RU2818036C1 - Method and system for controlling dialogue agent in user interaction channel - Google Patents

Abstract

FIELD: physics.

SUBSTANCE: invention relates to the field of computer technologies, in particular to dialogue systems providing processing and intelligent dialogue with the user. Disclosed technical solution is implemented by means of a system for controlling a dialogue agent in a channel for interaction with a user, comprising at least one processor; at least one data storage means; annotator module, selector module and post-processing module. Annotator module receives a user request in text form and processes it in order to determine characteristics of the intention of the user request. Selector module connects the skills of conducting a dialogue with the user, selects the skill of conducting a dialogue with the user based on certain characteristics of the intention of the user request and the associated response to the request. Post-processing module generates a response selected by the selector module, which is transmitted to the user interaction channel.

EFFECT: wider variability of conducting a dialogue with a user.

12 cl, 4 dwg

Description

TECHNICAL FIELD

[0001] This technical solution relates to the field of computer technology, in particular to dialogue systems that provide processing and intelligent dialogue with the user.

BACKGROUND OF THE ART

[0002] One of the most promising areas in the application of modern neural network methods are dialogue systems. Year after year, there are more messenger users than social network users. Also, many companies strive to provide their services through text communication channels, when communication occurs through services in the form of dialogue. Such applications include the following areas:

ensuring the ordering of goods and/or services using users’ personal devices;

support services;

personal assistants, voice assistants, etc.

[0003] Today, there are several approaches to developing conversational agents, or conversational dialogue systems, or conversational agents. Conversational agent development primarily uses a modular architecture for goal-directed dialogue in which a script unfolds. However, the user often needs to combine targeted dialogue, for example, with other functionality, such as answering questions or searching for information, as well as maintaining a conversation. Thus, the ideal conversational agent is a personal assistant that combines different types of agents and switches between its functionality, between its characters, depending on what task it is used for. At the same time, the agent must accumulate information about its essence and adjust its algorithms to a specific user. On the other hand, it must be able to integrate with external services (for example, make queries to external databases, receive information from there, process it, highlight what is important and transfer it to the user).

[0004] A method and system for creating dialog agents using artificial intelligence are known, disclosed in patent application RU 2015155633 A (RISOFTDEV, INC., 08/10/2017). The advantages of this development include the possibility of creating the personality of the dialogue system in the process of conducting a dialogue with the user. In this way, the system learns how to behave and how to respond correctly. However, this system supports only one type of interaction - a cheat-chat conversation and is not able to respond to specific user requests based on processing the parameters of the user's remarks.

[0005] US patent application US 20120016678 A1 (Apple Inc., 01/19/2012) discloses an intelligent automated assistant. The advantages of this development include the possibility of expanding internal functionality to achieve the set goal (goal-oriented dialogue, due to an active ontology), integration into various devices, as well as connecting external services with which the system is compatible. However, this system supports only one type of interaction - to solve a given goal and does not allow the use of other types of agents, i.e. This functionality of the system is limited in the variability of response formation when conducting a dialogue using an intelligent system.

[0006] Patent RU 2637874 C2 (Google Inc. (US), 12/07/2017) discloses a method for generating interactive recommendations for chat information systems. The advantages of this development include the ability to take a comprehensive approach to the question of what type of information may be requested by the user at a particular point in time, and solve it according to possible schemes. However, this dialogue system has limited functionality due to the lack of internal switching of the dialogue system and personality, supported by weak variability in dialogue, which narrows the possibilities for creating different scenarios for dialogue depending on the type of remarks supplied to the input of the system.

DISCLOSURE OF INVENTION

[0007] The technical problem solved with the help of the claimed invention is the creation of a new method and system for expanding the functionality of dialogue, allowing you to switch personalization depending on the user’s intentions and the purpose of communication with the intelligent system, by isolating the required entities from the text stream and storing the interaction history with the user, decide on a response, and perform response generation.

[0008] Such a system uses a number of methods to select the desired response, applies various approaches to processing the input utterance, and is capable of supporting more than one dialogue state.

[0009] The technical result consists in creating a self-learning scalable system that combines various types of agents and provides dynamic switching between functionality related to the dialogue scenario.

[0010] The claimed technical solution is implemented through a dialog agent control system in the user interaction channel, which contains:

at least one processor;

at least one data storage means;

an annotator module that ensures the receipt of a user request in the user interaction channel, wherein the request is presented in text form;

preliminary processing of the user's text request;

classification of replicas in a user's text request;

extracting parameters from said user request and determining characteristics of the intent of the user request; a selector module that provides a choice of skill for conducting a dialogue with the user based on certain characteristics of the intent of the user request;

selecting a response to the user's request associated with the selected skill, wherein the selection of the response is accomplished by processing characteristics of the intent of the user's request; a post-processing module that ensures the generation of a response selected by the selector module;

transmitting the response to the user interaction channel.

[0011] In one particular example of a system implementation, pre-processing of a user request includes at least one of: correcting typos, normalizing text, splitting the request text into individual words, identifying named entities, identifying and classifying intent in replicas of the user request.

[0012] In another particular example of a system implementation, when identifying named entities, information about the position of names in the text and information about the classes corresponding to them is searched.

[0013] In another particular example of a system implementation, after identifying named entities, their subject matter is determined using a classification method, which includes assigning a request to one of the known categories based on semantic content.

[0014] In another particular example of a system implementation, when classifying replicas, they are divided into receiving specific information, performing actions, receiving information in discussion mode, or combinations thereof.

[0015] In another particular example of the implementation of the system, based on the identified characteristics of the intent of the user request and their semantic content, the coefficient of compliance with one or more skills is calculated.

[0016] In another particular example of the implementation of the system, the skill of conducting a dialogue is selected from the group: a targeted answer to a question, an answer in a free form, or maintaining a free conversation.

[0017] In another particular example of a system implementation, the selector module accesses one or more external information systems to select a response to a user request.

[0018] In another particular example of a system implementation, each of the modules is based on one or more machine learning methods.

[0019] The claimed technical solution is also implemented through a computer-implemented method for controlling a dialog agent in a user interaction channel, performed using at least one process and containing the steps of:

receiving a user request in a user interaction channel, the request being presented in text form;

processing the user's text request, during which they classify the replicas in the user's request;

extracting parameters from said user request, based on which the characteristics of the intent of the user request are determined; select the skill of conducting a dialogue with the user based on certain characteristics of the intent of the user request;

selecting a response to the user's request associated with the selected skill, and selecting the response is carried out by processing the characteristics of the intent of the user's request; generate a response; And

transmit the response to the user interaction channel.

[0020] In one of the particular embodiments of the method, the channel of interaction with the user is selected from the group: online chat session, mobile application, multimedia device with a voice assistant, call center.

[0021] In another particular embodiment of the method, at the stage of processing a user request, at least one of the following is performed: correcting typos, normalizing text, breaking the request text into individual words, identifying named entities, identifying and classifying intent in replicas of the user request.

[0022] In another particular embodiment of the method, when identifying named entities, information about the position of names in the text and information about the classes corresponding to them is searched.

[0023] In another particular embodiment of the method, a match factor for one or more skills is calculated from the identified characteristics of the user request intent.

[0024] In another particular embodiment of the method, the dialogue skill is selected from the group: a targeted answer to a question, a free-form answer, or an answer to a question on behalf of a person.

[0025] In another particular embodiment of the method, when generating a response to a user request, one or more external information systems are contacted.

BRIEF DESCRIPTION OF THE DRAWINGS

[0026] FIG. Figure 1 illustrates the overall information architecture of the solution.

[0027] FIG. 2 illustrates the hardware and software architecture of the claimed system.

[0028] FIG. 3 illustrates a block diagram of the implementation of the claimed method for controlling a dialog agent.

[0029] FIG. 4A - Fig. 4B illustrates an example of conducting a dialogue in a messenger chatbot channel using the claimed solution.

IMPLEMENTATION OF THE INVENTION

[0030] In FIG. Figure 1 shows the general information architecture (100) of the claimed technical solution, which includes a user (10), a user device (110) for generating a request, a server (200) that performs the main software logic for processing user requests (10), and an external knowledge base ( 130).

[0031] Computing and computer devices participating in the general architecture of the claimed solution interact with each other through a data transmission network (150), in particular, the Internet.

[0032] Various means can act as a user device (110) for interacting with a dialog agent located on a server (200), for example, a computer, laptop, smartphone, tablet, smart wearable device (watch, bracelet, glasses), smart -TV, vehicle multimedia device, game console, multimedia voice assistant (smart speaker, smart speaker), etc.

[0033] Formation of a user request for dialogue using the device (110) is carried out in one or another interaction channel, for example, a telephone call to a call center, an online chat session on a website or mobile application, a chat with support service, communication with multimedia devices equipped with voice assistant functions, etc. The user interaction channel may vary relative to the implementation and operating principle of the device (110), as well as the software that provides the desired interaction with the conversational agent.

[0034] The data network (150) can be implemented using communication techniques known in the art, for example, WLAN, LAN, PAN, GSM/3G/LTE/5G, satellite communication, TCP/IP connection, etc.

[0035] An external knowledge base (130) can be formed on the basis of a set of data describing a particular subject area and organized in the form of a centralized information repository, for example, a database, etc. For example, the knowledge base (130) can be one or several resources on the Internet that are used to generate a response to user requests (Wikipedia.org website, online telephone directory, etc.).

[0036] In FIG. Figure 2 shows the architectural design of the software and hardware platform of the server (200), which hosts the control system for the dialog agent that processes incoming user requests from devices (110). The server (200) is based on a standardized layout of computing devices and includes components such as: one or more processors (201), RAM (202), data storage (203), input/output interfaces (204), input/output (205) and network communication means (206).

[0037] The processor (201) is configured to perform the program logic and required computational operations required to operate the server (200). The processor (201) executes the necessary machine-readable commands and instructions contained in the RAM (202). The processor (201) (or multiple processors, multi-core processor, etc.) may be selected from a variety of devices commonly used today, for example, from manufacturers such as: Intel™, AMD™, Apple™, Samsung Exynos™, MediaTEK ™, Qualcomm Snapdragon™, etc. The processor or one of the processors used in the server architecture (200) must also include a graphics processor, for example an NVIDIA or Graphcore GPU, the type of which is also suitable for fully or partially implementing the methods for implementing the claimed solution, and can also be used for training and applying models machine learning.

[0038] Random access memory (202), as a rule, is made in the form of RAM and contains the necessary program logic that provides the required functionality. The data storage medium (203) can be in the form of HDD, SSD disks, raid array, flash memory, optical storage devices (CD, DVD, MD, Blue-Ray disks), etc. The means (203) allow long-term storage of various types of information. Interfaces (204) are standard means for connecting and operating multiple devices, for example, USB, RS232, RJ45, LPT, COM, HDMI, PS/2, Lightning, Fire Wire, etc. The choice of interfaces (204) depends on the specific design of the server device (200).

[0039] The following can be used as data input/output means (205): keyboard, joystick, display (touch display), projector, touchpad, mouse, trackball, light pen, speakers, microphone, etc. Network interaction means (206) are selected from devices that provide network reception and transmission of data, for example, an Ethernet card, WLAN/Wi-Fi module, Bluetooth module, BLE module, NFC module, IrDa, RFID module, GSM modem, etc. Using the means (205), the organization of data exchange is ensured via a wired and/or wireless data transmission channel, for example, WAN, PAN, LAN, Intranet, Internet, WLAN, WMAN or GSM.

[0040] The software implementation provided by the server (200) that allows control of the dialog agent is implemented using a set of modules, in particular, an annotator module (2031), a selector module (2032) and a post-processing module (2033).

[0041] As shown in FIG. 3, the claimed method (300) for controlling a dialog agent consists of successive stages carried out using program logic implemented on the server (200).

[0042] In the first stage (301), the user request is initially received and preprocessed. The request coming from the user (10) using the device (110) can be generated in various forms, for example, voice, text, gesture, etc. Regardless of the initial state of the user request by the device (110), it is necessary to convert it into text form, which will be transferred to the annotator module (2031). The conversion to text form may be accomplished by various known means, such as voice-to-text conversion, gesture-to-text conversion, etc.

[0043] The query generated by the user (10) is received by the module (2031) as a block of unstructured text. The annotator module (2031) at step (302) extracts the following data from the received text: concepts, keywords, phrases, classifications and named entities, and marks the received excerpts as annotations. The module (2031) may include multiple text annotation algorithms that may operate in parallel.

[0044] Using the algorithms of the annotator module (203 G), the following is performed:

Text pre-processing;

Classification of a replica, correlation with one of the classes known to the system;

Retrieving request parameters.

[0045] The annotator module (2031) processes the user's statements (10) in natural language, which may include one or more of the following functions implemented by the mentioned algorithms:

Splitting text into words.

Correcting typos (for example, using the DeepPavlov library element database - Automatic spelling correction model).

Replenishing text with morphological features - determining the normal form of words and parts of speech (for example, using elements of the DeepPavlov library - Morphological tagging model)

Definition of named entities (for example, using the element base of the DeepPavlov library - NER model).

Classification of intents (characteristics of intentions) in a replica using simultaneously working approaches: based on examples of phrases and algorithms based on machine learning (for example, using the DeepPavlov library element base - Classification model). Intents include: organizations, time, locations, persons, products, facts, etc.

Filling information “slots” - request parameters passed in the user’s phrase (for example, using the element base of the DeepPavlov library - Slot filling model). When implementing this stage, based on the provided text, named entities (NER) are identified, classified and structured according to the established form (list). For example, full name, passport details, company names, data from the Unified State Register of Legal Entities, etc.

[0046] It should be noted that the above is only an example of the functionality of the annotator module (2031), which can be expanded depending on the required implementation of the technical solution. As a result of the operation of the annotator module (2031), based on keywords tagged with one or more algorithms indicated above, at stage (303) the user's intent (10) is determined based on the processing of his request, as well as its parameters - characteristics that allow a more accurate description intention, or clarify the necessary details of the request.

[0047] The term “intent” must be understood as the format of the required communication specified by the user request and the response that must be generated by the dialog agent control system when processing the user request. In the implementation of the stated solution, the main types of agents for dialogue can be distinguished:

general purpose (chat bots), allowing dialogue in a free form;

task-oriented (goal-oriented systems), forming a specific answer to the user’s question;

capable of conducting a dialogue on any topic (chatterboxes), designed to maintain a conversation.

[0048] As an example, we can consider a request generated by the user (10) received by the server (200) in the following form: “What is the weather today?” The user's intent (10) is to find out the weather, the parameter is the date and location. Additional information, such as user location, current date, etc. may be obtained from services of the user device (110) or from external services with which the user device (110) interacts. At the output, the module (2031) generates an annotated expression with certain characteristics of the intent.

[0049] Next, at step (304), the request text processed by the module (2031) is transferred for processing to the selector module (2032), which determines the next most suitable state. The state is understood as an expression obtained after processing the dialog input information to extract at least one basic meaning corresponding to it and its further interpretation by various software skills that allow processing the basic meaning of the request.

[0050] The selector module (2032) allows you to connect a different number of skills that can be used to process a user request, with each software skill providing a unique state option in relation to others, which expands the variability of dialogue with the user using an automated dialogue system. The base meaning of the user's request (10) represents the key intent determined by the annotator module (2031). For example, as indicated in the example above, find out the weather.

[0051] The selector module (2032) uses the skill algorithms connected to it, for each of which the confidence factor in the answer is calculated regarding the user's intention in the generated request. In particular, when choosing to activate a particular skill, the selector module (2032) can preliminarily determine the states they generate, for example:

skill A - request for additional information, location of the requested weather - “in what place are you interested in the weather?”;

skill B - generate information on the current coordinates of the user’s device (110) - “today it’s 18 degrees Celsius”;

skill C - joke about the weather - “it wouldn’t hurt you to take an umbrella today.” Next, among the responses generated using skills, the selector module (2032) selects the expression that most fully corresponds to the user’s statement (10), based on the obtained coefficient for each of the states generated by a particular skill, as well as an analysis of the semantic content of the user request text. The coefficient is understood as the value that the skill produces after processing the intent; this is the confidence value in the answer generated by each of the connected skills.

[0052] The selector module architecture (2032) may implement the following approaches to intent processing, for example:

The skill of answering a question takes an intention about the world in free form and produces an answer based on its knowledge from Wikipedia (can be implemented based on elements of the DeepPavlov library - ODQA skill and KBQA skill).

The goal-oriented dialogue skill allows you to predict answers (can be implemented based on elements of the DeepPavlov - Goal-oriented library)

The skill of creating a conversational dialogue that allows you to answer questions in free form on behalf of a person (can be implemented based on elements of the DeepPavlov - Chit-chat library). The examples of skills presented above, selected using the selector module (2032), display only an approximate set, which can be expanded depending on the required implementation of the technical solution.

[0053] After calculating the confidence factor of each of the connected skills and the context of the resulting expression, at step (305) the selector module (2032) selects the most relevant state. Moreover, at this stage there may be several states at once if several proposals were submitted to the input.

[0054] The selector module (2023) can access one or more information systems, for example, an external or internal knowledge base (130) or other resources and services to generate the required response. In particular, to geo-location services, websites with weather forecast data, news resources, etc.

[0055] The selected response is passed to the post-processing module (2033) by the selector module (2032), which generates a text response using natural language word matching functions. The generation of a text response occurs due to the receipt of text data in a previously prepared template, in which the most likely characteristics of a given text are selected and/or the most statistically probable responses are made in the form of a replica to a given text. A previously prepared template is understood as text that was obtained by training a skill on test data.

[0056] At step (306), the response is transmitted to the user interaction channel (10), in particular, to the channel in which the original request was generated. The dialogue system of the server (200) generates a response in text form, but it can be converted into the required form, for example, speech (voice assistants), gesture images, video sequences, etc.

[0057] Each of these modules can be implemented on the basis of one or more methods and/or machine learning algorithms based on neural networks, which may include, for example, an error correction method, an error backpropagation method, a support vector machine, a gradient algorithm boosting, random forest method, regression, etc.

[0058] In FIG. Figure 4A shows an example of how the proposed solution works using a chatbot as an example. The user (10) creates a request in the messenger application channel using the device (110), for example, using the Telegram application. The initially generated request contains information characterizing the user's intention (10) to obtain general data regarding establishments providing a certain type of service (cafes, restaurants, snack bars). The dialogue system connects the required skill to process the corresponding user request (10), namely, a skill for goal-oriented dialogue implemented on the DeepPavlov library and a database of restaurants (cafes, snack bars) containing the necessary information to implement a search engine at the user's request (10).

[0059] When recognizing a received request containing general information, the connected skill type initially contains empty slots that must be filled in during the conversation with the user. Thus, at each iteration, the dialogue system generates a clarifying request to determine the user’s intention when communicating and fills the information slots. Such information may be the type of cuisine preferred by the user (10). Upon receipt of the appropriate response, the dialogue system generates subsequent procedures for processing the request, which may contain, for example, clarification of the approximate location of the cafe/restaurant.

[0060] When the information slots are completely filled, the system generates a specific response based on the prepared template (for example, the name of the restaurant). The determination of the place of visit is carried out by the system based on the calculation of the confidence coefficient and, upon affirmative acceptance of the proposal of the dialogue system by the user, the provision of information identifying the selected place of visit. This type of dialogue system can be implemented on the basis of recurrent neural networks trained on a sample of information on the relevant places where food services are provided (cafes, restaurants, bistros, etc.).

[0061] In FIG. 4B shows an example of how the proposed solution works using a chatbot as an example. The user (10) generates a request in the interaction channel using a device (110), for example, using a voice speaker. The initially generated request contains information characterizing the user's intention (10) to communicate with the system on well-known topics (for example, about cars, fashion, etc.). The dialogue system connects the required skill to process the corresponding user request (10), namely, the chit-chat skill, implemented on the DeepPavlov library and systems containing the necessary information to implement a search engine at the user's request (10).

[0062] Instead of following a chat script, they try to find an answer that most closely matches the meaning of the last thing the user said. Most often, the answer is selected from a set of ready-made phrases obtained during training of the skill - with this approach, the bot’s answers do not always correspond to the user’s remarks in meaning, but are competent complete sentences. In this case, it is possible to generate a response from scratch - this is a much more flexible approach, but it is fraught with errors.

[0063] An ordinary chatbot, trained on a corpus of dialogues, answers questions about itself inconsistently - for example, it can say that it does not have animals, and then - that it just walked the dog. If the chatbot doesn’t have any information about itself, it simply generates a line that best matches the user’s previous comment, and does not check it against what it said before. However, to solve this problem, personality and answers to questions about oneself are written into the dialogue system, which allows it to use one or another preset personality.

[0064] The presented description of the claimed solution discloses only preferred examples of its implementation and should not be interpreted as limiting other, particular examples of its implementation that do not go beyond the scope of legal protection, which are obvious to a specialist in the relevant field of technology.

Claims (37)

1. A system for managing a dialog agent in a user interaction channel, containing: at least one processor; at least one data storage means; annotator module providing receiving a user request in a user interaction channel, the request being presented in text form; preliminary processing of the user's text request; classification of replicas in a user's text request; extracting parameters from said user request and determining characteristics of the intent of the user request; selector module providing connecting the skills of conducting a dialogue with the user, allowing you to process the basic meaning of the user request; calculating the confidence coefficient in the answer regarding the user's intention for each of the connected skills of conducting a dialogue with the user; selecting a skill for conducting a dialogue with the user based on certain characteristics of the intent of the user request, and the skill for conducting a dialogue is selected from the group: purposeful answer to a question, answer in a free form, or maintaining a free conversation; selecting a response to a user request related to the selected skill, wherein the choice of answer is carried out by processing the characteristics of the intent of the user request, as well as on the basis of the obtained confidence coefficient for each of the skills of conducting a dialogue with the user and analyzing the semantic content of the text of the user request; post-processing module providing generating a response selected by the module selector; transmitting the response to the user interaction channel. 2. The system according to claim 1, characterized in that the preliminary processing of the user request includes at least one of: correcting typos, text normalization, splitting the request text into individual words, identifying named entities, identifying and classifying intents in user request replicas . 3. The system according to claim 2, characterized in that when identifying named entities, information about the position of names in the text and information about the classes corresponding to them is searched. 4. The system according to claim 2, characterized by the fact that after identifying named entities, their subject matter is determined using a classification method, including assigning a request to one of the known categories based on semantic content. 5. The system according to claim 1, characterized by the fact that when classifying replicas, they are divided into receiving specific information, performing actions, receiving information in discussion mode, or combinations thereof. 6. The system according to claim 1, characterized in that the selector module accesses one or more external information systems to select a response to the user’s request. 7. The system according to claim 1, characterized in that each of the modules is based on one or more machine learning methods. 8. A computer-implemented method for controlling a dialog agent in a user interaction channel, performed using at least one process and comprising the steps of: receiving a user request in a user interaction channel, the request being presented in text form; perform processing of the user's text request, during which they classify replicas in the mentioned user request; extracting parameters from said user request, based on which the characteristics of the intent of the user request are determined; connect the skills of conducting a dialogue with the user, allowing you to process the basic meaning of the user request; calculating the confidence coefficient in the answer regarding the user's intention for each of the connected skills of conducting a dialogue with the user; select a skill for conducting a dialogue with the user based on certain characteristics of the intent of the user request, and the skill for conducting a dialogue is selected from the group: a targeted answer to a question, a free-form answer, or maintaining a free conversation; selecting an answer to the user's request associated with the selected skill, and selecting the answer is carried out by processing the characteristics of the intent of the user's request, as well as on the basis of the obtained confidence coefficient for each of the skills of conducting a dialogue with the user and analyzing the semantic content of the text of the user's request; generate a response; And transmit the response to the user interaction channel. 9. The method according to claim 8, characterized in that the channel of interaction with the user is selected from the group: online chat session, mobile application, multimedia device with a voice assistant, call center. 10. The method according to claim 8, characterized in that at the stage of processing the user request, at least one of the following is performed: correcting typos, normalizing the text, splitting the request text into individual words, identifying named entities, identifying and classifying intents in replicas of the user request. 11. The method according to claim 10, characterized in that when identifying named entities, information about the position of names in the text and information about the classes corresponding to them is searched. 12. The method according to claim 8, characterized in that when generating a response to a user request, an appeal is made to one or more external information systems.