll back, that is, return the weight to the opposite position, and reduce the magnitude of the change. If the sign of the derivative has not changed, the magnitude of the change in weight, on the contrary, increases for greater convergence. If the basic parameters of the weight correction are fixed, then the settings of the global parameters can be avoided. And this will be another advantage of the current method over the previous one. For these options there are recommended values, however, there are no restrictions on their choice.

So that the weight does not assume too large or small values, the established limits for the correction values ​​are used. The correction value is also calculated according to a certain rule. That is, if the derivative of a function at a particular point changes sign from plus to minus, it means that the error increases and the weight requires correction and its reduction occurs; otherwise, increase. The sequence of actions in this case is as follows: the correction value is initialized, partial derivatives are calculated, the new correction value of the weights is calculated, corrected weights, if the condition of stopping the algorithm is not satisfied, the process proceeds to the calculation of derivatives and repeats the cycle again. Such an approach makes it possible to achieve convergence of a neural network several times faster, unlike the previous training variant [4].

• Genetic algorithm.

It is a simplified interpretation of a natural algorithm based on the crossing of results. That is, in essence, there is a crossing of the results, the selection of the best and the formation of a new generation on the basis of them. If the result does not suit the algorithm repeats until the generation becomes perfect. The algorithm can be completed without achieving the desired result if the number of attempts is exhausted or the time for a mutation is exhausted. This algorithm is applicable to the process of optimizing the weight of the neural network, with the default topology. With this weight is encoded with a binary code and each result is determined by the full weight gain. Quality assessment occurs by calculating the output error [10].

Deep Learning (DL) is the statistical technique for classifying patterns based on test data using multilayer neural networks. Neural networks that are taught by the depth method usually consist of a set of input modules that accept data such as pixels or words, sets of hidden layers (the more layers, the deeper the network) containing hidden modules (also known as nodes or neurons), and a set of output modules, taking into account the presence of connections between different nodes [1]. The scheme of such a neural network is shown in figure 5.3.

In the general case, the network shown in figure 5.3 can be trained, for example, on a large set of handwritten numbers (these are input data in the form of images) and tags (output data) that define the categories to which the input data belong (this is an image 2, this is 3, and so on). Over time, the error propagation algorithm allows the gradient descent to correct connections between the modules so that any input data leads to the output of the corresponding output data.

In general, the relationship between input and output, which the network studies, can be thought of as building a map of correspondence. Neural networks, especially those with many hidden layers, very well cope with the construction of the correspondence between input and output. These systems usually describe as neural networks, since input modules, hidden modules and output modules can be thought of as approximate models of biological neurons, albeit extremely simplified, and the communication between modules in some way represents the connections between neurons. A long-standing question outside of this the work concerns the degree of biological reliability of such artificial neural networks [1].

Most of the DL–network uses a technology called «convolution» that places such restrictions on network connections that they naturally have such a property as translational invariance. Essentially, this idea is that an object can move in the image, without losing their identity; the circle in the upper left corner can be considered the same object as the circle in the lower right corner (even without direct evidence of this).

In-depth training is also known for the ability to independently build intermediate representations — for example, internal modules that respond to such things as horizontal lines or more complex image elements. Basically, with an infinite amount of data DL–systems can cope with any finite deterministic comparison between any input sets and the corresponding output, although in practice their ability to learn a certain comparison depends on many factors. One common problem — please to the local minimum in which the system is stuck with a not exactly optimal solution when there is no better solution in the number of nearby solutions. In practice, results with large data sets usually turn out quite good, and on a wide range of potential comparisons. For example, in the field of speech recognition, a neural network learns to match a set of speech sounds and a set of tags (words or phonemes). When recognizing images, a neural network learns to match a set of images to a set of tags (for example, machine images are marked as machines). In the Atari game system, which is being taught by the DeepMind network, neural networks learn the correspondence between pixels and joystick positions. Systems GOs are most often used as classifications, in the sense that the mission of a typical network are the decisions about which set of categories (defined by the output modules of the neural network) this input belongs to [6].

In the course of this research work, some materials on the topic of master's work were collected and studied. Thus, various types of chatbots and their areas of application were considered. The main existing architectures for building neural networks and methods for their learning were also studied.

After analyzing the materials studied, the recurrent neural network was chosen as the architecture for building an ontological chatbot, combining the features of several subspecies to achieve the best result. The choice is based on the presence of such features in recurrent neural networks as, for example, the presence of short-term and long-term memory. For the network, the language structure is a long-term information to remember. It includes grammar, as well as stylistic features of the corpus of texts on which training is performed. In fact, the neural network remembers the order in which it is usually followed. words, and can add a sentence or construct your own.

Also, in the course of research work for the chatbot training, a method of deep learning was chosen, as it will allow to realize the potential of the recurrent neural network to the greatest extent, serving as the basis of the chatbot.

At the time of writing this essay the master's work is not yet completed. Estimated completion date: May 2019. Full text of the work, as well as materials on the topic can be obtained from the author or his manager after the specified date.

• Гудфеллоу, Я. Глубокое обучение / Я. Гудфеллоу, И. Бенджио, А. Курвилль. – М.: ДМК-Пресс, 2017. – 652 с.
• Савельев, А. В. На пути к общей теории нейросетей. К вопросу о сложности // Нейрокомпьютеры: разработка, применение / А.В. Савельев. – 2006. – № 4–5. – С. 4–14.
• Нейронные сети для начинающих. Часть 1 // Режим доступа: https://habr.com/post/312450.
• Нейронные сети для начинающих. Часть 2 // Режим доступа: https://habr.com/post/313216.
• Ясницкий, Л. Н. Введение в искусственный интеллект / Л. Н. Ясницкий. – М.: Издат. центр «Академия», 2005. – 176 с.
• Николенко, С. Глубокое обучение / С. Николенко, А. Кадурин, Е. Архангельская. – СПб.: Питер, 2018. – 480 с.
• Барцев, С. И. Адаптивные сети обработки информации / С.И. Барцев, В.А. Охонин. – Красноярск: Ин-т физики СО АН СССР, 1986. – 20 с.
• Миркес, Е. М. Логически прозрачные нейронные сети и производство явных знаний из данных // Нейроинформатика / Е. М. Миркес, А. Н. Горбань, В. Л. Дунин-Барковский, А. Н. Кирдин. – Новосибирск: Наука. Сибирское предприятие РАН, 1998. – С. 10–25.
• Терехов, В. А. Нейросетевые системы управления / В. А. Терехов, Д. В. Ефимов, И. Ю. Тюкин. – М.: Высшая школа, 2002. – 184 с.
• Вороновский, Г. К. Генетические алгоритмы, искусственные нейронные сети и проблемы виртуальной реальности / Г.К. Вороновский, К. В. Махотило, С. Н. Петрашев, С. А. Сергеев. – Харьков: Основа, 1997. – 112 с.
• Горбань, А.Н. Обучение нейронных сетей / А.Н. Горбань – М.: СССР-США СП «Параграф», 1990. – 160 с.
• Рекуррентные нейронные сети: типы, обучение, примеры и применение // Режим доступа: https://neurohive.io/ru/osnovy-data-science/rekurrentnye-nejronnye-seti.