Virtual assistant faq configuration method
Patent Information
- Application Number
- EP2022843376
- Authority / Receiving Office
- EP · EP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-12-28
- Filing Date
- 2022-12-15
- Publication Date
- 2024-11-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing Virtual Assistant platforms face complexity in managing multi-channel responses for FAQs without context, requiring specific configurations for diverse question-and-answer structures, which increases management difficulty and costs, especially in low-code environments.
A method and system that simplifies FAQ configuration using a single repository for storing and training FAQ data sets, allowing configuration through either graphical user interfaces or CSV files, with a template-based CSV file for defining implicit associations between questions and answers, and utilizing a machine learning serving system like TensorFlow for embedding vector generation and indexing.
This approach reduces configuration complexity, enabling faster and simpler management of large FAQ datasets, allowing for a generic configuration interface that supports diverse FAQ types without the need for tailored developments, thus lowering costs and improving usability.
Smart Images

Figure 1.1
Abstract
Description
[0001] VIRTUAL ASSISTANT FAQ CONFIGURATION METHOD
[0002] FIELD OF THE INVENTION
[0003] The present invention is enclosed in the technical area of Natural Language Processing . Speci fically, it belongs to field of conversational arti ficial intelligence and relates to Natural Language Understanding engines , in particular how to feed the information in the use of unsupervised models (pre-trained models ) , and the process of configuring the responses of the Virtual Assistant for FAQs ( Frequent Asked Questions ) / QA ( Question and Answer) questions .
[0004] PRIOR ART
[0005] Arti ficial Intelligence Virtual Assistants are usually created and managed by the System Administrator in a web interface that allows to be defined the Natural Language concepts inherent to the language understanding; namely, 1 ) the questions the user asks , and 2 ) the Virtual Assistant ' s responses to the user utterances . Low code web interfaces are becoming the standard in Virtual Assistant Platforms focused on non-technical users - interfaces designed to be used by anyone . But even in low code environment it is complex to manage the Virtual Assistant ' s multi-channel responses to the user utterances for situations where there is no need a speci fic context to follow a conversation flow . The existent models supervised and unsupervised, by themselves do not facilitate the FAQs configurations . The complexity of management increases with the growing number of subj ects that BOTs need to know and that requires speci fic intentions to understand the user ' s questions .
[0006] PROBLEM TO BE SOLVED
[0007] Developments in Arti ficial Intelligence associated to the Natural Language universe allowed the massi f ication of intelligent Virtual Assistants trying to make easier people li fe . This has created the challenge / pressure of making easy to provision FAQ information in Virtual Assistants , through easy-to-use graphical user interfaces or by other means .
[0008] Virtual Assistant ' s low code Platforms are facing the need to be easier to add knowledge to BOTs . For simple question-and-answer dialogs such as FAQs , where there is no need to collect user information or follow a dialog sequence , managing the Virtual Assistant dialog is challenging, either using a friendly graphical user interface , or by uploading files with pre-prepared FAQ information without the need to add question by question .
[0009] On the other hand, the lack of FAQ generic configuration information models , able to support the large diversity of questions-and-answers information structures , does not allow the construction of a generic configuration interface to apply in a large diversity of FAQ types . This imply that tailored developments for specific FAQs would be required in the configuration interface which would create additional complexity and costs . This is a paradox in the perspective of a low code system . SUMMARY OF THE INVENTION
[0010] The present invention is related to a Virtual Assistant platform ecosystem (here called system) , more speci fically in Virtual Assistant scenarios o f simple question-and-answer ( FAQ) dialogs , where responses to user utterances ( questions ) are configured without the need to have dialogue contexts .
[0011] The invention provides a method and a system to simpli fy system administrators ( those who create and manage virtual assistants ) the task of configuring and provisioning FAQ information using di f ferent means and to train a FAQ dataset to make a FAQ available to end users . For that purpose , a single repository is used to store FAQs configurations , independently of using a graphical user interface or a csv file to configure and provision FAQs configurations which are used by a single training process to train the FAQs datasets . The csv file follows a template , as illustrated in Figure 4 , where is possible to define implicit associations between questions and answers and channels .
[0012] DESCRIPTION OF FIGURES
[0013] Figure 1 - Diagram of FAQ configuration data model , wherein the reference numbers represent :
[0014] • 100 - Question data entity;
[0015] • 101 - Answer data entity;
[0016] • 102 - Channel data entity . Figure 2 - Example of configured questions list with the related information .
[0017] Figure 3 - Example of form to configure an answer .
[0018] Figure 4 - Template of CSV file for FAQ configuration information .
[0019] Figure 5 - Example that illustrates a CSV file for a FAQ Program configuration .
[0020] Figure 6 - Depicts a flow chart describing a method for creating FAQs in virtual assistants , wherein the reference numbers represent :
[0021] • 600 - fills the FAQ configuration information ;
[0022] • 601 - stores the FAQ configuration information in the FAQ configuration repository ( 701 ) ;
[0023] • 602 - initiates FAQ training process ;
[0024] • 603 - converts the FAQ configuration information into a file usable by a machine learning model serving system, such as a TensorFlow Serving;
[0025] • 604 - initiation of the indexing procedure with the generation of embedding vectors for the questions ;
[0026] • 605 - termination of the indexing procedure with the storing of the resulting index obj ect ;
[0027] • 606 - terminates the training process and places the FAQ Program ready to be used .
[0028] Figure 7 - I llustrates a diagram of a modules architecture to create FAQs in a virtual assistant , wherein the reference numbers represent :
[0029] • 700 - FAQ request server ;
[0030] • 701 - FAQs configurations repository; • 702 - message queue broker ;
[0031] • 703 - Semantic Textual Similarity service ;
[0032] • 704 - Network File System storage ;
[0033] • 705 - TensorFlow Serving;
[0034] • 706 - Train Res Worker ;
[0035] DETAILED DESCRIPTION
[0036] Virtual Assistant ' s low-code platforms are facing the need to be easier to manage , particularly when used in FAQ scenarios where large amounts of information need to be managed . The present innovation provides a method and system to surpass the di f ficulty of managing large amounts of FAQs configuration information on Virtual Assistant low-code Systems , allowing to reduce configuration complexity for a faster and simpler configuration experience .
[0037] In a first obj ect of the present disclosure , it is proposed a virtual assistant FAQ configuration method . The method has two main parts to uni fy the management of FAQs configuration information to simpl i fy the FAQs configuration process .
[0038] The method developed represents a sequence of steps applied to FAQs configuration process , which uses a single repository to store the FAQs configuration information and to train FAQs data sets , independently of the interface used to load FAQs configuration .
[0039] Part 1 :
[0040] The first part of the method developed impacts on the procedures to load FAQs configuration, into FAQs configuration information repository, through the available interfaces: the manual interface using a GUI and the batch interface using a csv file.
[0041] It is used, in the present disclosure, the concept of Program FAQ (a FAQ related to a specific theme) , based in unlabeled models, which is a set of questions related with each other in a "linguistic understanding domain", as defined in the Portuguese patent application No. 116967 - "METHOD FOR NATURAL LANGUAGE PROCESSING BASED ON MULTI-MODEL SEGMENTS USING AN HYBRID NLP", whose contents are hereby include by reference. The Program FAQ questions set is a partition of the entire set of questions q (the space of all FAQs) and the partitioning into Program FAQs is performed by the System Administrator (those who create and manage Virtual Assistants) since he has the ability to understand the linguistic domain.
[0042] For each FAQ (Program FAQ) the System Administrator must define the set of questions (user utterances) , as well as the respective responses and the channels available for the responses.
[0043] A single information model, as represented in Figure 1, is used to facilitate the adoption of a single repository to store the FAQ configuration information related to each FAQ Program. This model has 3 main entities that represent the model of a FAQ configuration, and the relevant information is the following:
[0044] • Question (100) - where is defined the user utterance that identify a question, the associated answer, and the capacity to enable / disable these questions for training the data set; • Answer ( 101 ) - where is defined the answer ' s message , the channels available to send it , the questions to which it is associated;
[0045] • Channel ( 102 ) - where is defined the channel to interact with BOT to answer the question .
[0046] In the manual interface a System Administrator fills forms available in GUI pages from the Virtual Assistant management , to configure question by question . A System Administrator has the capacity to manage and configure each FAQ Program, as i llustrated in Figure 2 and 3 by configuring the questions , the associated answers with the respective channels and its formats and questions availability to training .
[0047] In the batch interface a System Administrator uploads a csv file with the entire FAQ configuration in the Virtual Assistant . A System Admini strator populates the csv file with the questions user utterances , the associated answers , and the respective channels . The csv file follows a template , as il lustrated in Figure 4 , where is possible to define implicit associations between questions and answers and channels - the three entities are placed in the same line - and uses speci fic notation - #«answer name» - to indicate that a question shares the same answer to one that has the same answer name . In Figure 5 is presented an example of a CSV file for a FAQ Program configuration .
[0048] Part 2 :
[0049] The second part of the method developed uses the information stored in the FAQs configuration information repository to proceed to the training of FAQs datasets . The FAQ training process is composed by some activities which requires some degree of orchestration to coordinate the order of their execution and uses a message broker to ensure the chronological execution.
[0050] The machine learning serving system, such as a TensorFlow Serving, is used to generate embedding vectors for each question of a FAQs dataset. These embedding vectors are indexed and stored in an index object that is stored and used later at parsing time.
[0051] At parsing time, when the user's question arrives, it is generated an embedding vector of the user utterance and the index object previously stored is used to find the most similar FAQ previous indexed.
[0052] DESCRIPTION OF THE EMBODIMENTS
[0053] In the following an embodiment of the present invention being related to a generic method for FAQs configuration in Virtual Assistant scenarios will be described with respect to Figure 6. Figure 6 shows a flow describing a generic method for FAQs configuration in Virtual Assistant scenarios: i. In a first step 600, the System Administrator fills the FAQ's configuration information by using one of the interfaces: 1) in the batch interface it will populate a FAQ's csv file with the FAQ's configuration information, according to a CSV template file, as represented in Figure 4; 2) in the manual interface it will fill the GUI forms with the FAQ's configuration information question by question as illustrated in Figure 2 and 3. ii. In the next step 601, the System Administrator stores the FAQ's configuration information into the FAQs configuration repository (701) , any missing information will be stored with default values, using one of the two following options: 1) in the batch interface it will uploads the FAQ's csv file; 2) in the manual interface it will save the filled forms . iii. In the next step 602, the System Administrator initiates the training process of a FAQ dataset by sending a FAQ training event to the FAQ Request server (700) which will place the training event in the message queue broker (702) to be processed asynchronously . iv. In the next step 603, the Semantic Textual Similarity service (703) consumes the FAQ training event from the message queue broker (702) and requests to the FAQ request server (700) to retrieve the CSV FAQ configuration information file, from the FAQs configurations repository (701) , and the Semantic Textual Similarity service (703) converts the retrieved CSV file into a new CSV file usable by the TensorFlow Serving (705) and store it in the Network File System storage (704) . v. In the next step 604, the Semantic Textual Similarity service (703) initiates the indexing procedure by requesting the TensorFlow Serving (705) to generate embedding vectors for each question (user utterance) of the FAQ Program. vi . In the next step 605, the Semantic Textual Similarity service (703) uses the returned embedding vectors list to generate an index object, based on the vectors, and stores it in the Network
[0054] File System storage (704) . Terminates the indexing procedure by placing an event of indexing end in the message queue broker ( 702 ) . vii . In the last step 606 , the Train Res Worker ( 706 ) consumes the indexing end event to update the FAQ Program to be ready for use in the Network File System storage ( 704 ) ending the training process .
[0055] It is also disclosed a virtual assistant FAQ configuration system . In a preferred embodiment of the system, it is comprised by
[0056] A system administrator ;
[0057] A FAQ Request server ;
[0058] A FAQs configurations repository;
[0059] A message queue broker module ;
[0060] A machine learning serving system;
[0061] A train Res Worker module ;
[0062] Wherein the system administrator being further configured to implement a semantic textual similarity service .
[0063] In another embodiment of the system, the FAQs configurations repository is configured to store configuration information related to a FAQ program according to a single information model .
[0064] In another embodiment of the system, the single information model comprises at least three entities representing the model of a FAQ configuration and relevant information :
[0065] Question : configured to define a user utterance that identi fy a question, the associated answer and the capacity to enable / disable a question for training a data set ; Answer: configured to define the answer's message, the channels available to send the answer and the questions to which it is associated.
[0066] Channel: configured to define the channel to interact with BOT to answer the question.
[0067] In another embodiment of the system, the machine learning serving system is a TensorFlow serving.
Claims
CLAIMS1. Virtual Assistant FAQ configuration method, comprising the steps of:I. Filling FAQ configuration information related to at least a FAQ program;II. Storing FAQ configuration information in a FAQ configuration repository;III. Initiating FAQ training process;IV. Converting FAQ configuration information into a file usable by a machine learning model serving system;V. Initiation of the indexing procedure with the generation of embedding vectors for questions ;VI. Termination of the indexing procedure by storing the resulting index object;VII. Termination of the FAQ training process and placing the FAQ program ready to be used.
2. Method according to claim 1, wherein the FAQ configuration information uses a single information model to collect configuration information related to each FAQ program.
3. Method according to claim 2, wherein the single information model has at least three entities:— Question: to define a user utterance that identify a question, the associated answer and the capacity to enable / disable a question for training a data set;— Answer: to define the answer's message, the channels available to send the answer and the questions to which it is associated.— Channel: to define the channel to interact with BOT to answer the question.
4. Method according to any of the previous claims, wherein the step I is implemented by the system administrator, by using at least one of the following interfaces :— A batch interface, populating a FAQ's csv file with FAQ' s configuration information according to a csv template file;— A manual interface by filling GUI forms.
5. Method according to any of the previous claims, wherein the step II is implemented by the system administrator, and wherein, missing information is stored with default values using at least one of the following options :— In a bath interface, by uploading a FAQ's csv file;— In a manual interface by saving the filled forms .
6. Method according to any of the previous claims, wherein the step III is implemented by the system administrator, and includes:— Sending a FAQ training event to a FAQ Request server;— Placing the training event in a message queue broker to be processed asynchronously.7 . Method according to claim 6 , wherein the step IV is implemented by a semantic textual similarity service , and includes :— Consuming the FAQ training event from the message queue broker ;— Requesting the FAQ Request server to retrieve the csv FAQ configuration information file , from the FAQ configurations repository;— Converting the retrieved csv file into a new csv file usable by machine learning model serving system;— Storing the new csv file in a network file system storage .8 . Method according to according to any of the previous claims , wherein the step V is implemented by a semantic textual similarity service , and includes :— Requesting the machine learning model serving system to generate embedding vectors for each question of the FAQ program;— Generation of embedded vectors by the machine learning model serving system .
9. Method according to claim 8 , wherein step VI is implemented by a semantic textual similarity service , and includes :— Generating an index obj ect based on a list of embedded vectors ;— Storing the index obj ect in the network file system storage ;— Terminating the indexing procedure by placing an event of indexing end in a message queue broker .1410 . Method according to any of the previous claims , wherein the step VI I including :— Consuming the indexing end event ;— Updating the FAQ program to be ready for use in the network file system storage .11 . Virtual Assistant FAQ configuration system configured to implement the method of the claims 1 to 10 ; the system comprising :A system administrator ;A FAQ Request server ;A FAQs configurations repository;A message queue broker module ;A machine learning serving system;A train Res Worker module ;Wherein the system administrator being further configured to implement a semantic textual similarity service .12 . System according to claim 11 , wherein the FAQs configurations repository is configured to store configuration information related to a FAQ program according to a single information model .
13. System according to claim 12 , wherein the single information model comprises at least three entities representing the model of a FAQ configuration and relevant information :— Question : configured to define a user utterance that identi fy a question, the associated answer and the capacity to enable / disable a question for training a data set ;15— Answer: configured to define the answer's message, the channels available to send the answer and the questions to which it is associated .— Channel: configured to define the channel to interact with BOT to answer the question.
14. System according to claim 11, wherein the machine learning serving system is a TensorFlow serving.16
