EA038264B1 - Method of creating model for analysing dialogues based on artificial intelligence for processing user requests and system using such model - Google Patents

Abstract

The invention relates in general to the field of electronic data processing, and more particularly to machine learning methods for constructing natural language dialogue analysis models. A computer implemented method for creating a dialogue analysis model on the basis of artificial intelligence for processing user queries is carried out with the aid of at least one processor and comprises the steps of obtaining a set of primary data containing at least text data of dialogues between users and operators including user queries and operator responses; processing the set of data obtained, during which forming a training set for an artificial neural network containing positive and negative examples of user queries based on an analysis of the context of the dialogues, wherein said positive examples contain a semantically related set of operator replies in response to a user query; identifying and coding a vector representation of each reply from the positive and negative examples of the training set referred to in the previous step; using the resulting training set to train models for determining relevant replies from the context of user queries in dialogues.

Description

Technology area

This technical solution generally refers to the field of computational data processing, and in particular to machine learning methods for constructing dialogue analysis models in natural language.

State of the art

Currently, systems for automated natural language recognition are widely used in various branches of technology. The most widespread use of these technologies is observed in the user sector when used in various software applications, for example, search engines, navigators, applications for the selection of goods, etc., for example, when using intelligent assistants. A key feature in the work of such intelligent assistants is the ability to accurately recognize speech commands generated by users.

The existing difficulty is the formation of models for the analysis of speech messages, which, with a given accuracy and speed, allow you to quickly form and provide an answer to a user's request, especially when it comes to a specialized area of their application, which requires careful tuning and training of such models.

At the moment, from the prior art, quite a few approaches are known in the field of creating and training models for natural language processing (English NLP Natural Language Processing). The principle of creating models using a machine learning algorithm is known, which consists in applying a method for filtering sentences using a recurrent neural network and the Bag of words algorithm (patent application US 20180268298, applicant: Salesforce.com Inc., published 20.09. 2018). The well-known approach reveals the principle of sentiment analysis using two types of models - simple and complex, which classify the received message in natural language. The disadvantages of the known approach are low accuracy and speed of operation, which is due to the use of several models, selected depending on the type and complexity of the received request.

The essence of the technical solution

The claimed technical solution offers a new approach to the application of artificial intelligence (AI) by creating machine learning models for processing user requests in natural language.

The technical problem or technical problem to be solved is the creation of a new way of creating a model for analyzing calls in natural language, which has a high degree of accuracy in recognizing the context of a call and the speed of processing incoming calls.

The main technical result achieved when solving the above technical problem is the creation of a model for analyzing user requests in natural language, which has a high accuracy in recognizing the context of requests, due to the ability to rank responses to incoming requests from users.

The claimed result is achieved due to a computer-implemented method for creating a dialogue analysis model based on artificial intelligence for processing user requests, performed using at least one processor and containing the steps at which a set of primary data is obtained, and the set includes at least , text data of dialogs between users and operators, containing user requests and operators' responses;

processing the obtained data set, during which a training sample for an artificial neural network is formed, containing positive and negative examples of user requests based on the analysis of the context of dialogues, and the positive examples contain a semantically related set of operator replicas in response to the user's request;

performing selection and coding of vector representations of each replica from the positive and negative examples of the training sample mentioned in the previous step;

the generated training sample is used to train the model for determining the relevant replicas from the context of user calls in dialogues.

In one particular embodiment of the method, the model is at least one artificial neural network.

In another particular embodiment of the method, positive examples are formed on the basis of complete chains of dialogues between the operator and the client, and such a chain contains at least one interrogative sentence.

In another particular embodiment of the method, when selecting relevant replicas for responding to a customer's call phrase at the stage of training the model, a scoring score is calculated for each response.

In another particular embodiment of the method, in the step of encoding replicas into vectors, replica representing sentences are encoded as a matrix of semantic vectors.

Also, the specified technical result is achieved by implementing a system for processing user requests in an information channel using artificial intelligence, which

- 1,038264 contains at least one processor; at least one memory connected to the processor, which contains computer-readable instructions, which, when executed by at least one processor, provide: obtaining user access using the information channel; processing user requests using a machine learning model for automated processing of user requests created using the method described above; formation and transmission in the information channel of a response message to the user's request.

In a particular implementation, the system is a server, mainframe, or supercomputer.

In another particular implementation of the system, the information channel is a chat session, VoIP communication, or a telephone communication channel.

In another particular implementation of the system, the chat session is a chat using a mobile application or a chat on a website.

Description of drawings

Features and advantages of the present invention will become apparent from the following detailed description of the invention and the accompanying drawings.

FIG. 1 illustrates a block diagram of the implementation of the claimed method.

FIG. 2 illustrates an example of data processing to form a training sample.

FIG. 3 illustrates the architecture of the interrogative sentence definition model.

FIG. 4 illustrates a method for training a model to identify relevant replicas.

FIG. 5 illustrates the architecture of the Relevant Replica Definition Model.

FIG. 6 illustrates an example of using a trained model for determining relevant replicas.

FIG. 7 illustrates a general view of the claimed system.

Implementation of the invention

In this technical solution, for clarity of understanding of the work, such terms as operator, client, bank employee can be used, which in general should be understood as a user of the system.

The claimed method (100) for creating an AI-based dialog analysis model for processing user requests, as shown in FIG. 1 consists in performing a series of sequential steps carried out by a processor.

The initial step (101) for the formation of a dialogue analysis model is to obtain a primary (raw) data set, on which a training sample for an artificial neural network (ANN) will be built. A set of primary data can be an array of unlabeled text logs (records) of dialogs between operators and clients when processing incoming calls. The topic of text logs can be different and vary depending on the requirements for the final formation of the analysis model for a given industry of its final application.

Customer inquiries are understood as any requests coming to information channels of interaction with the operator of the contact center or support service, for example, a financial institution. As a rule, the primary data is transformed into a text form of records of conversations between customers and operators. Information channels for receiving data from conversations with operators can be, but are not limited to, a telephone channel, a VoIP channel, a chat session on a website or in a mobile banking application, a chat bot of a messenger, etc. Any type of channel through which the client can carry out a dialogue with the operator can be used to receive these dialogs for their subsequent translation into text form for the purpose of implementing the learning process of a machine learning model.

As mentioned above, the operator in this solution can mean both a person who processes the purposes of customer requests and a software algorithm, for example, a chat bot or an intelligent answering machine that can also provide information for processing customer requests.

Based on the received array of primary data with logs at step (101), it is further processed to form a training sample (102) of the ANN. From the resulting array of logs, question-answer pairs are formed. This procedure includes an algorithm for collecting data, using a model for determining an interrogative sentence, and an algorithm for forming question-answer pairs. Using the model for determining interrogative sentences, the client's questions are searched in text logs. The operator's replica following the found client's question (provided that it satisfies a number of requirements: not an interrogative, long enough, does not contain stop words) is considered the answer to this request.

The processing of the data array of dialogues between operators and clients consists in forming positive and negative examples based on question-and-answer pairs. As a rule, such examples are formed as follows: (the context of the conversation (2-5 replicas), the client's request or request, the operator's response to the request) - a positive example; (the context of the conversation (2-5 replicas), the client's request or request, the operator's response to some other request) is a negative example.

Below are examples of the implementation of question-answer pairs.

- 2 038264

Example 1 (positive).

ctx: ['Hello, Ivan Ivanovich!', 'How can I help?', 'Hello, can I get a credit card'] rsp: 'You can clarify the conditions for credit cards and apply at the link:

http: //www.sberbank.m/moscow/m/person/bank_cards/credit/ '

Example 2 (negative).

ctx: ['Hello, Kirill!', 'How can I help you?', 'Hello, how to activate the mobile banking service?'] rsp: 'Online you can order only Visa Gold and MasterCard Gold cards'.

The formation of question-answer pairs for creating a training sample for the ANN is carried out using the model of analysis of interrogative sentences, which is necessary for the correct partitioning of the context of the logs and the formation of a training dataset. In this case, context refers to a time-ordered set of operator and client replicas. The last cue in context is always the customer's question.

FIG. 2 shows an example of the application of the model for determining interrogative sentences for the formation at stage (102) of the training sample for the ANN. The interrogative sentence determination model (220) is a machine learning model such as an artificial neural network. To train the model (220), a dataset (sometimes called a dataset) OpenSubtitles (OPUS) (http://opus.nlpl.eu/) (221) can be used, as well as data from chats with an operator (222). The OPUS dataset (221) is an open dataset of subtitles for films in various languages, which is used as a source of colloquial vocabulary commonly found in feature films.

All sentences containing a question mark were selected as positive examples from data sets (221) - (222), and all other sentences were selected as negative examples. In OPUS (221), all sentences ending with a question mark are interrogative, since the punctuation in the subtitles is always correct. This also applies to interrogative sentences from these chats (222). The interrogative sentences extracted in this way also undergo additional filtering: short sentences or sentences containing stop words are discarded. Similar to positive examples, most of the sentences from OPUS (221) that do not contain a question mark are not interrogative and are chosen as negative examples.

Additionally, sentences containing interrogative words or too short, where the word size is predefined, are discarded.

Negative examples can be generated in any number, which allows you to achieve any ratio of positive and negative examples in the training set for ANN. Experiments have shown that the best quality is achieved when the ratio of positive and negative examples is 1: 1.

In an exemplary embodiment, the training set is balanced, all punctuation is cut out so that the model (220) builds its predictions based solely on the semantics of the words in the sentence. All obtained raw data were used, but the number of positive and negative examples was the same in each batch (data fragment) when training the model (220). For semantic analysis, the semantic word model fasttext and a recurrent neural network based on LSTM (Long short-term memory) are used to model the semantics of a sentence. FastText is a library for learning word embedding and text classification created by AI Research Lab at Facebook ™. The accuracy of this data set processing procedure is about 95%.

LSTM is a type of feedback loop recurrent neural network that is widely used in the industry for modeling time series and other sequences. This architecture is most widely used in computational linguistics, where it is used to model the semantics of sentences or entire paragraphs of text.

FIG. 3 shows the architecture of the model for determining interrogative sentences (220). The architecture of the model for determining interrogative sentences (220) is presented on the example of a neural network model in the form of an acyclic computational graph.

On the architecture of model (220), examples of dimensions of the input and output tensor for each block are indicated.

An example of a recording.

Entrance: (No, 20) Exit: (No, 20, 300).

This example means that the block accepts a tensor of dimension (batch_size, 20) as input and sends a tensor of dimension (batch_size, 20, 300). The size of the batch (data packet) for the trained model can be any (this only affects performance and depends on the runtime environment), for this, the notation indicates (No).

Model (220) contains an input node for text data (inp_ctx_0) (2201) and one output node for model prediction (relevance) (2211). As pre-trained embeddings (vector

- 3 038264 word representations) the FastText model is used, containing texts in Russian. A bidirectional LSTM module is used as an encoder.

The word vectorization module (2202) contains a pre-trained word embedding model for vectorization at the word level. Each of the sentences submitted to the model input is already broken down into tokens - presented as a list of words. Moreover, all sentences are presented as sequences of equal length (this is necessary for efficient batch processing). Short sentences are padded to this fixed length with a zero token, too long sentences are truncated. Hereinafter, the length of the sequences will be denoted as MAX_LEN. The experiments used the value MAX_LEN = 24, but not limited to.

Word embedding is a vector representation of a word, obtained using a distributive language model (usually word2vec, fasttext, or glove natural language semantics analysis software tools). This is a vector of dimensions of the order of several hundred (100-1000). A characteristic feature is that words similar in meaning are represented by similar (according to the Euclidean metric L2) vectors.

Each word is assigned a semantic vector - the so-called. word embedding (see source of information https://ru.wikipedia.org/wiki/Word2vec). For this, a fasttext model trained on a thematic, for example banking, dataset is used (see https://arxiv.org/abs/1607.04606). The advantage of the fasttext model over the basic word2vec is the ability to process (vectorize) words that were absent in the training set. Fasttext semantic vectors have a dimension of the order of several hundred, this dimension can be denoted as EMB_DIM. In the experiments performed with the architecture of the presented model (220), EMB_DIM = 300 was used.

As a result of these procedures, each sentence at the input of the model (220) is associated with a matrix of dimensions (MAX_LEN, EMB_DIM). This functionality is encapsulated in the word_embedding_model module (2202). The module (2202) is used to vectorize the context and response sentences. In this case, the word vectorization module (2202) is not trained in the process of setting up the model (220), since the word vectors in it are fixed and do not change anymore.

The proposal vectorization module (2203) contains a vectorization model for the entire proposal. Each of the sentences is represented as a matrix from (MAX_LEN, EMB_DIM) module (2202) and encoded into a vector of fixed dimension. For this, a recurrent neural network of the LSTM type is used. The matrix obtained by the module (2202) is processed from left to right by the LSTM module, as the vector representation of the sentence corresponds to the last internal state of the LSTM (that is, corresponding to the last word in the sentence).

As a result of the operation of the module (2203), each sentence will be represented in the form of a vector of fixed dimension LSTM_DIM. As an example of work, the cell dimension LSTM_DIM = 340 was used. Thus, the matrix (CTX_LEN, LSTM_DIM) is assigned to the query context, consisting of CTX_LEN replicas of maximum SEQ_LEN words (input inp_ctx). A candidate consisting of one sentence is assigned the vector LSTM_DIM. A candidate represents one of the possible responses for a given context. This functionality is encapsulated in the sentence vectorization module (2203). Module (2203) contains most of the model's trainable parameters (2-5 million depending on the configuration) and is the most computationally difficult.

The subsampling (pooling) modules (2204, 2205) receive a fixed-dimension vector from the internal state RNN of the sentence vectorization module (2203) as input. In a particular embodiment, the modules (2204, 2205) may be part of the sentence vectorization module (2203).

The concatenation module (2206) is designed to concatenate the vectors received from the modules (2204, 2205) into one, for their subsequent transfer to the multilayer perceptron (2207) as a single vector.

Multilayer perceptron (English MLP / Multilayer perceptron) (2207), in particular two-layer, in which fully connected (English Dense) layers are interspersed with regularization (English Dropout). The Dropout value for this MLP example is 0.3. Dropout is a way to regularize neural networks to combat overfitting (see, for example, http://imlr.org/papers/volumel5/srivastaval4a/srivastaval4a.pdf).

The module (2208) is an output neuron with a sigmoidal function of the model for determining interrogative sentences of activation (220) and contains the prediction of the model (220), made based on the processing of the input text data. In the example presented, the architecture of the presented model (220) reached 0.945 AUC (Area under the ROC Curve), 0.875 ACC (Accuracy) on the validation set. The area under the ROC Curve (AUC) is an aggregated characteristic of the quality of the classification, independent of the ratio of error prices. The higher the AUC value, the better the classification model. This indicator is often used for comparative analysis of several classification models.

Next, we will consider the stage of generating a training sample (102) for a reference analysis model. Generation of the training sample (102) is performed using a model (220) for extracting interrogative sentences from the input dataset (210), which represents unlabeled chat dialogs between clients and operators (210). Each client replica in each chat is processed with

- 4 038264 using the mentioned model (220).

At the (SW) stage, replicas are tokenized and presented as a sequence of word vectors, after which they are fed to the input of the model (220) for extracting interrogative sentences. During replica processing, model (220) estimates the likelihood that the replica is an interrogative sentence. If several operator messages follow the replica request, a positive training example is formed (231). If there are several replicas of the client in a row, the one for which the predicted probability turned out to be the highest is chosen as the interrogative. If there are several replica-responses of the operator in a row in response to the client's request, then a training example is formed with each of them.

Positive training example (231) includes all replicas up to and including the client request as context. The following operator's replica is used as a response. The context includes the last n replicas (of both the client and the operator) preceding the client's request, where n is a model parameter, for example, from 1 to 6. In an exemplary implementation, during the processing of the data set (210) by the model (220), a training a set that contained about 1,000,000 positive examples (231).

Negative examples (232) were formed by replacing the correct answer in the positive example with an arbitrary one from the set of all possible operator responses (of which there were about 1,000,000 at the time of training).

Based on the generated training sample, at stage (104), the model for determining the relevant replicas (240) is trained. FIG. 4 shows an example of training a model for determining relevant replicas (240). The input of the model (240) receives data from the training sample (230), formed on the basis of the obtained positive (231) and negative (232) examples of processing customer requests.

Training batches are formed from the training part of the training sample (230). The ratio of positive and negative examples in the batch is chosen to be approximately equal. Typical batch size is 256, 512. Model (240) is trained for 32 epochs, at the end of each it is validated on a deferred sample. Lazy sampling is a part of the dataset that is not used when training a model, but which is used to validate it (calculate metrics). As an example, the lazy sample can be 10% of the original generated training sample.

Optionally, in the validation process, in addition to the training set (230) obtained in automatic mode from raw (in other words, unlabeled) data, a manually labeled dataset of questions and answers can be used. If the manually labeled dataset is large enough (thousands of question-answer pairs), then in this way it is possible to additionally train the model (240) on these pairs. In this case, the training sample (230) is replaced with a manually labeled dataset, with the help of which further training of the model (240) is continued. This leads to a significant increase in quality metrics on questions from an additional dataset.

If there is little data, then the model (240) is validated on them by calculating the corresponding quality metrics. As a rule, recall @ k and precision @ k metrics are calculated for a question-answer system including model (240) - the model with the maximum values of these metrics can be selected for subsequent serialization in pickle (the pickle module implements the algorithm for serializing and deserializing Python objects). The value of this metric is determined by the frequency of getting the correct answer to a question in the top-K of the model's answers ranked by relevance. This value is calculated using the formula: (number of relevant responses up to the k-th position in the ranked list of responses) / (total number of relevant responses).

For example: models asked 10 questions, 5 times correct answer was the first in the list of sorted answers, and 8 times correct answer entered the top 3 sorted by relevance answers. In this case, for such a test recall@1=5/10=0.5, recall @ 3 = 8/10 (assuming that there is only one relevant answer for each question).

Training model (240) on average takes 2-3 hours, depending on the use of the GPU NVIDIA 1080TL graphics accelerator Model (240) with the maximum accuracy on deferred sampling is serialized into a binary format (pickle) for further use.

FIG. 5 shows the architecture of the model (240) for determining the relevant replicas. The relevance determination model (240) is designed to assess the relevance of a given context-response pair. The model has two input nodes - for the dialogue context (2401) and for the candidate replica (2402). The model has one output node (2407), which is a module for determining a relevance score, which can take values from 0 to 1.

The word vectorization module (2403) is similar in functionality to the module (2202), which also carries out vectorization at the word level. The notation Iterative means that the module (2404) can perform the prescribed data processing sequentially several times. In this case, the Iterative parameter for the word vectorization module (2404) indicates that the word vectorization module (240) is applied in turn for each replica of the context (of which there are 3 in this example). This is not required for the candidate's replica, since it consists of one sentence (corresponding

- 5 038264 vechno vectorization works once).

Module (2405) is designed to vectorize sentences and, in terms of its functionality, repeats the functionality of module (2203). In the example shown in FIG. 5 in the diagram, the subsampling and concatenation nodes are encapsulated inside the module (2405) and are not explicitly shown in the diagram. In this case, the Iterative parameter indicates that the sentence vectorization module (2406) is applied in turn for each replica of the context (of which there are 3 in this example).

Module (2407) is a relevance computation module. This module (2407) takes as input vector representations of the context and the candidate and returns a single number [0,1] the relevance score. The score is calculated based on the calculation of a number of factors including the dot product between the candidate vector and each of the vectors in the context, the dot product between the candidate vector and the sum of the context vectors; concatenation of context vectors and a candidate vector; calculating the dot product with the candidate vector.

The result of the concatenation of the context vectors and the candidate vector is fed to the input to the two-layer perceptron. The dimension of the output layer is LSTM_DIM. At the output, a matrix is formed (CTX_LEN, LSTM_DIM). The dot product is computed with the candidate vector, and the result is the CTX_LEN factors for the resulting context. The length of the context is denoted as CTX_LEN and can be from 1 (context - just a question) to infinity (context - the whole dialogue). Typical values: [1: 5]. As an example implementation, with CTX_LEN = 3, 7 factors are obtained for calculating relevance.

These factors are fed to the input of another two-layer perceptron with a sigmoidal activation function on the last layer. The output of the module (2406) is one number - the relevance score, which is the final output of the entire model (240). Model (240) is trained as a binary classifier, predicting whether a given candidate is relevant to a given context or not, that is, model (240) allows you to determine the answer to the received question in circulation.

The trained model (240) can be used to construct a question-answer system as follows. Of all the possible answers of the operator, a certain limited set of candidates is distinguished. During the selection process, too short, too long, meaningless and duplicate answers are excluded from candidates. This process can be either fully automated or semi-automated, in which the final list of candidates is additionally checked manually by a specialist, which makes it possible to obtain additional quality of the entire system. As a result, there are many candidates (usually from hundreds to thousands), each of which model (240) will be able to answer the query.

To answer the query, the model (240) evaluates the relevance of the query context to each of the loaded candidates. Top-K candidates (typical value k = 3) are returned as the most likely response options for a customer request.

After receiving the trained model (240) for determining the relevance of the replicas, this model (240) can be used in the future in automated systems for analyzing dialogs coming from the client. For example, such systems can be chat bots, intelligent assistants placed on websites, widgets, telephone robots, etc.

FIG. 6 shows an example of using the trained model (240) for determining the relevant replicas for processing customer requests (10), which can be received by the resource (20) upon request. Resource (20) can be used as a web resource (website, portal, etc.), call center, mobile application, etc. The client (10) can form his appeal in the form of a phone call, through a chat session, a VoIP call, using a specialized widget or software, etc. The resource (20), upon receiving the information of the request from the client (10), transfers the context of the request to the trained model (240), which determines the question-answer pair and transmits the data to the module for generating the response to the client's request (250) to generate the answer to the client's question (10 ). The response to the client's request (10) is transmitted from the module (250), as a rule, in the same information channel from which the request came. The response can be a chatbot response, a telephone robot, interactive information, a hyperlink or a combination of response options, etc.

FIG. 7 shows an example of a general view of a computing system (300) that implements the claimed method (100) or is part of a computer system, for example a server, a personal computer, a part of a computing cluster that processes the necessary data to implement the claimed technical solution.

In the general case, the system (300) contains one or more processors (301) united by a common bus of information exchange, memory means such as RAM (302) and ROM (303), input / output interfaces (304), input / output devices (305 ) and a device for networking (306).

The processor (301) (or multiple processors, multi-core processor, etc.) can be selected from a range of devices currently widely used, for example, such manufacturers as Intel ™, AMD ™, Apple ™, Samsung Exynos ™, MediaTEK ™, Qualcomm Snapdragon ™ etc. Under the processor or one of the used processors in the system (300), it is also necessary to take into account the graphic

- 6 038264 processor, for example, GPU NVIDIA or Graphcore, the type of which is also suitable for full or partial execution of the method (100), and can also be used for training and applying machine learning models in various information systems.

RAM (302) is a random access memory and is intended for storing machine-readable instructions executed by the processor (301) for performing the necessary operations for logical data processing. RAM (302), as a rule, contains executable instructions of the operating system and corresponding software components (applications, software modules, etc.). In this case, the available memory of the graphics card or graphics processor can act as RAM (302).

ROM (303) is one or more permanent storage devices such as a hard disk drive (HDD), solid state data storage device (SSD), flash memory (EEPROM, NAKD, etc.), optical storage media (CD-R / RW, DVD-R / RW, BlueRay Disc, MD), etc.

To organize the operation of system components (300) and to organize the operation of external connected devices, various types of I / O interfaces (304) are used. The choice of appropriate interfaces depends on the specific design of the computing device, which can be, but are not limited to, PCI, AGP, PS / 2, IrDa, FireWire, LPT, COM, SATA, IDE, Lightning, USB (2.0, 3.0, 3.1, micro, mini, type C), TRS / Audio jack (2.5, 3.5, 6.35), HDMI, DVI, VGA, Display Port, RJ45, RS232, etc. To ensure user interaction with the computing system (300), various means (305) I / O information are used, for example, a keyboard, display (monitor), touch display, touchpad, joystick, mouse manipulator, light pen, stylus, touch panel, trackball, speakers , microphone, augmented reality, optical sensors, tablet, light indicators, projector, camera, biometric identification (retina scanner, fingerprint scanner, voice recognition module), etc.

The means of networking (306) provides data transmission via an internal or external computer network, for example, Intranet, Internet, LAN, etc. One or more means (306) may include, but is not limited to, an Ethernet card, a GSM modem, a GPRS modem, an LTE modem, a 5G modem, a satellite communication module, an NFC module, a Bluetooth and / or BLE module, a Wi-Fi module, and etc. Additionally, satellite navigation means can also be used as part of the system (300), for example, GPS, GLONASS, BeiDou, Galileo.

As shown in FIG. 6, the resource (20) accessed by the client (10) can be organized using the system (300), which can be a server to provide the required functionality for processing incoming calls, recognizing response replicas using a trained model (240) and generating response messages using the module (250), which are transmitted over various wired and / or wireless data channels. Client requests (10) can also be generated using a device that contains partial functionality of the system (300), in particular, the client's device (10) can be a smartphone, computer, tablet, terminal and any other device that provides a communication channel with a resource ( 20) to generate and transmit an appeal and receive the required response, which may also include various types of digital information.

Thus, when applying the model for determining the relevant responses (240), created using the claimed method (100), a more accurate selection of the answer pairs in the automated mode according to the incoming context in user calls is achieved, which makes it possible to create a new, more improved way of training and applying the models. machine learning in AI-based systems.

The presented materials of the description disclose the preferred examples of the implementation of the technical solution and should not be interpreted as limiting other, particular examples of its implementation, not going beyond the scope of the requested legal protection, which are obvious to specialists in the relevant field of technology.

Claims (9)

CLAIM 1. A computer-implemented method for creating a dialogue analysis model based on artificial intelligence for processing user requests, performed using at least one processor and containing the steps at which a set of primary data is obtained, the set including at least text data dialogues between users and operators, containing user requests and operators' responses; processing the obtained data set, during which a training sample for an artificial neural network is formed, containing positive and negative examples of user requests based on the analysis of the context of dialogues, and the positive examples contain a semantically related set of operator replicas in response to the user's request; performing selection and coding of vector representations of each replica from the positive and negative examples of the training sample mentioned in the previous step; - 7 038264 apply the generated training sample to train the model for determining the relevant replicas from the context of user calls in dialogues. 2. The method according to claim 1, characterized in that the model is at least one artificial neural network. 3. The method according to claim 1, characterized in that positive examples are formed on the basis of complete chains of dialogues between the operator and the user, and such a chain contains at least one interrogative sentence. 4. The method according to claim 1, characterized in that when selecting relevant replicas for responding to a user's request phrase at the stage of training the model, a scoring score is calculated for each response. 5. The method according to claim 1, characterized in that at the stage of coding the replicas into a vector representation of the replica, whereby the representing sentences are encoded as a matrix of semantic vectors. 6. A system for processing user requests in an information channel using artificial intelligence, containing at least one processor; at least one memory connected to the processor, which contains computer-readable instructions that, when executed by at least one processor, provide a user access using the information channel; processing user requests using a machine learning model for automated processing of user requests, created using the method according to any one of claims 1-5; formation and transmission in the information channel of a response message to the user's request. 7. The system according to claim 6, characterized in that it is a server, mainframe or supercomputer. 8. The system according to claim 6, characterized in that the information channel is a chat session, VoIP communication or a telephone communication channel. 9. The system according to claim 6, characterized in that the chat session is a chat using a mobile application or a chat on a website.