CN110489517B - Automatic learning method and system of virtual assistant - Google Patents

Abstract

An automatic learning method and system for virtual assistants. The virtual assistant's automatic learning method includes: receiving audio input and identifying the audio to form corpus data; using natural language processing models to analyze the corpus data to generate language feature information corresponding to the corpus data; and performing functional contextualization of the language feature information based on the functional context information. Analysis to determine the operation corresponding to one of these intentions; if the functional situation analysis cannot determine the operation corresponding to one of these intentions, perform word segmentation processing on the corpus data; based on the results of the word segmentation processing, determine whether there is a new vocabulary Or new corpus data; if new vocabulary exists, the natural language processing model is updated according to the meaning of the new vocabulary; if new corpus data exists, the functional situation analysis is updated according to the intention of the new corpus data. In this way, users can use the ERP system more quickly and conveniently.

Description

Automatic learning method and system for virtual assistant

Technical field

This case relates to an automatic learning method and system, and in particular, to an automatic learning method and system for a virtual assistant.

Background technique

Enterprise Resource Planning (ERP), referred to as ERP system, refers to a management platform based on information technology that provides decision-making for enterprise decision-makers. Its main purpose is to uniformly manage the flow of people, logistics, information, and capital of the enterprise to maximize the use of the company's resources. The ERP system includes three major functions: production control, logistics management and financial management, so the scale of the ERP system is very large.

Applying virtual assistants to ERP systems can quickly help users communicate with the huge ERP system, and can save users the time spent using the ERP system. However, due to the different habits of each user using the ERP system, Therefore, there will be situations where the virtual assistant cannot understand the user's problem, which in turn causes users difficulty in using the ERP system.

Contents of the invention

The main purpose of the present invention is to provide an automatic learning method and system for a virtual assistant, which mainly allows the virtual assistant to have an automatic learning function, so that the virtual assistant can automatically learn the user's characteristics in the process of communicating with the user. Speaking habits, or special terms used in the industry, can make users use the ERP system faster and more conveniently.

In order to achieve the above purpose, the first aspect of this project is to provide an automatic learning method for virtual assistants. This method includes the following steps: receiving audio input and identifying the audio to form corpus data; using a natural language processing model to analyze the corpus data to Generate language feature information corresponding to the corpus data, where the language feature information includes multiple intentions, probabilities corresponding to the multiple intentions, and multiple words; perform functional context analysis on the language feature information based on the functional context information, and determine the multiple The operation corresponding to one of the intentions; if the functional situation analysis cannot determine the operation corresponding to one of the multiple intentions, perform word segmentation processing on the corpus data; according to the result of the word segmentation processing, determine whether there is a new vocabulary or New corpus data; if there is new vocabulary, update the natural language processing model according to the meaning of the new vocabulary; if there is new corpus data, update the functional situation analysis according to the intention of the new corpus data; among them, the operations include query data operations and execution instruction operations. one.

According to an embodiment of the present case, it also includes: generating a system domain vocabulary set based on an application knowledge database and a domain knowledge database; the system domain vocabulary set and multiple service application parameters form a key entity set, and the key entity set includes multiple system domain vocabulary; classifying multiple training corpora into one of the query data operation and the execution instruction operation; distinguishing the multiple training corpora corresponding to the query data operation according to the categories in the enterprise database to form multiple A query data operation intention, and the intention of distinguishing the plurality of training corpora corresponding to the execution instruction operation according to the service behavior provided by the enterprise resource system to form multiple execution instruction operation intentions; establishing a template for the plurality of query data operation intentions , and the templates of the plurality of execution instruction operation intentions; establishing the overall database according to the key entity set, the plurality of query data operation intention templates and the plurality of execution instruction operation intention templates; identifying the key entity set A plurality of first probabilities that the plurality of system domain words appear in the plurality of training corpora, and the plurality of sentence structures of the plurality of training corpora are analyzed through the identified plurality of system domain words. , and multiple correlations between the plurality of system domain words, and establishing a common vocabulary model based on the plurality of first probabilities and the plurality of correlations; and analyzing the plurality of query data operation intentions and a plurality of second probabilities that the plurality of system domain words appear in the plurality of execution instruction operation intentions, and a common semantic model is established based on the plurality of sentence structures and the plurality of second probabilities.

According to an embodiment of the present case, it also includes: using a classifier to classify the data in a historical database into strong and weak relationships to generate a functional situation model; and segmenting and analyzing the plurality of training corpora, and analyzing the data according to the history The data in the database generates a functional vocabulary model.

According to an embodiment of this case, the functional situation analysis also includes: using the corpus data and the functional situation information to compare with the functional situation model, and generating a functional situation identification result; and judging the multiple functions based on the functional situation identification result. One of the intentions corresponds to one of the query data operation and the execution instruction operation.

According to an embodiment of the present case, the word segmentation processing further includes: segmenting the corpus data according to the functional vocabulary model to generate multiple word segments; and calculating the frequencies of the multiple word segments.

According to an embodiment of this case, it also includes: determining whether the frequencies of the plurality of word segments calculated by the word segmentation processing are lower than a threshold; if one of the plurality of word segments is lower than the threshold, the plurality of word segments are One of the word segments is the new vocabulary, and the definition of the new vocabulary is received to update the common vocabulary model and the common semantic model; and if the plurality of word segments are higher than the threshold, then the corpus data is The new corpus data is received, and the intention of receiving the new corpus data is used to update the functional situation model.

According to an embodiment of the present case, the method further includes: determining whether the new corpus data is common corpus, and if so, updating the system domain vocabulary set according to the new corpus data; and updating the system domain vocabulary set according to the new vocabulary.

According to an embodiment of the present case, the natural language processing model analyzing the corpus data further includes: using the common vocabulary model to identify whether the corpus data contains the plurality of system domain words in the key entity set, and setting the identification results for the plurality of words, and analyze the probability of occurrence of the plurality of words; analyze the sentence structure of the corpus data according to the plurality of words; and use the common semantic model according to the probability of occurrence of the plurality of words and the The sentence structure of the corpus data identifies the plurality of intentions of the corpus data and the probabilities corresponding to the plurality of intentions.

The second aspect of this case is to provide an automatic learning system for a virtual assistant, which is connected to an enterprise database and an enterprise resource system respectively, and includes: a processor, a storage device, and an input/output device. The storage device is electrically connected to the processor and is used to store the overall database, application knowledge database, domain knowledge database and historical database. The input/output device is electrically connected to the processor and provides an interface for inputting audio. Among them, the processor includes: a speech recognition module, a corpus analysis module, a situation recognition module, an unknown corpus judgment module, and an update information module. The speech recognition module is used to recognize audio to form corpus data. The corpus analysis module is electrically connected to the speech recognition module, and is used to analyze the corpus data using a natural language processing model to generate language feature information corresponding to the corpus data, where the language feature information includes multiple intentions and probabilities corresponding to the multiple intentions. and multiple vocabulary words. The situation identification module is electrically connected to the corpus analysis module, and is used to perform functional situation analysis on the language feature information based on the functional situation information, and determine the operation corresponding to one of the plurality of intentions. The unknown corpus judgment module is electrically connected to the situation identification module, and is used to perform word segmentation processing on the corpus data when the situation identification module cannot identify the operation corresponding to one of the plurality of intentions, and based on the result of the word segmentation processing, Determine whether there are new vocabulary or new corpus data. The update information module is electrically connected to the unknown corpus judgment module, and is used to update the natural language processing model according to the meaning of the new vocabulary when a new vocabulary is generated, and to update the natural language processing model according to the intention of the new corpus data when the new corpus data is generated. The functional scenario analysis is updated; wherein the operation includes one of a data query operation and an instruction execution operation.

According to an embodiment of the present case, the processor further includes: a training module, electrically connected to the corpus analysis module, for generating a system domain vocabulary set based on the application knowledge database and the domain knowledge database. The system domain vocabulary set and Multiple service application parameters are formed into a key entity set. The key entity set includes multiple system domain vocabularies, and multiple training corpora are classified into one of the query data operation and the execution instruction operation. According to the enterprise database Classify the intentions of the plurality of training corpora corresponding to the query data operation to form multiple query data operation intentions, and differentiate the intentions of the plurality of training corpora corresponding to the instruction operation according to the service behavior provided by the enterprise resource system Form multiple execution instruction operation intentions; a template creation module, electrically connected to the training module, establishes the templates of the plurality of query data operation intentions, and the templates of the multiple execution instruction operation intentions, according to the key entity set , the plurality of query data operation intention templates and the plurality of execution instruction operation intention templates establish the overall database; a vocabulary model establishment module is electrically connected to the template establishment module to identify all the key entity sets in the key entity set. A plurality of first probabilities of the plurality of system domain words appearing in the plurality of training corpora, and a plurality of sentence structures of the plurality of training corpora are analyzed through the identified plurality of system domain words, and the Describe multiple correlations between multiple system domain words, and establish a common vocabulary model based on the multiple first probabilities and the multiple correlations; and a semantic model building module, electrically connected to the template building module sexual connection, analyze the plurality of second probabilities that the plurality of system domain words appear in the plurality of query data operation intentions and the plurality of execution instruction operation intentions, and based on the plurality of sentence structures and the plurality of execution instruction operation intentions, A second probability is to establish a common semantic model.

According to an embodiment of the present case, the processor further includes: a situation training module, electrically connected to the situation analysis module, for using a classifier to classify the data in the historical database into strong and weak relationships to generate a functional situation model. ; And a vocabulary training module, electrically connected to the unknown corpus judgment module, used to segment and analyze the plurality of training corpus, and generate a functional vocabulary model based on the data in the historical database.

According to an embodiment of the present case, the situation analysis module is further used to compare the corpus data and the functional situation information with the functional situation model, generate a functional situation identification result, and judge the multiple functions based on the functional situation identification result. One of the intentions corresponds to one of the query data operation and the execution instruction operation.

According to an embodiment of the present case, the unknown corpus judgment module is further used to segment the corpus data according to the functional vocabulary model to generate a plurality of word segments and calculate the frequencies of the plurality of word segments.

According to an embodiment of the present case, the update information module is further used to determine whether the frequencies of the plurality of word segments calculated by the word segmentation processing are lower than a threshold; if one of the plurality of word segments is lower than the threshold, One of the plurality of word segments is the new vocabulary, and the definition of the new vocabulary is received to update the common vocabulary model and the common semantic model; if the plurality of word segments are higher than the threshold, then the The corpus data is the new corpus data, and the intention of the new corpus data is received to update the functional situation model.

According to an embodiment of the present case, the update information module is further used to determine whether the new corpus data is common corpus, and if so, update the system domain vocabulary set according to the new corpus data; and update the system domain vocabulary set according to the new vocabulary.

According to an embodiment of the present case, the corpus analysis module is further used to use the common vocabulary model to identify whether the corpus data contains the plurality of system domain words in the key entity set, and set the identification result to the plurality of system domain words in the key entity set. Vocabulary, and analyze the probability of occurrence of the multiple vocabulary, analyze the sentence structure of the corpus data based on the multiple vocabulary, and use the common semantic model to analyze the probability of occurrence of the multiple vocabulary and the sentence pattern of the corpus data The structure identifies the plurality of intentions of the corpus data and the probabilities corresponding to the plurality of intentions.

The automatic learning method of the virtual assistant and the automatic learning system of the virtual assistant of the present invention mainly allow the virtual assistant to have an automatic learning function, so that the virtual assistant can automatically learn the user's speaking habits during the process of communicating with the user, or It is a special term used in the industry to make users use the ERP system faster and more conveniently.

Description of the drawings

In order to make the above and other objects, features, advantages and embodiments of the present invention more apparent and understandable, the accompanying drawings are described as follows:

Figure 1 is a schematic diagram of an automatic learning system for a virtual assistant according to some embodiments of the present case;

Figure 2 is a schematic diagram of a processor according to some embodiments of the present application;

Figure 3 is a flow chart of an automatic learning method for a virtual assistant according to some embodiments of the present case;

Figure 4 is a flow chart of a training data model according to some embodiments of this case;

Figure 5 is a flow chart of step S320 according to some embodiments of the present case;

Figure 6 is a flow chart of step S330 according to some embodiments of the present case;

Figure 7 is a flow chart of step S340 according to some embodiments of this case; and

Figure 8 is a flowchart of step S360 according to some embodiments of the present application.

Detailed ways

The following disclosure provides many different embodiments or illustrations for implementing various features of the invention. Particular illustrations of components and arrangements are used in the following discussion to simplify the present disclosure. Any examples discussed are for illustrative purposes only and do not limit in any way the scope and significance of the invention or its examples. In addition, this disclosure may repeatedly refer to numerical symbols and/or letters in different examples. These repetitions are for simplicity and explanation, and do not themselves specify the relationship between different embodiments and/or configurations in the following discussion.

Unless otherwise noted, the terms used throughout the specification and claims generally have their ordinary meanings as used in the field, in the disclosure and in the specific context. Certain terms used to describe the present disclosure are discussed below or elsewhere in this specification to provide those skilled in the art with additional guidance in describing the present disclosure.

As used herein, "coupling" or "connected" may refer to two or more elements that are in direct physical or electrical contact with each other, or that are in indirect physical or electrical contact with each other, and "coupled" or "connected" "Connected" can also refer to the mutual operation or action of two or more elements.

It will be understood that the terms first, second, third, etc. are used herein to describe various elements, components, regions, layers and/or blocks. However, these elements, components, regions, layers and/or blocks should not be limited by these terms. These terms are limited to identifying a single element, component, region, layer and/or block. Therefore, a first element, component, region, layer and/or block below can also be termed as a second element, component, region, layer and/or block without departing from the spirit of the present invention. As used herein, the term "and/or" includes any combination of one or more of the associated listed items. The "and/or" mentioned in this document refers to any one, all or any combination of at least one of the listed elements.

See Figure 1. Figure 1 is a schematic diagram of an automatic learning system 100 for a virtual assistant according to some embodiments of the present application. As shown in FIG. 1 , the virtual assistant automatic learning system 100 includes a processor 110 , a storage device 130 and an input/output device 150 . The storage device 130 is used to store the overall database 131 , the application knowledge database 132 , the domain knowledge database 133 and the history database 134 . The storage device 131 , the application knowledge database 132 , the domain knowledge database 133 and the history database 134 are electrically connected to the processor 110 . The input/output device 150 is electrically connected to the processor 110 to provide an interface for inputting audio. In one embodiment, the input/output device 150 may be a keyboard, a touch screen, a microphone, a speaker, or other suitable input/output devices. Users can input audio through the interface provided by the input/output device.

In various embodiments of the present invention, the processor 110 may be implemented as an integrated circuit such as a microcontroller, a microprocessor, a digital signal processor, or an application specific integrated circuit. ASIC), logic circuits or other similar components or a combination of the above components. The storage device 150 can be implemented as a memory, a hard disk, a pen drive, a memory card, etc.

Please refer to FIG. 2. FIG. 2 is a schematic diagram of a processor 110 according to some embodiments of the present application. The processor 110 includes a speech recognition module 111, a corpus analysis module 112, a situation recognition module 113, an unknown corpus judgment module 114, an update information module 115, a training module 121, a template establishment module 122, a semantic model establishment module 123, and a vocabulary model establishment module 124. , situation training module 125 and vocabulary training module 126. The corpus analysis module 112 is electrically connected to the speech recognition module 111, the situation recognition module 113 is electrically connected to the corpus analysis module 112, the unknown corpus judgment module 114 is electrically connected to the situation judgment module 113, and the update information module 115 is to the unknown corpus judgment module 114. Electrical connection. The training module 121 is electrically connected to the corpus analysis module 112, the template building module 122 is electrically connected to the training module 121, the semantic model building module 123 and the vocabulary model building module 124 are electrically connected to the template building module 122, and the situation training module 125 is connected to the situation The recognition module 113 is electrically connected, and the unknown corpus judgment module 114 and the vocabulary training module 126 are electrically connected.

Please refer to Figures 1 to 3 together. Figure 3 is a flow chart of an automatic learning method 300 for a virtual assistant according to some embodiments of the present case. As shown in Figure 3, the automatic learning method 300 of the virtual assistant includes the following steps:

Step S310: Receive audio input and recognize the audio to form corpus data;

Step S320: Use a natural language processing model to analyze the corpus data to generate language feature information corresponding to the corpus data;

Step S330: Perform functional situation analysis on the language feature information based on the functional situation information, and determine the operation corresponding to one of these intentions;

Step S340: If the functional situation analysis cannot determine the operation corresponding to one of these intentions, perform word segmentation processing on the corpus data;

Step S350: According to the result of word segmentation processing, determine whether there is new vocabulary or new corpus data; and

Step S360: If there is a new vocabulary, update the natural language processing model according to the meaning of the new vocabulary. If there is new corpus data, update the functional situation analysis according to the intention of the new corpus data.

In step S310, audio input is received and the audio is recognized to form corpus data. In one embodiment, the audio received through the input/output device 150 can be speech recognized by the speech recognition module 111 of the processor 110 to convert the user's natural language into corpus data. In another embodiment, speech recognition can also transmit the audio to the cloud speech recognition system through the Internet. After the cloud speech recognition system recognizes the audio, the recognition result is used as corpus data. For example, the cloud speech recognition system can Implemented as Google's speech recognition system.

Before executing step S320, a common vocabulary model and a common semantic model need to be established first. Therefore, please refer to Figure 4, which is a flow chart of a training data model according to some embodiments of the present application. As shown in Figure 4, the training data model phase includes the following steps:

Step S410: Generate a system domain vocabulary set according to the application knowledge database and the domain knowledge database;

Step S420: The system domain vocabulary set and multiple service application parameters are formed into a key entity set;

Step S430: Classify multiple training corpora into one of query data operations and instruction execution operations;

Step S440: Differentiate the intentions of the training corpus corresponding to the query data operation according to the categories in the enterprise database to form multiple query data operation intentions, and differentiate the intentions of the training corpus corresponding to the execution of the instruction operation according to the service behavior provided by the enterprise resource system to form multiple An intention to execute the instruction operation;

Step S450: Establish a template of the query data operation intention and a template of the instruction operation intention;

Step S460: Establish an overall database based on the key entity set, the template of the query data operation intention, and the template of the instruction operation intention;

Step S470: Identify multiple first probabilities that the system domain vocabulary in the key entity set appears in the training corpus, and analyze multiple sentence structures of the training corpus through the identified system domain vocabulary, as well as the relationships between the system domain vocabulary and each other. Multiple correlations, and establish a common vocabulary model based on the first probability and correlation; and

Step S480: Analyze multiple second probabilities of system domain words appearing in the query data operation intention and the execution instruction operation intention, and establish a common semantic model based on the sentence structure and the second probabilities.

In step S410 and step S420, a system domain vocabulary set is generated according to the application knowledge database 132 and the domain knowledge database 133, and then a key entity set is formed using the system domain vocabulary set and multiple service application parameters. The key entity set includes multiple system domains. vocabulary. For example, the key entity set contains information such as enterprise domain vocabulary and service application parameters of the enterprise system. Enterprise domain vocabulary refers to the vocabulary that companies in different fields may need to use. For example, the vocabulary used in the medical industry must be different from the vocabulary used in the transportation industry. Therefore, the enterprise domain vocabulary will be based on each company using the ERP system. It varies from company to company. The service application parameters of the enterprise system are the parameters corresponding to the services provided by the enterprise system. For example, the leave function in the enterprise system may require leave time, leave type and other information, and the system domain vocabulary in the key entity set will need Contains personal leave, annual leave, sick leave, business trip leave and other information.

In detail, the key entity set also includes the data field names that will be used when accessing data, the service names provided by the enterprise system to users, the parameter values of the restrictions set by users when querying, and the parameters of service applications. Values and operating functions of the enterprise system, etc. The operating functions of the enterprise system can be leave, overtime application, business trip application, expense reimbursement and other operating functions. The above information may also have corresponding aliases, which need to be input when training the database. For example, a shipping order may have different names such as shipping details or sales orders for manufacturers in specific fields.

In step S430, multiple training corpora are classified into one of query data operations and instruction execution operations. The training corpus can be natural language data such as instructions the user may give or questions he may ask. After the key entity set is established, the training corpus will be classified according to intention. In one embodiment, the user's intention is divided into Query data operations and execute command operations, but the user's intentions can also be classified more precisely, and the present invention is not limited to this. For example, if the user says to the virtual assistant: "Please help me find the shipment order of XX company", it will be classified as a query data operation in the intention classification of the present invention, and the virtual assistant will go to the enterprise database to help the user Check the shipment order of XX company. If the user says to the virtual assistant: "Please help me ask for leave for a business trip on January 30th", it will be classified as executing the command operation in the intention classification of the present invention, and the virtual assistant will enter the enterprise resource system to help the user ask for leave.

In step S440, the intentions of the training corpus corresponding to the query data operation are distinguished according to the categories in the enterprise database to form multiple query data operation intentions, and the intentions of the training corpus corresponding to the execution of the instruction operation are distinguished according to the service behavior provided by the enterprise resource system. Form multiple execution instruction operation intentions. In one embodiment, the intent of the query data operation is first distinguished according to the enterprise database in each different field. For example, an enterprise database in the medical industry will store different data fields than an enterprise database in the transportation industry, so the user needs of the two may not be the same. For example, users in the medical industry may query medical record data, query ward vacancies, etc., which are all query data operations. Users in the transportation industry may query shipping records, query package delivery status, etc., which are all queries. Different intentions for data operations. Of course, the intention of executing the instruction operation will also be distinguished according to the service behavior provided by the enterprise resource system in each different field. As mentioned above, the services provided by the enterprise resource system in the medical industry will of course be different from those in the transportation industry. Each different field The query data operations or service behavior operations provided by enterprises may not necessarily be universal. Therefore, it is also necessary to distinguish the intentions of the services provided by enterprises in different fields. For example, users in the medical industry may provide registration services. , providing services such as ordering healthy meals for hospitalization, etc. are all different intentions of service behavior operations. For users in the transportation industry, there may be services such as automatically classifying goods, arranging services for shipment of goods, etc., which are all different intentions of service behavior operations. .

In steps S450 and S460, a template of the query data operation intention and a template of the instruction operation intention are established, and the overall database 131 is established based on the key entity set, the template of the query data operation intention, and the template of the instruction operation intention. For example, after distinguishing the user's intention to query data and execute instructions when operating a virtual assistant of a company in a certain field, a corresponding template can be generated for each intention. According to the example above, medical The industry will have four templates for querying medical record data, querying ward vacancies, providing registration services, and providing services for ordering healthy meals in hospital. The transportation industry will have corresponding queries for shipping records, querying package delivery status, and providing automatic classification of goods. 4 templates for services and services for arranging goods shipment sequence, and then the overall database 131 will be established based on these templates and key entity sets.

In step S470, multiple first probabilities of the system domain vocabulary appearing in the training corpus in the key entity set are identified, and multiple sentence structures of the training corpus are analyzed through the identified system domain vocabulary, as well as the relationship between the system domain vocabulary and each other. multiple correlations between them, and establish a common vocabulary model based on the first probability and correlation. In one embodiment, two algorithms, n-gram (n-GRAM) and context-free grammar (CFG), are used to calculate the probability of each system domain vocabulary appearing in the training corpus, and the system domain vocabulary is used to calculate the probability of occurrence of each system domain vocabulary in the training corpus. Analyze the sentence structure of the training corpus and the correlation between vocabulary in the system domain to establish a common vocabulary model. For example, if the training corpus contains "I want to inquire about the quotation of XX company" and "I want to inquire about the shipping order of XX company", and "XX company", "quotation" and "shipping order" are all System domain vocabulary, but in the above example, since "XX Company" may appear on average in each intention of querying data operations, the probability of "XX Company" is almost the same in each intention of querying data operations, and "Quotation" and "Shipping Order" only appear in large numbers in the training corpus for the purpose of querying certain specific data, and will not appear in the training corpus for the purpose of querying other data. Therefore, "Quotation" and "Shipping Order" The probability of "Manifest Order" will be particularly high in the corresponding intent, and lower in other intents.

In step S480, a plurality of second probabilities of system domain vocabulary appearing in the query data operation intention and the execution instruction operation intention are analyzed, and a common semantic model is established based on the sentence structure and the second probabilities. In one embodiment, a Hidden Markov Model (HMM) algorithm is used to calculate the probability of system domain words appearing in the query data operation intention and the instruction operation intention to establish a common semantic model. For example, In the training data model stage, a lot of training corpus is input, and the hidden Markov model algorithm must calculate the probability of occurrence of words in the system domain for different purposes. Combined with the above example, if there is "I want to query the shipment order of XX company" in the training corpus, according to n-gram and context-free grammar, it can be found that "XX company" and "shipment order" are both system domain words, and The hidden Markov model algorithm can further determine the "shipping order" and query shipping data based on the probability of all identified system domain words in the query data operation intention and the execution instruction operation intention, as well as the relationship between the system domain words. Associated with the intention, combined with the system domain vocabulary of "XX company", it can automatically help users query XX company's shipment-related data in the enterprise database.

After the common vocabulary model and the common semantic model are established, step S320 is then performed to analyze the corpus data using the natural language processing model to generate language feature information corresponding to the corpus data. The language feature information includes multiple intentions, probabilities corresponding to the intentions, and Multiple words. For the detailed process of step S320, please refer to Figure 5. Figure 5 is a flow chart of step S320 according to some embodiments of this case. As shown in Figure 5, step S320 includes the following steps:

Step S321: Use the common vocabulary model to identify whether the corpus data contains system domain vocabulary that matches the key entity set, set the identification result to the vocabulary in the language feature information, and analyze the probability of occurrence of the vocabulary in the language feature information;

Step S322: Analyze the sentence structure of the corpus data according to the vocabulary in the feature information; and

Step S323: Use the common semantic model to identify the intention of the corpus data and the probability corresponding to the intention based on the probability of word occurrence in the feature information and the sentence structure of the corpus data.

In steps S321 and S322, the common vocabulary model is used to identify whether the corpus data contains system domain vocabulary that matches the key entity set, the identification result is set to the vocabulary in the language feature information, and the occurrence of the vocabulary in the language feature information is analyzed probability, and then analyze the sentence structure of the corpus data based on the vocabulary in the feature information. For example, the corpus data input by the user is used to identify words containing system domain vocabulary in the corpus data using a common vocabulary model, and then the sentence structure of the corpus data is further determined. For example, if the user says to the virtual assistant: "I want to check the shipment order of XX company last month," the common vocabulary model can identify "XX company", "last month" and "shipment order" ” and other words that match the vocabulary of the system domain.

In step S323, the common semantic model is used to identify the intention of the corpus data and the probability corresponding to the intention according to the probability of word occurrence in the feature information and the sentence structure of the corpus data. According to the example above, after identifying words such as "XX company", "last month" and "shipping note", the probability of these words in all intentions will be further determined. All intentions referred to here include all query data operation intentions and the probability of executing instruction operation intentions.

In step S330, functional context analysis is performed on the language feature information based on the functional context information, and the operation corresponding to one of these intentions is determined. Before performing functional situation analysis, it is necessary to establish a functional situation model and a functional vocabulary model. When performing functional situation analysis, the functional situation model first converts the data in the historical database 134 into feature vectors, and then uses a machine learning algorithm to The data in the historical database 134 is classified according to various situations, and then the strength relationship between the feature vector and each situation is calculated, and then a functional situation model is generated. Machine learning algorithms suitable for establishing the above functional scenarios include: Support Vector Machine (SVM) commonly used in traditional machine learning, and Convolutional Neural Networks (CNN) related to current deep learning (Deep Learning), Algorithms such as Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM).

Following the above, the functional vocabulary model is based on the analysis of a large amount of input training corpus using the Hidden Markov Model algorithm and then performs word segmentation processing. Then the frequency of word segmentation is counted to generate a word segmentation frequency table, and then the functional vocabulary model is established. Please refer to Figure 6 for the detailed process of step S330. Figure 6 is a flow chart of step S330 according to some embodiments of this case. As shown in Figure 6, step S330 includes the following steps:

Step S331: Use corpus data and functional situation information to compare with the functional situation model, and generate functional situation identification results; and

Step S332: Determine that one of these intentions corresponds to one of the query data operation and the instruction execution operation according to the functional context identification result.

In step S331, the corpus data and functional situation information are compared with the functional situation model, and a functional situation identification result is generated. Functional context information includes the user's identity, user's position, user's department, time and location. Part of the functional context information may be sensed by the input/output device 150, for example, the user's current status may be detected (eg, whether the user is back from a business trip). Based on the probabilities and vocabulary corresponding to all intentions obtained after previously identifying the user's corpus data, combined with the functional situation information, the similarity between the user's corpus data and the data in the training data model can be further estimated as the probability of the corresponding intention.

In step S332, it is determined according to the functional context recognition result that one of the intentions corresponds to one of the query data operation and the execution instruction operation. Since there will be multiple query data operation intentions and multiple execution instruction operation intentions in the training data model, and the probability corresponding to each intention will be generated after the calculation of the aforementioned common semantic model, the intention with a lower probability value can Use threshold filtering to get the most likely intent and confirm the corresponding action. As can be seen from the above examples, when words such as "XX company", "last month" and "shipping order" are recognized, these words will be judged and matched with the functional context information to find the most suitable query data operation intention or execute the command operation. Intention, after the above operation, it will be determined that the user said to the virtual assistant: "I want to check the shipment order of XX company last month", and most likely want to check the shipment order of XX company, so it can It corresponds to the query data operation that the user wants to perform. Functional context judgment is needed because users have different needs due to different information such as position, department, operating time, operating location, etc. For example, purchasing personnel and financial personnel will both look at [Manufacturer Monthly Statistical Table] , but the statistical goals of the two [Manufacturer Monthly Statistical Tables] may be different: one is to count the manufacturer's purchase status, and the other is to count the company's payment to the manufacturer. However, when the user talks to the virtual assistant, he may not necessarily specify what he needs [manufacturer's monthly statistics]. He may only say something as simple as: "I need last month's manufacturer's monthly statistics." Therefore, It is even more necessary to match the user's functional situation information to make further accurate judgments.

In step S340, if the functional situation analysis cannot determine the operation corresponding to one of these intentions, word segmentation processing is performed on the corpus data. For the detailed process of step S340, please refer to FIG. 7. FIG. 7 is a flow chart of step S340 according to some embodiments of this case. As shown in Figure 7, step S340 includes the following steps:

Step S341: segment the corpus data according to the functional vocabulary model to generate multiple word segments; and

Step S342: Calculate the frequencies of these word segments.

In steps S341 and S342, the corpus data is segmented according to the functional vocabulary model to generate multiple word segments; and then the frequencies of these segmentations are calculated. If the functional situation analysis cannot determine the operation corresponding to the input corpus data in step S330, it is necessary to perform word segmentation processing on the corpus data. First, the corpus data will be segmented based on the vocabulary stored in the previously pre-established functional vocabulary model, and then the frequency of multiple word segments generated after segmentation will be calculated.

In steps S350 and S360, based on the results of word segmentation processing, it is determined whether there are new words or new corpus data; if there are new words, the natural language processing model is updated according to the meaning of the new words; if there is new corpus data, the natural language processing model is updated according to the new words. Functional context analysis of intent updating of corpus data. Please refer to Figure 8 for the detailed process of step S360. Figure 8 is a flow chart of step S360 according to some embodiments of this case. As shown in Figure 8, step S360 includes the following steps:

Step S361: Determine whether the frequencies of these word segments calculated by the word segmentation processing are lower than the threshold;

Step S362: If one of the word segments is lower than the threshold, one of the word segments is a new vocabulary, and the definition of the new vocabulary is received to update the common vocabulary model and the common semantic model; and

Step S363: If these word segments are higher than the threshold, the corpus data is new corpus data, and the intention of the new corpus data is received to update the functional situation model.

In steps S361 and S362, it is judged whether the frequencies of these word segments calculated by the word segmentation processing are lower than the threshold value. If one of these word segments is lower than the threshold value, one of these word segments is a new vocabulary, and a new word segmentation is received. Definition of vocabulary to update the common vocabulary model and common semantic model. In one embodiment, after calculating the frequency of these word segmentations through word segmentation processing, the word segmentations below the threshold are set as new words. The virtual assistant will ask the user for the definition of the new word and assign the new word and the definition of the new word to the user. Store them together in a common vocabulary model and a common semantic model. For example, the corpus input by the user is "I want to find a contact person from XX company", and if the virtual assistant cannot determine the meaning of "I want to find a contact person from XX company", it will separate "I" after word segmentation processing. , "want to find", "XX company", "of", "contact person" and other words. If "XX company" is lower than the threshold, the virtual assistant will ask the user what "XX company" means, and then add the user The answers and "XX company" are stored in the common vocabulary model and common semantic model; and the new vocabulary also needs to be stored in the system domain vocabulary collection and shared with everyone.

In step S363, if these word segments are higher than the threshold, the corpus data is new corpus data, and the intention of the new corpus data is received to update the functional situation model. Continuing the example of "I want to find a contact person of XX company" above, after word segmentation processing, separate the words "I", "want to find", "XX company", "of", "contact person" and other words. If there are no words in any of them If it is lower than the threshold, it means that the virtual assistant does not understand the intention of the corpus. It is possible that the training corpus used when training the intelligent assistant is all about "Help me check the contact person of XX company", so the virtual assistant will not be able to understand " "I want to find a contact person for XX company", the virtual assistant needs to ask the user what "I want to find a contact person for XX company" means, and then combine the user's answer with "I want to find a contact person for XX company". The new corpus of "people" is stored in the functional situation model together. Before storing the functional model, it is necessary to determine whether the new corpus is common corpus. If so, it means that other people will also use the new corpus when using the virtual assistant. Therefore, the new corpus needs to be stored in the system domain vocabulary collection so that everyone can Shared; but if not, it means that the new corpus is just a different terminology based on the user's own speaking habits, so only the functional situation model needs to be updated, and there is no need to update the system domain vocabulary set.

It can be seen from the above implementation method of this case that the main purpose is to allow the virtual assistant to have an automatic learning function, so that during the process of communicating with the user, the virtual assistant can update the virtual assistant after asking the user if there are words that the intelligent assistant does not understand. The database allows the virtual assistant to automatically learn the user's speaking habits or special terms used in the industry, making it faster and more convenient for users to use the ERP system.

In addition, the above illustrations include sequential exemplary steps, but these steps need not be performed in the order shown. It is contemplated by this disclosure to perform these steps in a different order. These steps may be added, substituted, changed in order and/or omitted as appropriate within the spirit and scope of embodiments of the present disclosure.

Although the implementation of this case has been disclosed as above, it is not used to limit this case. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of this case. Therefore, the scope of protection of this case should be Subject to the scope defined by the appended claims.

Claims (14)

1. An automatic learning method for virtual assistants, characterized by including: receiving an audio input and identifying the audio to form a corpus data; Utilize a natural language processing model to analyze the corpus data to generate a language feature information corresponding to the corpus data, wherein the language feature information includes multiple intentions, probabilities corresponding to the multiple intentions, and multiple words; Perform a functional context analysis on the language feature information based on a functional context information to determine an operation corresponding to one of the plurality of intentions, wherein the functional context information includes an identity, a position, a department, At least one of a time and a place, wherein the functional situation analysis also includes: using the corpus data and the functional situation information to compare with a functional situation model and generating a functional situation identification result; and based on the functional situation The identification result determines that one of the plurality of intentions corresponds to one of a query data operation and an execution instruction operation, wherein the functional situation model is related to a feature vector converted from data in a historical database and a plurality of situations. The strong and weak relationship between them; If the functional situation analysis cannot determine the operation corresponding to one of the multiple intentions, then perform one-word segmentation processing on the corpus data; Based on the result of the word segmentation processing, determine whether there is a new vocabulary or new corpus data; and If the new vocabulary exists, update the natural language processing model according to the meaning of the new vocabulary; if the new corpus data exists, update the functional situation analysis according to the intention of the new corpus data; The operation includes one of a data query operation and an instruction execution operation. 2. The automatic learning method of a virtual assistant according to claim 1, further comprising: Generate a system domain vocabulary set according to an application knowledge database and a domain knowledge database; The system domain vocabulary set and multiple service application parameters form a key entity set, and the key entity set includes multiple system domain vocabulary; Classify multiple training corpora into one of the query data operation and the execution instruction operation; Differentiating the intentions of the plurality of training corpora corresponding to the query data operation according to categories in an enterprise database to form multiple query data operation intentions, and differentiating the plurality of training corpora corresponding to the execution of the instruction operation according to the service behavior provided by an enterprise resource system. The intention of each training corpus forms multiple execution instruction operation intentions; Establishing templates for the plurality of query data operation intentions, and templates for the plurality of execution instruction operation intentions; Establish an overall database based on the key entity set, the plurality of templates of query data operation intentions, and the plurality of templates of execution instruction operation intentions; Identify a plurality of first probabilities that the plurality of system domain words in the key entity set appear in the plurality of training corpora, and analyze the plurality of training corpora through the identified plurality of system domain words. A plurality of sentence structures, and a plurality of correlations between the plurality of system domain words, and establishing a common vocabulary model based on the plurality of first probabilities and the plurality of correlations; and Analyze a plurality of second probabilities of occurrence of the plurality of system domain words in the plurality of query data operation intentions and the plurality of execution instruction operation intentions, and analyze the plurality of second probabilities of occurrence of the plurality of system domain words according to the plurality of sentence structures and the plurality of second Probability establishes a common semantic model. 3. The automatic learning method of a virtual assistant according to claim 2, further comprising: Use a classifier to classify the data in the historical database into strong and weak relationships to generate the functional situation model; and The multiple training corpora are segmented and analyzed, and a functional vocabulary model is generated based on the data in the historical database. 4. The automatic learning method of virtual assistant according to claim 3, characterized in that the word segmentation processing also includes: Segment the corpus data according to the functional vocabulary model to generate multiple word segments; and Calculate the frequency of the plurality of word segments. 5. The automatic learning method of a virtual assistant according to claim 4, further comprising: Determine whether the frequencies of the plurality of word segments calculated by the word segmentation processing are lower than a threshold; If one of the plurality of word segments is lower than the threshold, one of the plurality of word segments is a new vocabulary, and the definition of the new vocabulary is received to update the common vocabulary model and the common semantic model. ;as well as If the plurality of word segments are all higher than the threshold, the corpus data is new corpus data, and the intention of the new corpus data is received to update the functional situation model. 6. The automatic learning method of a virtual assistant according to claim 5, further comprising: Determine whether the new corpus data is common corpus, and if so, update the system domain vocabulary set based on the new corpus data; and The system domain vocabulary set is updated according to the new vocabulary. 7. The automatic learning method of virtual assistant according to claim 2, characterized in that the natural language processing model analyzing the corpus data also includes: Use the common vocabulary model to identify whether the corpus data contains the multiple system domain words that match the key entity set, set the identification results to the multiple words, and analyze the probability of occurrence of the multiple words; Analyze the sentence structure of the corpus data according to the plurality of words; and The common semantic model is used to identify the plurality of intentions of the corpus data and the probabilities corresponding to the plurality of intentions based on the occurrence probabilities of the plurality of words and the sentence structure of the corpus data. 8. An automatic learning system for virtual assistants, which is connected to an enterprise database and an enterprise resource system respectively, and is characterized by including: a processor; A storage device electrically connected to the processor for storing an overall database, an application knowledge database, a domain knowledge database and a historical database; an input/output device electrically connected to the processor for providing an interface for inputting audio; Among them, the processor contains: a speech recognition module for identifying the audio to form a corpus data; A corpus analysis module, electrically connected to the speech recognition module, is used to analyze the corpus data using a natural language processing model to generate a language feature information corresponding to the corpus data, wherein the language feature information includes multiple intentions, The probabilities and multiple words corresponding to the plurality of intentions; A situation identification module, electrically connected to the corpus analysis module, is used to perform a function situation analysis on the language feature information based on a function situation information, and determine an operation corresponding to one of the plurality of intentions, wherein the function The situational information includes at least one of an identity, a position, a department, a time and a place of a user, and a situational analysis module is further configured to use the corpus data, the functional situational information and a functional situational model to perform Compare and generate a functional context identification result; and determine according to the functional context identification result that one of the plurality of intentions corresponds to one of a query data operation and an execution instruction operation, wherein the functional context model is related to The strong and weak relationship between the feature vector converted by the data in a historical database and each person in multiple situations; An unknown corpus judgment module is electrically connected to the situation identification module, and is used to perform word segmentation processing on the corpus data when the situation identification module cannot identify the operation corresponding to one of the plurality of intentions, and perform word segmentation processing on the corpus data according to The result after word segmentation processing is used to determine whether there is a new vocabulary or new corpus data; and An update information module, electrically connected to the unknown corpus judgment module, is used to update the natural language processing model according to the meaning of the new vocabulary when the new vocabulary is generated, and when the new corpus data is generated, update the natural language processing model according to the new vocabulary when the new corpus data is generated. The intention of corpus data is to update the functional situation analysis; The operation includes one of a data query operation and an instruction execution operation. 9. The automatic learning system of virtual assistant according to claim 8, characterized in that the processor further includes: A training module, electrically connected to the corpus analysis module, used to generate a system domain vocabulary set based on the application knowledge database and the domain knowledge database. The system domain vocabulary set and multiple service application parameters form a key entity set, The key entity set includes a plurality of system domain vocabulary, and the plurality of training corpus is classified into one of the query data operation and the execution instruction operation, and the plurality of training corpus corresponding to the query data operation is distinguished according to the categories in the enterprise database. The intention of each training corpus forms multiple query data operation intentions, and the intention of distinguishing the plurality of training corpora corresponding to the instruction operation is formed according to the service behavior provided by the enterprise resource system to form multiple execution instruction operation intentions; A template creation module, electrically connected to the training module, establishes templates of the plurality of query data operation intentions, and templates of the plurality of execution instruction operation intentions, based on the key entity set, the plurality of query data operations The template of the intention and the templates of the plurality of execution instruction operation intentions establish the overall database; A vocabulary model building module, electrically connected to the template building module, identifies a plurality of first probabilities that the plurality of system domain words in the key entity set appear in the plurality of training corpora, and passes the identified The multiple system domain words analyze multiple sentence structures of the multiple training corpora and multiple correlations between the multiple system domain words, and analyze the multiple system domain words according to the multiple first probabilities and the multiple correlations between the multiple system domain words. Multiple associations build a common vocabulary model; and A semantic model building module, electrically connected to the template building module, analyzes the plurality of query data operation intentions and the plurality of execution instruction operation intentions for multiple second probabilities of occurrence of the plurality of system domain words, and A common semantic model is established based on the plurality of sentence structures and the plurality of second probabilities. 10. The automatic learning system of virtual assistant according to claim 9, characterized in that the processor further includes: A situation training module, electrically connected to the situation identification module, is used to use a classifier to classify the data in the historical database into strong and weak relationships to generate the functional situation model; and A vocabulary training module is electrically connected to the unknown corpus judgment module for segmenting and analyzing the plurality of training corpus, and generating a functional vocabulary model based on the data in the historical database. 11. The automatic learning system of the virtual assistant according to claim 10, characterized in that the unknown corpus judgment module is further used to segment the corpus data according to the functional vocabulary model to generate a plurality of word segments to calculate all Describe the frequency of multiple participles. 12. The automatic learning system of the virtual assistant according to claim 11, wherein the update information module is further used to determine whether the frequencies of the plurality of word segments calculated by the word segmentation processing are lower than a threshold; if If one of the plurality of word segments is lower than the threshold, one of the plurality of word segments is the new vocabulary, and the definition of the new vocabulary is received to update the common vocabulary model and the common semantic model; if If the plurality of word segments are all higher than the threshold, the corpus data is new corpus data, and the intention of receiving the new corpus data is to update the functional situation model. 13. The automatic learning system of the virtual assistant according to claim 12, characterized in that the update information module is further used to determine whether the new corpus data is common corpus, and if so, update the system domain vocabulary according to the new corpus data. collection; and updating the system domain vocabulary collection according to the new vocabulary. 14. The automatic learning system of the virtual assistant according to claim 9, wherein the corpus analysis module is further configured to use the common vocabulary model to identify whether the corpus data contains the plurality of key entity sets. System domain vocabulary, set the recognition results to the multiple vocabulary, analyze the probability of occurrence of the multiple vocabulary, analyze the sentence structure of the corpus data based on the multiple vocabulary, and use the common semantic model to analyze the sentence structure of the corpus based on the multiple vocabulary. The occurrence probabilities of the plurality of words and the sentence structure of the corpus data are used to identify the plurality of intentions of the corpus data and the probabilities corresponding to the plurality of intentions.