RU2759090C1 - Method for controlling a dialogue and natural language recognition system in a platform of virtual assistants - Google Patents

Abstract

FIELD: computing technology.

SUBSTANCE: technical result is achieved due to receiving a dialogue initiation request from the user, containing a unique virtual assistant identifier and a dialogue identifier; searching in the virtual assistant database based on the assistant identifier received at the previous stage; searching in the database of a previously initiated dialogue based on the dialogue identifier received at the previous stage, wherein if the dialogue was found, the search is conducted in the database of the active session associated with the found dialogue; if the session is not found, a new session is created in the database for the found dialogue, determined by a new unique session identifier, identifier of the created dialogue, empty session context; wherein a request is received from the user in the dialogue, containing the identifier of the dialogue with the virtual assistant received during initiation of the dialogue, the text message of the user, the context of the request; the intents of the user are detected using machine learning algorithms receiving the virtual assistant identifier and the text message from the request of the user at the input, as a result, forming a list of intents and the degree of confidence that said intents are present in the message of the user; searching for the most appropriate rule for forming a response based on the intents of the user received at the previous stage.

EFFECT: increase in the accuracy of forming responses to the user by a virtual assistant.

9 cl, 4 dwg

Description

FIELD OF TECHNOLOGY

[001] This technical solution generally relates to the field of computing, and in particular to methods for managing dialogue and natural language understanding systems in a platform of virtual assistants.

LEVEL OF TECHNOLOGY

[002] To date, information dialogue systems have become widespread and are used in various areas of public life, for example, to organize an automatic test of knowledge, an automated user support service, to diagnose diseases, and so on. However, the existing information dialogue systems are designed to solve problems of a narrow profile, that is, they are able to support only a dialogue on a given topic. In addition, most of them do not have the ability to form a response in a natural language, to give an emotional coloring to the formed response, to perform any additional actions, including interacting with other information systems and subsystems. The presence of such capabilities would allow not only two-way exchange of information, instructions and commands between the user and the system, but also to conduct a full-fledged dialogue, creating the impression of communication with a live interlocutor for the user, as well as to solve the tasks set by the user much more efficiently. Therefore, at the moment, it remains relevant to develop such a method of communication between the user and the information dialogue system, which would expand the possibilities of interaction between the user and the information dialogue system.

[003] The closest analogue of the claimed invention is an adaptive natural language interface and a method for receiving, interpreting and performing user input in natural language, described in US patent 7216080 B2 "Natural-language voice-activated personal assistants copyright holder: INDNER ROBERT D JR Nuance Communications Inc Vlingo Corp, publ. 08.05.2007. The method includes a user inputting a request, receiving and converting the user's request into text, processing the text and generating a response in the form of an output command, converting the output command into an executive command, outputting the executive command to an additional system and / or subsystems for execution.

SUMMARY OF THE INVENTION

[004] The technical problem or the technical problem solved in this technical solution is the implementation of the dialogue control method and the natural language understanding system in the platform of virtual assistants.

[005] The technical result achieved by solving the above technical problem is to improve the accuracy of the formation of responses to the user by the virtual assistant through the use of machine learning algorithms.

[006] The specified technical result is achieved by implementing a method for managing a dialogue, executed by at least one processor, in which at least one request for initiating a dialogue is received from at least one user, containing a unique identifier of a virtual assistant and an identifier of the dialogue; searching the database for the virtual assistant based on the assistant ID obtained in the previous step; searching the database for a previously initiated dialogue based on the dialogue identifier obtained in the previous step, and if the dialogue was found, then searching the database for an active session associated with the found dialogue; if the session is not found, then a new session is created in the database for the found dialog, which is determined by a new unique session identifier, the identifier of the created dialog, and an empty session context; receive from the user a request in the dialogue, containing the identifier of the dialogue with the virtual assistant, received when initiating the dialogue, at least one text message of the user, the context of the request; detecting user intents by means of machine learning algorithms that receive the virtual assistant's identifier and a text message from the user's request as input, as a result of which a list of intents and the degree of confidence that these intents are present in the user's message are generated; search for the most appropriate rule for generating a response based on the user's intents obtained in the previous step.

[007] In some embodiments of the technical solution, the virtual assistant contains a unique assistant identifier and a set of named parameters.

[008] In some embodiments of the technical solution, upon receipt of a request from a user in a dialog, the search for occurrences of dictionaries of the DL language in the user's message is performed according to the following method:

• define all occurrences of text elements of strings of dictionaries DL using a pre-prepared prefix tree;

• selects all lines of dictionaries that meet the following heuristics:

a. the user's message contains all the text elements included in the DL dictionary line,

b. there is an order of found occurrences of text elements corresponding to the DL dictionary string,

если нет текстовых элементов, располагающихся в самом начале сообщения пользователя, то в строке словаря DL в начале должна стоять звездочка или суперзвездочка,

if there are no text elements located at the very beginning of the user's message, then an asterisk or a superstar should be at the beginning in the DL dictionary line,

если нет текстовых элементов, располагающихся в самом конце сообщения пользователя, то в строке словаря DL в конце должна стоять звездочка или суперзвездочка.

if there are no text elements located at the very end of the user's message, then an asterisk or superstar should be at the end in the DL dictionary line.

[009] In some implementations of the technical solution, when intents are found, if their classification confidence is lower than a configuration parameter, they are discarded.

[0010] In some implementations of the technical solution, for each intent, the classification confidence is multiplied by the configuration parameter and rounded to the accuracy determined by the configuration parameter.

[0011] In some implementations of the technical solution, once the intents are discovered, they are grouped by the name of the role assigned to them.

[0012] In some embodiments of the technical solution, when searching for a suitable rule for generating an answer for each question template and intent included in the DL language template, the weight of its match with the user's message is determined.

[0013] In some embodiments, the weight is a 32-bit unsigned integer.

[0014] In some embodiments of the technical solution, templates of the same weight are sampled by selecting from all found question templates and intents with the maximum weight, while the selected templates are removed from the search list.

[0015] In some embodiments of the technical solution, when searching for a suitable rule to generate a response

составляют список шаблон-ответов, относящихся к выбранным шаблон-вопросам и интентам;

make a list of template-answers related to the selected template-questions and intents;

проверяют возможность генерации непустого ответа, где пустым ответом считается ответ, не содержащий ни одного текстового элемента вывода;

check the possibility of generating a non-empty response, where an empty response is a response that does not contain any textual output element;

осуществляют рандомизацию, причем выбирают случайно один из отобранных ранее шаблон-ответов и определяют его ответом на сообщение пользователя;

carry out randomization, and select at random one of the previously selected template-responses and determine it as a response to the user's message;

формируют ответ по выбранному правилу, причем для выбранного шаблон-ответа обрабатываются все функциональные элементы языка DL в соответствии с их описанием;

form a response according to the selected rule, and for the selected response template, all functional elements of the DL language are processed in accordance with their description;

записывают журнал диалога и сформированный ответ записывается в базу данных.

write the dialogue log and the generated response is written to the database.

BRIEF DESCRIPTION OF DRAWINGS

[0016] The features and advantages of the present technical solution will become apparent from the following detailed description and the accompanying drawings, in which:

[0017] FIG. 1 shows an embodiment of a method for managing a dialogue in the platform of virtual assistants.

[0018] FIG. 2 shows an embodiment of the interaction between the user and the virtual assistant in the form of a block diagram.

[0019] FIG. 3 shows an example of implementation of the context of a user dialogue with a virtual assistant.

[0020] FIG. 4 shows an example of the implementation of a natural language understanding system in the platform of virtual assistants.

DETAILED DESCRIPTION OF THE INVENTION

[0021] Below will be discussed in detail the terms and their definitions used in the description of the technical solution.

[0022] In this invention, a system means a computer system, a computer (electronic computer), a CNC (numerical control), a PLC (programmable logic controller), computerized control systems and any other devices capable of performing a given, well-defined sequence of operations. (actions, instructions), centralized and distributed databases, smart contracts.

[0023] By a command processor is meant an electronic unit or an integrated circuit (microprocessor) executing machine instructions (programs), a smart contract, an Ethereum virtual machine (EVM), or the like. A command processor reads and executes machine instructions (programs) from one or more storage devices. The role of data storage devices can be, but are not limited to, hard disks (HDD), flash memory, ROM (read only memory), solid state drives (SSD), optical drives.

[0024] A program is a sequence of instructions for execution by a computer control device or command processing device.

[0025] NLU (English “Natural Language Understanding))) - understanding of natural language or interpretation of natural language

[0026] the NLU part of this technical solution consists of a component for determining the intentions of the user (classification of intents, English "lntent Classification", abbreviated IC) and extraction of named entities (English "Named Entity Recognition)), abbr. NER) from the user's message text.

[0027] The task of classification (categorization) is a task in which there are many objects, divided in some way into classes (categories). A finite set of objects is given, for which it is known what classes they belong to. This set is called a sample. The class of the remaining objects is unknown. It is required to construct an algorithm capable of classifying (categorizing) an arbitrary object from the original set. To classify an object means to indicate the number (or name) of the class to which this object belongs.

[0028] Object Classification is a class number or name given by the classification algorithm as a result of its application to that particular object.

[0029] Precision and recall are metrics that are used in evaluating most information retrieval algorithms.

[0030] The accuracy of an algorithm within a class is the proportion of classification objects that actually belong to a given class relative to all objects that the algorithm has assigned to this class. The completeness of the algorithm is the proportion of objects belonging to a class found by the classifier relative to all objects of this class in the test dataset.

[0031] The F-measure is a generally accepted quality metric for categorization problems and is a harmonic average between accuracy and completeness.

[0032] Determining the user's intent - categorizing what the user wanted to know when he entered the request text in the call to the chatbot. Replicas are categorized according to a predetermined list of possible user intentions. It is a classification algorithm at the level of short texts (user replicas). For example, a custom response “I want to open a deposit with your bank” can be categorized as the intent to “open deposits”.

[0033] The intent of a search query is the intent of the user that he puts into the query; the goal that drives a person when he enters a query into the search bar.

[0034] Named entity extraction is the process of separating boundaries and categorizing words or phrases in the text of user calls to the chatbot that are meaningfully related to one of the predefined categories that are interesting for possible further processing by the chatbot. It is also a classification algorithm at the level of individual words and phrases in user replicas. For example, in the custom response “I want to transfer 1000 rubles from a savings account to a debit card”, you can select the entities “1000 rubles” - “amount of money”, “savings account” - “bank product”, “debit card” - “bank cards”.

[0035] Dialogue refers to the exchange of related messages between two participants in this process, for example, between two users.

[0036] The platform of virtual assistants (hereinafter referred to as the platform) is a software package that implements the functionality of conducting dialogues between users and in an automated mode.

[0037] Dialogues are conducted in a text format in a natural language between the user of the platform and dialogue robots (chat bots), called virtual assistants (hereinafter referred to as assistants) within the platform. Virtual assistant is a software agent that can perform tasks for the user based on information entered by the user, data on his location, as well as information obtained from various Internet resources.

[0038] Channels (transports) for conducting dialogues within the technical solution can be various text communication channels (such as a widget on a website, instant messengers, etc.) connected to the platform via a server API based on the HTTP protocol.

[0039] This technical solution allows the simultaneous conduct of multiple dialogues between different users and different assistants.

[0040] A method for managing a dialogue in a virtual assistant platform, shown in a series of steps, will be described in detail below.

[0041] Step 110: receiving from at least one user at least one request to initiate a conversation containing a unique virtual assistant ID and a conversation ID.

[0042] In this technical solution, the user 210 interacts with the virtual assistant 230 through the communication device of the user 220, as shown in FIG. 2.

[0043] Each assistant 230 is defined by the following data:

• unique assistant identifier 230,

• a set of named parameters called the assistant context 230.

[0044] All context parameters are a pair of values:

уникальное имя параметра;

unique parameter name;

значение параметра, которое может быть различным для разных ассистентов 230 и может быть изменено.

a parameter value that can be different for different assistants 230 and can be changed.

[0045] The knowledge base of the assistant (hereinafter - the knowledge base) is a set of rules and dictionaries, including:

описания намерений (интентов) в виде набора примеров для обучения нейронной сети и правил на специализированном языке описания диалога DL,

descriptions of intentions (intents) in the form of a set of examples for training a neural network and rules in a specialized language for describing the dialogue DL,

описания выделяемых сущностей в виде набор размеченных примеров для обучения нейронной сети.

descriptions of the selected entities in the form of a set of marked examples for training a neural network.

[0046] The parameters of the context of the assistant 230 can be defined as system parameters - such parameters are used in the dialogue algorithm and directly affect the choice of priorities when generating responses to requests from the user 210.

[0047] This technical solution uses the following system parameters (the use of such parameters is described later in the current document):

• that_anchor.

[0048] All data of virtual assistants 230 is stored in the solution database.

[0049] The dialogue between the user 210 and the assistant 230 consists of the following steps: 1. Initiating a dialogue between the user 210 and the assistant 230,

2. Exchange of messages in text form between the participants of the dialogue, and is determined by the following data:

• a unique identifier of the dialogue, defined at the stage of dialogue initiation.

• the unique identifier of the assistant 230 with whom the dialogue is being conducted.

• a unique identifier of the user 210 conducting a dialogue with the assistant 230.

• a unique session identifier, where session is a segment of the dialogue determined by the fact that during it the time between messages of the user 210 did not exceed a predetermined value, called the maximum allowable time of inactivity (hereinafter referred to as the inactivity limit).

[0050] Each dialog identifier may contain dialog variables — these are variables used within the dialog. Their values persist throughout the conversation, but are cleared when the conversation ends. The criterion for the end of the dialogue is the set timeout for the user silence period.

[0051] Examples of rules are shown below:

[0052] // rule 1

$ my name is [- ** -]

# Nice to meet you [& 1]. [% var1 = "[& 1]"]

// rule 2

$ * what is * my * name *

# [if (% var1)] {You are [% var1].} [else] {fl don't know your name.}

[0053] Rule 1 in the instruction for generating the response (line #) specifies the command to assign the variable var1 a value that was extracted from the user's replica at the stage of its processing by the rule (line $).

[0054] In rule 2, in the instruction for generating a response, it is checked that the variable var1 is not empty (has a value). If the value of the var1 variable exists at this particular moment of the dialogue, then it is used to generate a response.

[0055] An example of a dialogue using these rules:

• User: my name is Vasya

• Bot (rule 1 worked): Nice to meet you, Vasya.

• User: what is my name?

• Bot (rule 2 worked): You are Vasya.

[0056] The dialogue ended. If after some time the same user 210 starts a new dialog, then the values of the variables will be "reset" and the bot will not "remember" the username:

• User: what is my name?

• Bot (rule 2 worked, part of the answer in else, because the value of var1 was reset at the end of the last dialogue): I don't know your name.

[0057] A conversation session is defined by the following data:

• unique session identifier.

• a unique identifier of the dialogue to which the session belongs.

• a set of named parameters called the session context.

[0058] All context parameters are a pair of values:

уникальное имя параметра.

the unique name of the parameter.

значение параметра, которое может быть изменено в ходе диалога.

parameter value that can be changed during the dialog.

[0059] Session context parameters can be defined as dialog-level parameters - such parameters retain their values between sessions, other parameters lose them, as shown in FIG. 3.

[0060] A session in which there are no messages from user 210, or for which the time elapsed since the last message from user 210 is less than the inactivity limit is called an active session.

[0061] All data of dialogues and sessions are stored in the database of the technical solution.

[0062] Dialogue initiation is the first stage in a dialog that defines all the dialog data and forms it within the technical solution, the result of which is the provision of a dialog identifier required for exchanging messages with an assistant.

[0063] To initiate a dialogue, the following decision parameters are required:

• assistant ID (required parameter),

• dialogue identifier.

[0064] The Dialogue ID is transmitted if the dialog being initiated is a continuation of a conversation that took place earlier between the same user 210 and assistant 230.

[0065] When a request to initiate a dialog is received, the following steps are performed.

[0066] Step 120: the database is searched for the virtual assistant based on the assistant ID obtained in the previous step.

[0067] In the first step, an assistant 230 is searched. Thus, a search is made in the database of the technical solution for the assistant 230 with the obtained identifier. If the assistant 230 is not found in the database, then the technical solution returns information about the error and completes the initiation.

[0068] Step 130: searching the database for a previously initiated conversation based on the conversation ID obtained in the previous step.

[0069] Dialogue search is then performed. This step is carried out only if the dialog ID was received. It searches the database for a technical solution for a previously initiated dialogue with the corresponding identifier.

[0070] Each user 210 has its own variables. User variables 210 are variables that store values between dialogs with the same user. Variable name format: user X.

[0071] Examples of templates are shown below.

// template 1

$ my name is [- ** -]

# Nice to meet you [& 1]. [% user_name = "[& 1]"]

// template 2

$ * what is * my * name *

# [if (% user_name)] {You are [% user_name].} [else] {I don’t know your name.}

[0072] An example of a dialogue is shown below:

[0073] User: my name is Vasya

[0074] Virtual assistant (rule 1 triggered): Nice to meet you, Vasya.

[0075] User: What is my name?

[0076] Virtual assistant (rule 2 triggered, part of the response under the condition for a non-empty username variable): You are Vasya.

[0077] The dialogue ended. One day later, the user 210 comes from the same computer to talk to the virtual assistant 230, who “understands” that this is the same user 210.

[0078] User: What is my name?

[0079] Virtual assistant (rule 2 worked, part of the response under the condition for a non-empty variable user name, because the value of this variable is saved between dialogues): You Vasya.

[0080] If the dialog was found, then a search is made in the database for the active session associated with the found dialog. If the session is not found, then a new session (which is automatically active) is created in the database for the found dialogue, determined by the following data:

• a new unique session identifier.

• the identifier of the created dialogue.

• empty session context.

[0081] The technical solution returns the ID of the found dialog and completes the initiation.

[0082] The next step is to create a dialog. This step is performed if the previous step was skipped or the target dialog was not found on it. The following actions are performed on it.

[0083] A new dialog is created in the technical solution database, defined by the following data:

• a new unique identifier for the dialogue.

• received assistant ID.

[0084] For the created dialogue, a new session is created in the technical solution database, defined by the following data:

• a new unique session identifier.

• the identifier of the created dialogue.

• empty session context.

[0085] The technical solution returns the ID of the created dialog and completes the initiation.

[0086] Messaging will then be discussed in more detail below.

[0087] Step 140: receive from the user a request in the dialogue containing the ID of the dialogue with the virtual assistant obtained when initiating the dialogue, at least one text message of the user, the request context.

[0088] The user 210 of the technical solution sends a request that contains the following data:

• the ID of the dialogue with the assistant 230, obtained when the dialogue was initiated.

• text message of user 210.

• a set of named parameters called the user 210 request context (hereinafter the request context).

[0089] All context parameters are a pair of values:

• unique name of the parameter,

• parameter value.

[0090] The task of the technical solution is to process such a request, generate the most appropriate response to it and transmit it to the user.

[0091] Within this task, the following steps are performed.

[0092] The first step is to search the database for a previously initiated dialog with the corresponding identifier.

[0093] If the dialog was not found, then the technical solution returns an error and ends the processing of the request of the user 210.

[0094] Next, the session is received.

[0095] This step searches the database for the active session associated with the found dialog. If the session is not found, then a new session (which is automatically active) is created in the database for the found dialogue, determined by the following data:

• a new unique session identifier.

• the identifier of the created dialogue.

• empty session context.

[0096] Next, the session context is updated.

[0097] This step is only performed if the request context is not empty. On it, the session context is updated according to the following algorithm:

• all parameters whose names are absent in the session context but are present in the request context are added to the session context with the corresponding values from the request context.

• all parameters whose names are present both in the context of the session and in the context of the request, in the context of the session are assigned the corresponding values from the context of the request.

[0098] Next, the search context is generated. A search context is a set of named parameters obtained from the context of the assistant 230 and the session context according to the following algorithm.

[0099] The context of the assistant 230 is copied into the search context. All parameters whose names are absent in the search context but are present in the session context are added to the search context with the corresponding values from the session context. All parameters whose names are present both in the search context and in the session context are assigned corresponding values from the session context in the search context.

[00100] Next, a search is made for occurrences of the dictionaries DL.

[00101] This step searches for all occurrences of the DL language dictionaries in the user 210 message using the following algorithm:

• all occurrences of textual elements of strings of dictionaries DL (not containing elements of the DL language) are searched using a pre-prepared prefix tree.

• all lines of dictionaries are selected that correspond to the following heuristics:

в сообщении пользователя 210 есть все текстовые элементы входящие в строку словаря DL,

user 210's message contains all the text elements included in the DL dictionary line,

существует порядок найденных вхождений текстовых элементов, соответствующий строке словаря DL,

there is an order of found occurrences of text elements corresponding to the DL dictionary string,

если нет текстовых элементов, располагающихся в самом начале сообщения пользователя 210, то в строке словаря DL в начале должна стоять звездочка или суперзвездочка,

if there are no text elements located at the very beginning of the message of user 210, then an asterisk or superstar should be at the beginning in the DL dictionary line,

если нет текстовых элементов, располагающихся в самом конце сообщения пользователя, то в строке словаря DL в конце должна стоять звездочка или суперзвездочка.

if there are no text elements located at the very end of the user's message, then an asterisk or superstar should be at the end in the DL dictionary line.

• all selected lines of the dictionary are checked for exact correspondence with the user's message.

[00102] The beginning of the formation of the pool of dedicated named entities is described below.

[00103] For each DL dictionary, the occurrences of which were found in the previous step, the occurrences are selected for which the number of words in the DL template when instead of substituting all inline dictionaries with their specific values is the maximum for this dictionary. Such occurrences are called distinguished named vocabulary entities, the name of which is the same as the name of the DL vocabulary of which they belong.

[00104] A set of dedicated named vocabulary entities forms a plurality of dedicated entities for a message from user 210.

[00105] Step 150: discovery of user intents is carried out by means of machine learning algorithms that receive the ID of the virtual assistant and a text message from the user's request as input, as a result of which a list of intents and the degree of confidence that these intents are present in the user's message is generated.

[00106] Next, detection of possible intentions of the user 210 (intents) is performed.

[00107] In this step, the neural network algorithms described below are used to detect possible intentions, which receive the assistant ID 230 and a text message from the user's request 210 as input, and the result of their execution is a list of possible intentions and the degree of confidence that this intention is valid. is present in user 210's message.

[00108] All intents whose classification confidence is below a configuration parameter (eg, Threshold value) are discarded.

[00109] For each intent, the confidence is multiplied by a specific configuration parameter (Multiplier) and rounded to the accuracy specified by the configuration parameter (Accuracy). If the confidence exceeds the value of the configuration parameter (Limit), then the confidence value is set to Limit. The resulting value is considered further as the confidence of this intent.

[00110] Next, the extraction of named entities is performed based on the neural network. At this step, the neural network algorithms described below are used to extract named entities, which receive the assistant ID 230 and a text message from the user's request 210 as input, and the result of their execution is a list of selected entities with an indication of their roles and the degree of confidence that they were highlighted correctly.

[00111] All selected entities are grouped by the name of the role assigned to them. For each name, the entity with the highest confidence score is selected. For each such entity, it is checked whether there is an entity with the same name in the set of selected entities formed earlier. If there is no such entity, then it is added to the set of selected entities.

[00112] Step 160: search for the best rule to generate a response based on user intents obtained in the previous step.

[00113] The next step is to search for the most appropriate rule to generate a response. For each template-question and intent included in the template of the DL language, the weight of its coincidence with the message of user 210 is determined. The weight is a 32-bit unsigned integer determined according to the following principles:

• if the DL template contains conditions for the special variable that_anchor, then the 31st bit takes on the value 1, otherwise 0.

• for intents and question patterns that contain asterisks and superstars, the 30th bit is always 0, for question patterns that do not contain asterisks and superstars, the 30th bit is 1.

• A 16-bit number located from the 14th to the 29th bit is defined as follows:

для интентов оно равно числу слов в сообщении пользователя, умноженное на уверенность интента. При превышении значения 2^16-1, значение устанавливается в 2^16-1.

for intents, it is equal to the number of words in the user's message multiplied by the confidence of the intent. When the value is exceeded 2 ^ 16-1, the value is set to 2 ^ 16-1.

для шаблон-вопросов оно равно числу слов в шаблон-вопросе в раскрытом виде (после подстановки конкретных значений вместо элементов языка DL таких как словари или inline-словари), включая звездочки и суперзвездочки. При превышении значения 2^16-1, значение устанавливается в 2^16-1.

for pattern-questions, it is equal to the number of words in the pattern-question in the expanded form (after substitution of specific values instead of elements of the DL language such as dictionaries or inline dictionaries), including asterisks and superstars. When the value is exceeded 2 ^ 16-1, the value is set to 2 ^ 16-1.

• 8-bit number, located from the 6th to the 13th bit, is defined as the number of pattern-conditions in the pattern. If the number of conditional patterns exceeds 2 ^ 8-1, then the value is set to 2 ^ 8-1.

[00114] Next, the following selection algorithm is executed.

[00115] In the first step, templates of the same weight are sampled by selecting from all found question templates and intents with the maximum weight, and the selected templates are removed from the search list. If the list is empty, then an empty answer is considered generated and the next two steps are skipped.

[00116] Next, a sample of all possible response patterns is sampled.

[00117] A list of template-responses related to the template-questions and intents selected in the previous step is compiled, the possibility of generating a non-empty response is checked, where an empty response is a response that does not contain any textual output element (text, inline-dictionary, dictionary, etc. .NS.). If the list turns out to be empty, then return to the previous step.

[00118] Next, randomization is performed. Choose at random one of the previously selected template-responses and determine it as a response to the user's message.

[00119] Then, a response is generated according to the selected rule. For the selected response template, all functional elements of the DL language are processed in accordance with their description.

[00120] Next, a dialogue log is recorded. The generated answer is recorded in the database of the technical solution.

[00121] Then save the context of the session. For each instruction to change the session, change the context of the session in accordance with this instruction. The updated session context is saved to the technical solution database.

[00122] At the last step, the response is transmitted to the user 210. The generated response is transmitted to the user 210.

[00123] Machine learning algorithms used in this technical solution will be described below.

[00124] Modern algorithms for classifying intents and extracting named entities traditionally consist of two parts, namely, extracting features from the text in some numerical vector form and applying a special decision function to these numerical vectors, which results in the class number in the list of intent classes for the problem classifications of intents and labels with numbers of entity names from the list of entities for all words of the original replica of the user.

[00125] Feature extraction is a kind of abstraction, a dimensionality reduction process in which the original set of original variables is reduced to more manageable groups (features) for further processing, while remaining a sufficient set to accurately and completely describe the original dataset.

[00126] Text data vectorization is a method of representing selected features through transformation into numerical vectors of a certain multidimensional space.

[00127] The function that vectorizes text data is called a vectorizer (embedder), and the received features corresponding to the original phrase are called a vector (sentence embedding).

[00128] Currently, ELMo (Embeddings from Language Models) and BERT (Bidirectional Encoder Representations from Transformers) are among the best vectorizers in terms of qualitative indicators (according to the final F-measure of the complete intent classification algorithm or extracting named entities from arbitrary datasets).

[00129] ELMo is a more productive version (capable of processing more text cues per second), built on the basis of Bi-LSTM networks (Bidirectional Long-Short-Term-Memory, bidirectional recurrent artificial neural networks consisting of LSTM layers, long short-term memory). ELMo looks at the whole sentence before assigning each word its embedding. The neural network of such a vectorizer is trained (selects and optimizes the coefficients of some complex numerical function to better approximate the expected behavior of the algorithm) specifically for the task of creating such embeddings. The model, during training on a large set of texts, tries to predict the next word in sentences, as well as the previous one during the reverse passage through the sentence. Preliminary learning takes place on texts without markup, automatically generating pairs "part of a sentence" - "next or previous word". Thus, the model remembers co-occurring words and stable language constructs.

[00130] BERT is a heavier version (with lower performance), the architecture of which consists of parts with sequence-to-sequence neural networks (models that take as input a sequence of elements, words, letters, other features, and returning another sequence of elements), based on the attention mechanism. It is a transforming neural network consisting of a whole stack of sequence-to-sequence network encoders. It is possible to train such a model by "masking" individual words in sentences without any markup. A model that has broken several records for the success of solving a number of NLP problems. There is also a large number of pre-trained models for different languages in the public domain, including multilingual ones for solving all kinds of problems.

[00131] In the NLU part of this technical solution, a lightweight implementation of BERT (DistilBERT) for classifying intents and ELMo for extracting named entities can be used as vectorizers based on quality and performance indicators. There are no special preprocessing of texts, since both of these models work better with source texts and are able to take into account parts of speech, word forms, punctuation marks and other additional information that may be lost during preprocessing, and based on knowledge of word combinations and language constructions, they make their own predictions.

[00132] After receiving the embedding for the entire user's replica, it remains to classify this embedding in vector space. This problem in this technical solution can be solved both by means of classical machine learning and by using shallow neural networks consisting of 2-3 fully connected layers. Empirically, on representative datasets (about 500 intents, at least 50 examples for each intent), a comparison was made in terms of the quality and performance of algorithms (k-NN, Random Forest, SVM, XGBoost, their compositions, shallow neural networks), as a result of which an approach with Feed-Forward Neural Network (FFNN) was selected from three fully connected layers with an output layer equal to the number of required classes and architecture parameters obtained using Bayesian optimization.

[00133] Bayesian optimization is a global optimization technique for an unknown function. The neural network architecture search (NNAS) task involves the automatic selection of the neural network architecture. This is the task of neural network morphism, when the functionality of the neural network is preserved, but the architecture of the neural network is selected differently. The key idea of the proposed method is to explore the search space of architectures using a Bayesian optimization (BO) algorithm. Traditionally, Bayesian optimization consists of a loop with 3 steps: updating, generating and observing. In the context of searching for a neural network architecture, the cycle consists of the following steps:

• Update: learns a Gaussian process on existing architectures and their results;

• Generation: a new architecture is generated through the optimization of the acquisition function;

• Observation: record the results of the new architecture.

[00134] This is how the number of layers, neurons on each layer, the type of normalization used, activation, the probability of turning off neurons in Dropout layers is selected.

[00135] The final optimal neural network for classifying intents is automatically selected for each dataset; in the example with the considered representative dataset, an architecture consisting of a sequence of layers was obtained: the first fully connected layer, batch normalization, Dropout layer, activation Elu layer, second fully connected layer , batch normalization, Dropout layer, Elu activation layer, third fully connected layer, in total - 233 neurons in fully connected layers. In essence, such a neural network is a hybrid of the FFNN network. At the output from this neural network, a vector of probability distribution is obtained for each of the classes (the number of elements of the vector is equal to the number of classes).

[00136] As a result, the following sequence of processing user replicas is obtained: replica -> DistilBERT -> embedding -> FFNN -> probability distribution of the replica belonging to intents.

[00137] After receiving the embeddings of the words of the user replica, they are further classified according to a given set of entity labels (tags). When marking up text for training, each word is marked with a special tag - the name of the class to which the word belongs, and the position of the word in relation to the whole marked entity in the form of a prefix:

[00138] B - from the word beginning - the first word in an entity consisting of more than 1 word.

[00139] I - from the word inside - words that are in the middle of an entity consisting of more than 1 word.

[00140] O - from the word out - if the word does not refer to any entity, it is marked with this label.

[00141] For each word, you need to get a tag. This can be done in a number of ways. A simpler and more obvious way is to use some neural network, the last layer of which will be a fully connected layer of dimension d, where d is the number of possible word labels. Thus, we get the probabilities of a word to have each of the possible labels (and we can choose the most probable one). But to take into account their mutual dependence in the final tags, this is not enough.

[00142] After comparing the quality metrics of various algorithms of classical machine learning (k-NN, Random Forest, SVM, XGBoost, their composition) and neural network architectures (Bi-LSTM Encoder + CRF, sequence-to-sequence model) as a classifier Bi-LSTM Encoder was chosen with an additional CRF (conditional random field) layer, taking into account the relationship of tags to each other for the correct placement of prefixes. The output from this neural network is a probability distribution vector for each of the tags for each word (the number of vector elements is equal to the number of tags).

[00143] As a result, the following sequence of processing user replicas is obtained: replica -> ELMo -> word embeddings -> Bi-LSTM Encoder + CRF -> probability distribution of replica words belonging to entity tags.

[00144] Vectorization models are not further trained within the platform, but are prepared and supplied by developers for each specific area of application. Such models are trained in advance without markup on large text arrays of target content. For example, large corpuses of the Russian language, such as Wikipedia and news sites, extended with the logs of communication between bank customers in chats with operators for training vectorizers for banking topics. Then, in the case of the BERT model, for use in classification, a one-time additional training can be performed in the composition together with the classifier on the target training data (pairs "replica" - "intent"), which significantly increases the final quality metrics.

[00145] Models for classifying intents and entity tags are trained as part of the operator's interaction with the platform on user data marked up by operators, entered through the platform IDE interface in the "examples" sections. The platform has the following adjustable learning hyperparameters that can be changed in a text configuration file: learning rate, choice of loss function, metrics for optimization, number of learning epochs and batch size. When you start training from the IDE, all examples are collected with marks of their intents and entity tags, the data passes through vectorizers, and phrase embeddings are collected for subsequent intent classification and word embeddings for extracting named entities. The data is stratified into training and validation samples (the proportion of partitioning into these parts is also set in the configuration file with training hyperparameters, and this proportion is observed for each intent and entity tag), then for the task of classifying intents, the classes are automatically balanced by the number of examples. For this task, oversampling (artificial expansion of the number of examples in classes with a small number of examples) of the training part of the sample is used using the SMOTE (Synthetic Minority Oversampling TEchnique) technique of small classes to the size of the largest. This is a special technique for generating synthetic (fake) data in the embedding space obtained from the vectorizer (and not at the level of generating texts fed to the vectorizer). Thus, we increase the sensitivity of the classifier to small classes.

[00146] Next is the direct training of the classifier (intents and entities). During training, the learning rate is automatically adjusted, and an early stop can be made in case of fast convergence (so as not to pass all the given epochs in vain), or if the algorithm cannot be significantly optimized on the current dataset.

[00147] Upon completion of training, all received model parameters are serialized into model files that can be used to create new instances of intent classifiers and named entity extraction models.

[00148] All NLU components of a part of a technical solution work inside separate services, which are wrappers around neural network modules. They include a separate service with vectorizers (for optimization purposes, more and more heavy neural networks, vectorizers, were moved to a separate service), services for classifying intents, training intent classifiers, extracting named entities and training models for extracting named entities.

[00149] To initialize runtime services, you must specify the path to the files of the corresponding models, from which the services are raised. Service requests must contain replicas of users, and service responses are lists of the most likely intents with their respective probability levels from 0.0 to 1.0, and sets of words that may be representatives of the sought named entities. For each such highlighted word in the response of the named entity extraction service, there is its position in the text (indentation from the beginning of the replica in characters) and a list of entities with their probability levels. The most likely intents and entities can be used in further processing of the user's replica by the platform.

[00150] The training services in the body of requests to start training receive paths to text files with sets of examples marked up by IDE replica operators and generate model files. If necessary, training can be stopped, requesting the status of training (find out at what stage of the process), or at the end use the generated model files to initialize new runtime services.

[00151] Referring to FIG. 4, this technical solution can be implemented as a computer system 400 for implementing a dialogue control method and a natural language understanding system in a virtual assistant platform that contains one or more of the following components:

• a processing component 401 comprising at least one processor 402,

• memory 403,

• component 405 multimedia,

• component 406 audio,

• interface 407 input / output (I / O),

• sensory component 408,

• data transfer component 409.

[00152] The processing component 401 generally controls all of the operations of the system 400, such as processing user data or a chat request, and also controls the display, phone call, data transfer, camera operation, and recording operation of the mobile communication device. The processing component 401 may include one or more processors 402 that implement instructions to complete all or part of the steps of the above methods. In addition, the processing component 401 may include one or more modules for convenient communication between other processing modules 401 and other modules. For example, the processing component 401 may include a media module for convenient, lightweight interaction between the media component 405 and the processing component 401.

[00153] The memory 403 is configured to store various types of data to support the operation of the system 400, such as a user profile database. Examples of such data include instructions from any application or method, contact information, address book data, messages, images, videos, etc., all of which operate on system 400. Memory 403 can be implemented as any type of volatile storage device, nonvolatile memory, or a combination thereof, such as static random access memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (EPROM), Read Only Memory device (ROM), magnetic memory, flash memory, magnetic disk or optical disk and others, but not limited to.

[00154] The multimedia component 405 includes a screen that provides an output interface between a system 400 that may be installed on a user's mobile communication device and a user. In some implementations, the screen may be a liquid crystal display (LCD) or a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen for receiving input from a user. The touch pad includes one or more touch sensors in terms of gestures, touching, and sliding across the touch pad. The touch sensor can not only sense the subject's touch boundary or swipe gesture, but also detect the length of time and pressure associated with touch and slide behavior. In some embodiments, the media component 405 includes one front camera and / or one rear camera. When the system 400 is in an operating mode, such as shooting mode or video mode, the front camera and / or the rear camera can receive media data from the outside. Each front camera and rear camera can be one fixed lens optical system or can have focal length or optical zoom.

[00155] The audio component 406 is configured to output and / or input an audio signal. For example, the audio component 406 includes a single microphone (MIC) that is configured to receive an external audio signal when the system 400 is in a mode of operation, such as a call mode, a recording mode, and a speech recognition mode. The resulting audio signal can be further stored in memory 403 or routed to data transfer component 409. In some embodiments, the audio component 406 also includes a single speaker configured to output an audio signal.

[00156] The input / output (I / O) interface 407 provides an interface between the processing component 401 and any peripheral interface module. The above peripheral interface module can be keyboard, steering wheel, button, etc. These buttons may include, but are not limited to, a start button, a volume button, a start button, and a lock button.

[00157] The sensor component 408 contains one or more sensors and is configured to provide various aspects of assessing the state of the system 400. For example, the sensor component 408 can detect on / off states of the system 400, the relative position of components, such as a display and a keypad, of a single component. system 400, presence or absence of contact between a subject and system 400, and orientation or acceleration / deceleration and temperature change of system 400. Sensor component 408 comprises a proximity sensor configured to detect the presence of an object in the vicinity when there is no physical contact. Touch component 408 includes an optical sensor (eg, CMOS or CCD image sensor) configured for use in rendering an application. In some embodiments, sensor component 408 includes an acceleration sensor, gyroscope sensor, magnetic sensor, pressure sensor, or temperature sensor.

[00158] The communication component 409 is configured to facilitate wired or wireless communication between the system 400 and other devices. System 400 can access a wireless network based on a communication standard such as WiFi, 2G, 3G, 5G, or a combination thereof. In one exemplary embodiment, communication component 409 receives a broadcast signal or broadcast associated information from an external broadcast control system via a broadcast channel. In one embodiment, communication component 409 comprises a Near Field Communication (NFC) module to facilitate near field communication. For example, the NFC module may be based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[00159] In an exemplary embodiment, system 400 may be implemented by one or more Application-Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Units (DSPs), Programmable Logic Devices (PLDs), a logic chip programmed into operating conditions (FPGA), controller, microcontroller, microprocessor, or other electronic components, and can be configured to implement the dialogue control method 100 and the natural language understanding system in the virtual assistant platform.

[00160] In an exemplary embodiment, a nonvolatile computer-readable medium includes a memory 403 that includes instructions where instructions are executed by the processor 401 of system 400 to implement the above-described conversation control methods.

[00161]. For example, non-volatile computer-readable media can be ROM, random access memory (RAM), compact disc, magnetic tape, floppy disks, optical storage devices, and the like.

[00162] Computing system 400 may include a display interface that transmits graphics, text, and other data from a communications infrastructure (or from a framebuffer, not shown) for display on a multimedia component 405. Computing system 400 further includes input devices or peripherals. Peripheral devices may include one or more devices for interacting with a user's mobile communications device, such as a keyboard, microphone, wearable device, camera, one or more audio speakers, and other sensors. Peripheral devices can be external or internal to the user's mobile communications device. A touchscreen can display, typically graphics and text, and also provides a user interface (such as, but not limited to, a graphical user interface (GUI)) through which a subject can interact with a user's mobile communication device, such as access and interact with applications running on the device.

[00163] The elements of the proposed technical solution are in a functional relationship, and their joint use leads to the creation of a new and unique technical solution. Thus, all blocks are functionally linked.

[00164] All blocks used in the system can be implemented with electronic components used to create digital integrated circuits, which is obvious to a person skilled in the art. Not limited to, microcircuits can be used, the logic of which is determined during manufacture, or programmable logic integrated circuits (FPGA), the logic of which is set through programming. For programming, programmers and debugging environments are used that allow you to set the desired structure of a digital device in the form of a circuit diagram or a program in special hardware description languages: Verilog, VHDL, AHDL, etc. An alternative to FPGAs can be programmable logic controllers (PLCs), basic matrix crystals ( BMK) requiring a factory production process for programming; ASICs are specialized custom large integrated circuits (LSI), which are significantly more expensive for small-scale and single-piece production.

[00165] Typically, the FPGA itself consists of the following components:

• configurable logic blocks that implement the required logic function;

• programmable electronic links between configurable logic blocks;

• programmable input / output blocks providing connection of the external output of the microcircuit with the internal logic.

[00166] Blocks can also be implemented using read-only memory devices.

[00167] Thus, the implementation of all the blocks used is achieved by standard means based on the classical principles of implementing the foundations of computing.

[00168] As one skilled in the art will appreciate, aspects of the present technical solution may be embodied in a system, method, or computer program product. Accordingly, various aspects of the present technical solution may be implemented solely as hardware, as software (including application software, and so on), or as an embodiment combining software and hardware aspects, which may generally be referred to as a "module" , "System" or "architecture". In addition, aspects of the present technical solution may take the form of a computer program product implemented on one or more computer-readable media having computer-readable program code that is implemented thereon.

[00169] Any combination of one or more computer readable media can also be used. The computer-readable storage medium can be, without limitation, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, device, or any suitable combination thereof. More specifically, examples (non-exhaustive list) of a computer-readable storage medium include: an electrical connection using one or more wires, a portable computer diskette; hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), fiber optic connection, compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any combination of the above. As used herein, a computer-readable storage medium can be any flexible storage medium that can contain or store a program for use by the system itself, device, apparatus, or in connection therewith.

[00170] The program code embedded in a computer-readable medium can be transmitted using any medium, including, without limitation, wireless, wired, fiber optic, infrared, and any other suitable network or a combination of the above.

[00171] Computer program code for performing operations for the steps of the present technical solution may be written in any programming language or combinations of programming languages, including an object-oriented programming language such as Python, R, Java, Smalltalk, C ++, and so on, and common procedural programming languages such as "C" programming language or similar programming languages. The program code can be executed on the user's computer in whole, in part, or as a separate software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer. In the latter case, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN), a wide area network (WAN), or a connection to an external computer (for example, via the Internet using Internet service providers).

[00172] Aspects of the present technical solution have been described in detail with reference to block diagrams, circuit diagrams, and / or diagrams of methods, devices (systems), and computer program products in accordance with embodiments of the present technical solution. It should be appreciated that each block from the block diagram and / or diagrams, as well as combinations of blocks from the block diagram and / or diagrams, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other data processing device to create a procedure, such that instructions executed by a computer processor or other programmable data processing device create means to implement the functions / actions specified in block or blocks of flowchart and / or diagram.

[00173] These computer program instructions may also be stored on a computer-readable medium that can control a computer other than a programmable data processing device or devices that function in a particular way, such that the instructions stored on the computer-readable medium create a device including instructions that perform the functions / actions specified in the block diagram and / or diagram.

Claims (22)

1. A method for managing a dialog, executed by at least one processor and comprising the following steps: • receive from at least one user at least one request to initiate a dialogue, containing a unique identifier of the virtual assistant and the identifier of the dialogue; • search the database for a virtual assistant based on the assistant's ID obtained in the previous step; • search the database for a previously initiated dialog based on the dialog identifier obtained in the previous step, and a) if the dialog was found, then a search is performed in the database of the active session associated with the found dialog, b) if the session is not found, then a new session is created in the database for the found dialog, which is determined by a new unique session identifier, the identifier of the created dialog, an empty session context; • receive from the user a request in the dialogue, containing the identifier of the dialogue with the virtual assistant, received at the initiation of the dialogue, at least one text message of the user, the context of the request; • detecting user intents by means of machine learning algorithms that receive the ID of the virtual assistant and a text message from the user's request as input, as a result of which they form a list of intents and the degree of confidence that these intents are present in the user's message; • search for the most appropriate rule for generating a response based on the user's intents obtained in the previous step. 2. The method according to claim 1, characterized in that the virtual assistant contains a unique assistant identifier and a set of named parameters. 3. The method according to claim 1, characterized in that when intents are detected, if their classification confidence is lower than the configuration parameter, they are discarded. 4. The method according to claim 1, characterized in that, for each intent, the classification confidence is multiplied by the configuration parameter and rounded with an accuracy determined by the configuration parameter. 5. The method according to claim 1, characterized in that after the discovery of intents, they are grouped by the name of the role assigned to them. 6. The method according to claim 1, characterized in that when searching for a suitable rule for generating an answer for each question template and intent included in the DL language template, the weight of its coincidence with the user's message is determined. 7. The method of claim 6, wherein the weight is a 32-bit unsigned integer. 8. The method according to claim 7, characterized in that templates of the same weight are selected by selecting from all found templates-questions and intents with the maximum weight, while the selected templates are removed from the search list. 9. The method according to claim 1, characterized in that when searching for a suitable rule for generating a response

make a list of response templates related to the selected question templates and intents;

check the possibility of generating a non-empty response, where an empty response is a response that does not contain any textual output element;

carry out randomization, and select at random one of the previously selected response templates and determine it as a response to the user's message;

form a response according to the selected rule, and for the selected response template, all functional elements of the DL language are processed in accordance with their description;

write the dialogue log and the generated response is written to the database.

Images