A method and system for user intention recognition and precise answer generation

Abstract

The present invention discloses a method and system for user intention recognition and precise answer generation, belonging to the technical field of data analysis. The method includes obtaining the historical intention demands of a preset user based on big data, generating a training set from the voice data corresponding to the historical intention demands, extracting each sample text in the training set, generating a recognition model, and inputting the sample text into the recognition model. In the recognition model, the training set is divided into a test set and a validation set. The present invention utilizes deep learning analysis technology to generate data training for voice data by obtaining the historical intention demands of a preset user, and generates a recognition model based on this to train and recognize the voice data of users in real time. At the same time, further dividing the voice text data of users, analyzing and recognizing the user intention according to the divided text data, improving the accuracy of user intention recognition and the analysis and recognition efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and particularly to a method and system for user intention recognition and precise answer generation. Background Art

[0002] In modern society, analyzing users' intention demands according to their intentions and providing intelligent responses have been applied in many industries. Among them, various voice customer service services are widely used in many scenarios. Most of the existing technologies mainly rely on human-machine voice interaction. Currently, intelligent assistants mainly use human-machine voice interaction. Usually, word vectors and context word vectors in sample texts are obtained for training to obtain an intention recognition model. Then, when recognizing a user's intention, the trained intention recognition model can be used to extract the word vector and context word vector matrix of each word in the text to be classified. Then, based on the extracted word vector and context word vector matrix of each word, the corresponding intention of the user input is determined, and then a series of behaviors and strategies are generated and executed to achieve interaction with the user. However, the text to be classified is often obtained by speech recognition of the user's input speech. Due to the limited accuracy of speech recognition, it is sometimes difficult to analyze and determine the user's intention and intention demands based on the recognized speech, which will not only reduce the accuracy of user intention recognition, but sometimes even give the user an analysis reply that is far from their intention, making it difficult to improve the analysis and recognition efficiency. Summary of the Invention

[0003] In view of the above problems existing in the existing technical field of data analysis, the present invention is proposed.

[0004] Therefore, one of the purposes of the present invention is to provide a method and system for user intention recognition and precise answer generation, which uses deep learning analysis technology to generate data training for speech data by obtaining the historical intention demands of preset users, and generates a recognition model based on this to train and recognize the user's speech data in real time. At the same time, the user's speech text data is further divided, and the user's intention is analyzed and recognized based on the divided text data, improving the accuracy of user intention recognition and the analysis and recognition efficiency.

[0005] To solve the above technical problems, the present invention provides the following technical solutions:

[0006] On the one hand, the present invention provides a method for user intention recognition and precise answer generation, including:

[0007] Obtaining the historical intention demands of preset users based on big data, generating a training set from the speech data corresponding to the historical intention demands, extracting each sample text in the training set, generating a recognition model, and inputting the sample text into the recognition model;

[0008] In the recognition model, divide the training set into a test set and a validation set, and mark the acquisition time of each sample text in the training set with the intended demands of a preset user, and analyze the relevance between the sample text acquisition time and the preset user's intended demands, including:

[0009] Obtain the voice data sent by the preset user to the customer service at the beginning of the month, and count the historical intended demands of the preset user during the beginning-of-the-month period according to the obtained voice data. The historical intended demands include the services that the preset user needs to handle and / or consult.

[0010] Collect the word vector features between the different services that the preset user needs to handle and / or consult and the sample text of the voice data, and divide the different services into category A services, category B services, and category C services according to the historical intended demands of the preset user. At the same time, match the collected word vector features according to different service types.

[0011] Generate a response text according to the service type matched by the word vector features, and verify the generated response text with the intended demands of the preset user. If the verification result is correct, store the response text and upload it to the recognition model. Otherwise, do not store it and regenerate the response text until the verification result is correct.

[0012] Analyze the text data features between the response texts with correct verification results and the response texts with incorrect verification results. When replying to the intended demands of the preset user according to the service type required by the preset user in the future period, automatically eliminate the text data features included in the response texts with incorrect verification results from the response text.

[0013] Mark the obtained sample texts separately, and input the differentiated data information into the recognition model, and at the same time update the recognition model.

[0014] Obtain the voice text data of the preset user, divide the voice text data, and analyze each one by one.

[0015] Collect the intended data information of the preset user according to the divided voice text data, and perform judgment processing according to the obtained user intended data information and the corresponding number of voice text data.

[0016] As a preferred solution of the present invention, among them: differentiating the obtained sample texts includes differentiating them into corresponding service handling texts, complaint and suggestion texts, and service consultation texts. Among them, marking the obtained sample texts includes counting the input times of the historical voice texts statistically in the differentiated sample text data of the preset user, and sorting the differentiated sample text data according to the input times.

[0017] As a preferred embodiment of the present invention, the speech text data is divided into θ 1 , θ 2 ,..., θ n , where n represents the nth divided speech text data, the context word vectors between different pieces of speech text data are analyzed, and the currently divided speech text data is recorded and stored.

[0018] As a preferred embodiment of the present invention, the semantic features between each piece of speech text data are analyzed, and the semantic features are divided into first semantic features, second semantic features, and third semantic features, the association rules between each speech feature are obtained, and the association rules are input into the recognition model, where the semantic features include image semantic features.

[0019] As a preferred embodiment of the present invention, according to the obtained user intention data information, a judgment process is performed on the corresponding number of speech text data to verify the intention analysis and recognition result of the corresponding number of speech text data. When the intention analysis and recognition result of the corresponding number of speech text data is consistent with the obtained user intention data information, the data features between the corresponding number of speech text data and the user intention data information are further obtained, and the data features are summarized as correctly recognized data features. On the contrary, when the intention analysis and recognition result of the corresponding number of speech text data is inconsistent with the obtained user intention data information, the data features at the time of inconsistency are also obtained, and the data features are summarized as incorrectly recognized data features.

[0020] As a preferred embodiment of the present invention, among the data features summarized as incorrectly recognized data features, the speech text data corresponding to the data features is further divided into two parts, new semantic analysis is performed according to the secondarily divided speech text data, and at the same time, it is matched with the user intention data information, the intention analysis result of the user is obtained again, and an answer text is generated.

[0021] As a preferred embodiment of the present invention, in the training set, the sample texts therein are trained using the stochastic gradient descent method, and the association features between the preset user intentions and the sample texts are collected to generate a prediction model based on the association features.

[0022] On the other hand, the present invention provides a system for a method of user intention recognition and accurate answer generation, including:

[0023] A big data information acquisition module, configured to acquire the historical intention demands of a preset user, generate a training set from the speech data corresponding to the historical intention demands, and generate a recognition model;

[0024] The comprehensive data processing module is used to distinguish and label the acquired sample texts, input the distinguished data information into the recognition model, update the recognition model at the same time, divide the training set into a test set and a validation set, and mark the acquisition time of each sample text in the training set with the intended demands of the preset user.

[0025] The fusion analysis module is used to acquire the preset user voice text data, divide the voice text data and analyze its semantic features and user intention data information one by one, and the fusion analysis module includes a data statistics module, an analysis module, a verification module and a correction module.

[0026] The data statistics module is used to statistically analyze the historical intention demands of the preset user in the early month period according to the voice data sent by the preset user to the customer service at the beginning of the month, and the historical intention demands include the services that the preset user needs to handle and / or consult.

[0027] The analysis module is used to analyze the word vector features between the different services that the preset user needs to handle and / or consult and the voice data sample texts, divide the different services into type A services, type B services and type C services according to the historical intention demands of the preset user, and match the collected word vector features according to different service types at the same time.

[0028] The verification module is used to generate a reply text according to the service type matched by the word vector features, and verify the generated reply text with the intended demands of the preset user at the same time. If the verification result is correct, store the reply text and upload it to the recognition model. Otherwise, do not store it and regenerate the reply text until the verification result is correct.

[0029] The correction module is used to analyze the text data features between the reply text with a correct verification result and the reply text with an incorrect verification result. When replying to the intended demands of the preset user according to the service type required by the preset user in the future period, automatically remove the text data features included in the reply text with an incorrect verification result from the reply text.

[0030] The data discrimination module, the data discrimination module responds to the voice text data divided item by item by the fusion analysis module, and performs judgment processing according to the acquired user intention data information and the corresponding number of voice text data to verify the intention analysis and recognition result of the voice text data of this number.

[0031] A terminal includes a processor, an input interface, an output interface and a memory. The processor, the input interface, the output interface and the memory are connected to each other. Among them, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 7.

[0032] A computer-readable storage medium stores a computer program, and the computer program includes program instructions which, when executed by a processor, cause the processor to execute the method according to any one of claims 1 to 7.

[0033] Beneficial effects:

[0034] The present invention utilizes deep learning analysis technology. By obtaining the historical intention demands of a preset user, data training is generated for voice data, and an identification model is generated based on this to train and identify the user's voice data in real time. At the same time, the user's voice text data is further divided, and the user's intention is analyzed and judged according to the divided text data, and an accurate response is made to the user's intention demand. When a certain piece of divided voice text data fails to correctly identify and judge the user's intention, the piece of voice text data can be further divided to re-identify and judge the user's intention, thereby continuously improving and optimizing the intention recognition and answer quality, and improving the accuracy of user intention recognition and the analysis and recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts. Among them:

[0036] Figure 1 is a modular structure diagram of the user intention recognition and accurate answer generation system according to an embodiment of the present invention;

[0037] Figure 2 is a schematic flowchart of the method according to an embodiment of the present invention;

[0038] Figure 3 is a schematic flowchart of the method according to an embodiment of the present invention;

[0039] Figure 4 is a schematic flowchart tree structure diagram according to an embodiment of the present invention;

[0040] Reference numerals in the drawings: 110 - big data information acquisition module; 120 - comprehensive data processing module; 130 - fusion analysis module; 1301 - data statistics module; 1302 - analysis module; 1303 - verification module; 1304 - correction module; 140 - data discrimination module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention fall within the scope of protection of the present invention.

[0042] Since the prior art uses the method of classifying the to-be-classified speech text to perform speech recognition on the user input speech, it is sometimes difficult to analyze and determine the user's intention and intention requirements based on the recognized speech, which will not only lead to a decrease in the accuracy of user intention recognition, but sometimes even cause the user to obtain an analysis reply that is quite different from their intention, making it difficult to improve the analysis and recognition efficiency.

[0043] Based on this, the present invention proposes a method and system for generating user intention recognition and accurate answers. It uses deep learning analysis technology to generate data training for speech data by obtaining the historical intention requirements of preset users, and generates a recognition model based on this to perform training recognition on the user's speech data in real time. At the same time, further divide the user's speech text data, analyze and identify the user's intention according to the divided text data, and give a precise reply to the user's intention requirements, improving the accuracy of user intention recognition and the analysis and recognition efficiency.

[0044] The following further specifically describes this solution through embodiments and with reference to the accompanying drawings.

[0045] Refer to Figures 1 to 4 , which is an embodiment of the present invention. This embodiment provides a method for generating user intention recognition and accurate answers, including:

[0046] S10: Obtain the historical intention requirements of preset users based on big data, generate a training set from the speech data corresponding to the historical intention requirements, extract each sample text in the training set, generate a recognition model, and input the sample text into the recognition model;

[0047] It should be noted in this embodiment that the obtained sample texts are distinguished, including being distinguished into corresponding business handling texts, complaint and suggestion texts, and business consultation texts. Among them, the obtained sample texts are marked, including the input times of the historical speech texts statistically in each of the distinguished sample text data by the preset user, and the distinguished sample text data are sorted according to the input times;

[0048] On this basis, in this embodiment, further in the training set, the sample texts therein are trained using the stochastic gradient descent method, and the correlation features between the preset user intention and the sample texts are collected to generate a prediction model based on the correlation features;

[0049] S20: In the recognition model, divide the training set into a test set and a validation set, mark the acquisition time of each sample text in the training set with the intended demands of a preset user, and analyze the relevance between the acquisition time of the sample text and the intended demands of the preset user, including:

[0050] S201: Obtain the voice data sent by the preset user to the customer service at the beginning of the month, and count the historical intended demands of the preset user at the beginning of the month according to the obtained voice data. The historical intended demands include the services that the preset user needs to handle and / or consult;

[0051] S202: Collect the word vector features between the different services that the preset user needs to handle and / or consult and the sample text of the voice data, divide the different services into type A services, type B services, and type C services according to the historical intended demands of the preset user, and match the collected word vector features according to different service types;

[0052] S203: Generate a response text according to the service type matched by the word vector features, and verify the generated response text with the intended demands of the preset user. If the verification result is correct, store the response text and upload it to the recognition model. Otherwise, do not store it and regenerate the response text until the verification result is correct;

[0053] S204: Analyze the text data features between the response text with a correct verification result and the response text with an incorrect verification result. When replying to the intended demands of the preset user according to the service type required by the preset user in the future period, automatically eliminate the text data features included in the response text with an incorrect verification result from the response text;

[0054] S30: Mark the obtained sample text for differentiation, input the differentiated data information into the recognition model, and update the recognition model at the same time;

[0055] Specifically in this embodiment, divide the voice text data into θ 1 , θ 2 ,..., θ n , where n represents the nth divided voice text data, analyze the context word vectors between different voice text data, and record and store the currently divided voice text data;

[0056] S40: Obtain the voice text data of the preset user, divide the voice text data and analyze it one by one;

[0057] S50: Collect the intended data information of the preset user according to the divided voice text data, and make a judgment and processing according to the obtained user intended data information and the corresponding number of voice text data;

[0058] In this embodiment, further, the obtained user intention data information is judged and processed with the corresponding number of speech text data to verify the intention analysis and recognition result of the speech text data of this number. When the intention analysis and recognition result of the speech text data of the corresponding number is consistent with the obtained user intention data information, the data features between the speech text data of the corresponding number and the user intention data information are further obtained, and the data features are summarized as the correctly recognized data features. On the contrary, when the intention analysis and recognition result of the speech text data of the corresponding number is inconsistent with the obtained user intention data information, the data features at the time of inconsistency are also obtained, and the data features are summarized as the incorrectly recognized data features;

[0059] On the above basis, among the data features summarized as incorrectly recognized data features, the speech text data corresponding to the number of data features is further divided into two parts, and new semantic analysis is performed according to the speech text data after the two-part division. At the same time, it is matched with the user intention data information, and the intention analysis result of the user is obtained again, and a response text is generated;

[0060] Further, the semantic features between each piece of speech text data are analyzed, and the semantic features are divided into first semantic features, second semantic features and third semantic features. The association rules between the speech features are obtained, and the association rules are input into the recognition model, where the semantic features include image semantic features.

[0061] Based on the above, this application uses deep learning analysis technology. By obtaining the historical intention demands of preset users, data training is generated for speech data, and a recognition model is generated based on this to train and recognize the speech data of users in real time. At the same time, the speech text data of users is further divided, and the user intention is analyzed and judged according to the divided text data, and the intention demands of users are accurately answered, improving the accuracy and analysis and recognition efficiency of user intention recognition.

[0062] Combined with the above method for user intention recognition and accurate answer generation, this embodiment also proposes a system applied to this method, including:

[0063] The big data information acquisition module 110 is used to obtain the historical intention demands of preset users, generate a training set from the speech data corresponding to the historical intention demands, and generate a recognition model;

[0064] The comprehensive data processing module 120 is used to distinguish and mark the obtained sample texts, input the distinguished data information into the recognition model, update the recognition model at the same time, divide the training set into a test set and a validation set, and mark the acquisition time of each sample text in the training set with the intention demands of the preset user;

[0065] The fusion analysis module 130 is used to obtain preset user voice text data, divide the voice text data, analyze its semantic features and user intention data information one by one, and the fusion analysis module includes a data statistics module 1301, an analysis module 1302, a verification module 1303 and a correction module 1304;

[0066] The data statistics module 1301 is used to statistically analyze the historical intention demands of the preset user in the early month period according to the voice data sent by the preset user to the customer service at the beginning of the month. The historical intention demands include the services that the preset user needs to handle and / or consult;

[0067] The analysis module 1302 is used to analyze the word vector features between the different services that the preset user needs to handle and / or consult and the voice data sample text, divide the different services into category A services, category B services and category C services according to the historical intention demands of the preset user, and match the collected word vector features according to different service types;

[0068] The verification module 1303 is used to generate a reply text according to the service type matched by the word vector features, and at the same time verify the generated reply text with the intention demands of the preset user. If the verification result is correct, the reply text is stored and uploaded to the recognition model. Otherwise, it is not stored, and a new reply text is generated until the verification result is correct;

[0069] The correction module 1304 is used to analyze the text data features between the reply text with a correct verification result and the reply text with an incorrect verification result. When replying to the intention demands of the preset user according to the service type required by the preset user in the future period, the text data features included in the reply text with an incorrect verification result are automatically excluded from the reply text;

[0070] The data discrimination module 140. The data discrimination module 140 responds to the voice text data divided item by item by the fusion analysis module 130, and performs judgment processing according to the obtained user intention data information and the corresponding number of voice text data to verify the intention analysis and recognition result of the voice text data of this number.

[0071] A terminal includes a processor, an input interface, an output interface and a memory. The processor, the input interface, the output interface and the memory are connected to each other. Among them, the memory is used to store a computer program. The computer program includes program instructions. The processor is configured to call the program instructions to execute the method according to any one of claims 1 to 7.

[0072] A computer-readable storage medium stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, the processor is caused to execute the method according to any one of claims 1 to 7.

[0073] In summary, the present invention utilizes deep learning analysis technology. By obtaining the historical intent demands of a preset user, data training is generated for voice data, and an identification model is generated based on this to train and identify the user's voice data in real time. At the same time, the user's voice text data is further divided, and the user's intent is analyzed and judged according to the divided text data, and a precise response is given to the user's intent demand. When a certain piece of divided voice text data fails to correctly identify and judge the user's intent, the piece of voice text data can be further divided to re-identify and judge the user's intent, thereby continuously improving and optimizing the intent recognition and answer quality, and enhancing the accuracy of user intent recognition and the analysis and recognition efficiency.

[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying user intention and generating accurate answers, characterized in that: include: Based on big data, the historical intention demands of preset users are obtained, and a training set is generated from the voice data corresponding to the historical intention demands, each sample text in the training set is extracted, a recognition model is generated, and the sample text is input into the recognition model; In the recognition model, the training set is divided into a test set and a validation set, and the acquisition time of each sample text in the training set is marked with the preset user's intention demand, and the correlation between the sample text acquisition time and the preset user's intention demand is analyzed, including: Acquire voice data sent by a preset user to a customer service representative at the beginning of the month, and collect historical intentions and demands of the preset user at the beginning of the month based on the acquired voice data, wherein the historical intentions and demands include the services that the preset user needs to handle and / or consult; Collect word vector features between different services that the preset user needs to handle and / or consult and the voice data sample text, and classify the different services into Class A services, Class B services, and Class C services according to the historical intention demands of the preset user, and match the collected word vector features according to different service types; Generate a reply text according to the business type matched by the word vector feature, and verify the generated reply text with the preset user's intention. If the verification result is correct, store the reply text and upload it to the recognition model. Otherwise, do not store it and regenerate the reply text until the verification result is correct. Analyze the text data features between the reply text with a correct verification result and the reply text with a verification error. When replying to the intended request of the user according to the type of business that the user needs to handle in the future period, automatically remove the text data features contained in the reply text with a verification error. Marking the acquired sample texts, inputting the distinguished data information into the recognition model, and updating the recognition model; The obtained sample texts are distinguished, including distinguishing into corresponding business handling texts, complaint and suggestion texts, and business consultation texts, wherein the obtained sample texts are marked, including counting the number of inputs of historical voice texts by a preset user in each distinguished sample text data, and sorting each distinguished sample text data according to the number of inputs; Acquire preset user voice and text data, divide the voice and text data and analyze them one by one; At the same time, the speech text data is divided into θ1, θ2, ..., θ n , where n represents the nth divided voice text data, analyzes the context word vectors between different voice text data, and records and stores the currently divided voice text data; The intention data information of the preset user is collected according to the divided voice and text data, and judgment and processing are performed based on the acquired user intention data information and the corresponding number of voice and text data.

2. A method for identifying user intention and generating accurate answers as claimed in claim 1, characterized in that: Analyze the semantic features between each voice text data, and divide the semantic features into a first semantic feature, a second semantic feature and a third semantic feature, obtain the association rules between each voice feature, and input the association rules into the recognition model, wherein the semantic features include image semantic features.

3. A method for identifying user intention and generating accurate answers as claimed in claim 1, characterized in that: A judgment process is performed based on the acquired user intention data information and the corresponding number of voice and text data to verify the intention analysis and recognition results of the corresponding number of voice and text data. When the intention analysis and recognition results of the corresponding number of voice and text data are consistent with the acquired user intention data information, data features between the corresponding number of voice and text data and the user intention data information are further acquired, and the data features are summarized as correctly recognized data features. Conversely, when the intention analysis and recognition results of the corresponding number of voice and text data are inconsistent with the acquired user intention data information, the data features of the inconsistency are also acquired, and the data features are summarized as incorrectly recognized data features.

4. A method for identifying user intention and generating accurate answers as claimed in claim 3, characterized in that: Among the data features summarized as incorrectly identified, the voice text data corresponding to the data features are further divided for a second time, and a new semantic analysis is performed based on the secondarily divided voice text data. At the same time, the data is matched with the user intention data information to re-obtain the analysis result of the user's intention and generate a reply text.

5. A method for identifying user intention and generating accurate answers as claimed in claim 1, characterized in that: In the training set, the sample texts are trained using the stochastic gradient descent method, and correlation features between the preset user intentions and the sample texts are collected to generate a prediction model based on the correlation features.

6. A system for the method of identifying user intention and generating accurate answers as claimed in claim 1, characterized in that: include: A big data information acquisition module is used to obtain the historical intention demands of a preset user, and generate a training set with the voice data corresponding to the historical intention demands to generate a recognition model; A comprehensive data processing module is used to distinguish and mark the acquired sample texts, input the distinguished data information into the recognition model, update the recognition model, divide the training set into a test set and a validation set, and mark the acquisition time of each sample text in the training set with the preset user's intention appeal; A fusion analysis module is used to obtain preset user voice and text data, divide the voice and text data and analyze its semantic features and user intention data information one by one, and the fusion analysis module includes a data statistics module, an analysis module, a verification module and a correction module; The data statistics module is used to collect historical intention demands of the preset user in the period at the beginning of the month based on the voice data sent by the preset user to the customer service in the period at the beginning of the month, and the historical intention demands include the services that the preset user needs to handle and / or consult; The analysis module is used to analyze the word vector features between the different services that the preset user needs to handle and / or consult and the voice data sample text, and classify the different services into Class A services, Class B services and Class C services according to the historical intention demands of the preset user, and match the collected word vector features according to different business types; The verification module is used to generate a reply text according to the business type matched by the word vector feature, and verify the generated reply text with the preset user's intention and demand. If the verification result is correct, the reply text is stored and uploaded to the recognition model. Otherwise, it is not stored and the reply text is regenerated until the verification result is correct; The correction module is used to analyze the text data features between the reply text with a correct verification result and the reply text with a verification error, and when replying to the intended request of the user according to the type of business that the user needs to handle in the future period, the text data features contained in the reply text with a verification error are automatically removed from the reply text; The data identification module responds to the voice and text data divided piece by piece by the fusion analysis module, and performs judgment and processing based on the acquired user intention data information and the corresponding number of voice and text data to verify the intention analysis and recognition results of the number of voice and text data.

Images