Intelligent message reply processing method and device based on semantic analysis large model

By adopting a message intelligent reply processing method based on a large semantic analysis model, the problem of low message reply intelligence is solved, achieving efficient and accurate replies and improving user experience.

CN118939778APending Publication Date: 2024-11-12SHENZHEN BORUI WORLD TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411117857.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Current technologies have low levels of intelligence in message replies, resulting in inefficient reply methods and poor consistency and accuracy.

Method used

A message intelligent reply processing method based on a semantic analysis big data model is adopted. By reading a pre-defined reading and extraction plan, user messages are read and extracted to generate user action strings. The reply message data set is then filtered, deduplicated, and matched with reply templates. The reply priority is sorted by combining a user satisfaction prediction function, and finally an intelligent reply is generated.

Benefits of technology

It enables efficient and accurate responses to user messages, improves the efficiency and quality of intelligent message reply processing, and enhances the user's message reply experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939778A_ABST
    Figure CN118939778A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent message reply processing method and device based on a semantic analysis large model, and relates to the technical field related to data processing.The method comprises the steps that a first reply message set is obtained by obtaining a first user action string and traversing and screening in a reply message data set, a first reply template set is obtained for duplicate removal, and a second reply message set is obtained; and obtaining a first target reply template set. And generating a first candidate reply, and introducing a user satisfaction prediction function to perform user satisfaction prediction on the first candidate reply to obtain a first predicted satisfaction. And performing reply priority ranking on the first candidate reply set according to the first predicted satisfaction to obtain a first candidate reply list. And extracting a target candidate reply in the first candidate reply list to intelligently reply the first user message. The technical problems of low intelligence of message reply, low reply mode efficiency and poor reply consistency and accuracy in the prior art are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to fields related to data processing technology, and specifically to a message intelligent reply processing method and device based on a semantic analysis large model. Background Art

[0002] With the rapid development of communication technology, the speed of information transmission and the amount of data exchange have been greatly improved, bringing users an unprecedented efficient communication experience. However, faced with a large number of user messages, how to quickly and accurately process and reply to messages has become an urgent problem to be solved in the field of communication services. The traditional manual reply method is not only inefficient, but also difficult to ensure the consistency and accuracy of the reply.

[0003] Therefore, in the prior art, message replies are less intelligent, the reply method is inefficient, and there are technical problems such as poor reply consistency and accuracy. Summary of the invention

[0004] This application provides a method and device for intelligent message reply processing based on a semantic analysis large model, which solves the technical problems of low intelligence of message reply, low efficiency of reply mode, and poor reply consistency and accuracy in the prior art. It realizes efficient and accurate reply to user messages, improves the efficiency and quality of intelligent message reply processing, and improves the user's message reply experience.

[0005] The present application provides a method for intelligent reply processing of messages based on a semantic analysis large model, the method comprising: reading a predetermined reading and extraction plan, and performing reading and extraction on a first user message based on the predetermined reading and extraction plan to obtain a first user action string. Traverse and filter the first user action in the first user action string in the reply message data set to obtain a first reply message set. De-duplicate the first reply template set obtained by analyzing the first reply message set to obtain a first target reply template set. Generate a first candidate reply based on the first target reply template in the first target reply template set and in combination with the first user action. Introduce a user satisfaction prediction function to perform user satisfaction prediction on the first candidate reply to obtain a first predicted satisfaction. According to the first predicted satisfaction, the first candidate reply set formed based on the first candidate reply is sorted by reply priority to obtain a first candidate reply list. Extract the target candidate reply in the first candidate reply list to intelligently reply to the first user message.

[0006] In an implementation, the reply message data set includes multiple reply message data sets of predetermined message types, including: analyzing the first user action string through a user message type discrimination model trained based on big data to obtain a first message type. A first reply message data set matching the first message type in the multiple reply message data sets, wherein the first reply message data set is used to perform traversal screening of the first user action. The predetermined message types include at least consultation, inquiry, and complaint.

[0007] In an implementation, a second user message is extracted from a user message data set, and the second user message has an identifier of a second message type. The second user message is read and extracted based on the predetermined reading and extraction plan to obtain a second user action string. A training data set composed of the second user action string and the second message type is supervised and tested based on the support vector machine principle to obtain the user message type discrimination model.

[0008] In an implementation, the first reply message set includes a plurality of reply message data with a reply template identifier. The plurality of reply message data with a reply template identifier are matched and screened based on the first target reply template to obtain first target message data, the first target message data including multiple reply messages with a user satisfaction identifier. The user satisfaction prediction function combines the multiple reply messages with the user satisfaction identifier to obtain the first predicted satisfaction.

[0009] In an implementation, the expression of the user satisfaction prediction function is as follows: ; in, It refers to the reply message among the multiple reply messages with user satisfaction indicators. Reply to the first candidate The first satisfaction support of , k refers to the magnitude of the user satisfaction prediction function, and , is the attenuation factor, and .

[0010] In the implementation method, it is determined whether the average of the satisfaction support obtained by calculating the first satisfaction support meets the support limit. If it meets the support limit, the reply message matched in the multiple reply messages with user satisfaction identifiers is The corresponding user satisfaction is used as the first predicted satisfaction.

[0011] In an implementation, if there is non-compliance, a manual reply instruction is issued, and a manual reply is activated based on the manual reply instruction.

[0012] In an implementation, templates in the first target reply template set are newly added or modified based on the first user action string.

[0013] The present application also provides a message intelligent reply processing device based on a semantic analysis large model, including: The action string acquisition module is used to read the predetermined reading and extraction plan, and perform reading and extraction on the first user message based on the predetermined reading and extraction plan to obtain the first user action string.

[0014] The message set acquisition module is used to traverse and filter the first user action in the first user action string in the reply message data set to obtain a first reply message set.

[0015] The template set acquisition module is used to deduplicate the first reply template set obtained by analyzing the first reply message set to obtain a first target reply template set.

[0016] A candidate reply acquisition module is used to generate a first candidate reply according to a first target reply template in the first target reply template set and in combination with the first user action.

[0017] The satisfaction prediction module is used to introduce a user satisfaction prediction function to predict the user satisfaction of the first candidate reply to obtain a first predicted satisfaction.

[0018] A reply list acquisition module is used to sort the reply priorities of a first candidate reply set formed based on the first candidate reply according to the first predicted satisfaction level to obtain a first candidate reply list.

[0019] The intelligent reply module is used to extract the target candidate reply from the first candidate reply list and intelligently reply to the first user message.

[0020] The method and device for intelligent reply processing of messages based on a semantic analysis large model proposed in this application are proposed to read a predetermined reading and extraction plan, and read and extract the first user message based on the predetermined reading and extraction plan to obtain a first user action string. The first user action in the first user action string is traversed and screened in the reply message data set to obtain a first reply message set. The first reply template set obtained by analyzing the first reply message set is deduplicated to obtain a first target reply template set. According to the first target reply template in the first target reply template set, a first candidate reply is generated in combination with the first user action. A user satisfaction prediction function is introduced to predict user satisfaction for the first candidate reply to obtain a first predicted satisfaction. According to the first predicted satisfaction, the first candidate reply set formed based on the first candidate reply is sorted by reply priority to obtain a first candidate reply list. The target candidate reply in the first candidate reply list is extracted to intelligently reply to the first user message. The technical problems of low intelligence of message reply, low efficiency of reply mode, and poor reply consistency and accuracy in the prior art are solved. Efficient and accurate reply to user messages is achieved, the efficiency and quality of message intelligent reply processing are improved, and the user's message reply experience is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solution of the embodiment of the present disclosure, the accompanying drawings of the embodiment of the present disclosure will be briefly introduced below. A flow chart is used in the present application to illustrate the operations performed by the device according to the embodiment of the present application. It should be understood that the previous or following operations are not necessarily performed precisely in order. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or more operations can be removed from these processes.

[0022] Figure 1 A schematic diagram of a flow chart of a message intelligent reply processing method based on a semantic analysis large model provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a message intelligent reply processing device based on a semantic analysis large model provided in an embodiment of the present application.

[0023] Explanation of the accompanying drawings: action string acquisition module 11, message set acquisition module 12, template set acquisition module 13, candidate reply acquisition module 14, satisfaction prediction module 15, reply list acquisition module 16, intelligent reply module 17. DETAILED DESCRIPTION

[0024] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0025] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0026] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, device, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0027] The present application embodiment provides a message intelligent reply processing method and device based on a semantic analysis large model, such as Figure 1 As shown, the method includes: The predetermined reading and extraction plan is read, and based on the predetermined reading and extraction plan, the first user message is read and extracted to obtain a first user action string.

[0028] The first user action in the first user action string is traversed and screened in the reply message data set to obtain a first reply message set.

[0029] The first reply template set obtained by analyzing the first reply message set is deduplicated to obtain a first target reply template set.

[0030] With the rapid development of communication technology, the speed of information transmission and the amount of data exchange have been greatly improved, bringing users an unprecedented efficient communication experience. However, in the face of a large number of user messages, how to quickly and accurately process and reply messages has become an urgent problem to be solved in the field of communication services. The traditional manual reply method is not only inefficient, but also difficult to ensure the consistency and accuracy of the reply. By reading the predetermined reading and extraction plan, the predetermined reading and extraction plan is a pre-set message extraction plan, such as extracting all verbs such as received, unpacked, and found in the message, and all nouns such as cosmetics, food, etc. Based on the predetermined reading and extraction plan, the first user message is read and extracted, and the first user message is any message obtained in real time, and the first user action string is obtained by reading and extracting it. Further, the first user action in the first user action string is traversed and screened in the first reply message data set obtained based on the reply message data set, and the reply message data corresponding to the first user action in the first reply message data set is obtained, and the reply messages corresponding to all actions in the first user action string are traversed to obtain the first reply message set, and the first reply message set includes a plurality of reply message data with a reply template identifier in the first reply message set. Furthermore, the first reply template set obtained by analyzing the first reply message set is deduplicated. When analyzing the first reply message set, the reply messages corresponding to each action in the first reply message set are extracted, and the reply messages that appear in each action reply message are obtained as common reply messages. The template set corresponding to the common reply messages is extracted to obtain the first reply template set, and the same templates are deduplicated for the first reply template set to obtain the first target reply template set.

[0031] The first user action string is analyzed by a user message type discrimination model trained based on big data to obtain a first message type.

[0032] A first reply message data set matching the first message type in the multiple reply message data sets, wherein the first reply message data set is used for traversal and screening of the first user action.

[0033] The predetermined message types include at least consultation, inquiry and complaint.

[0034] The reply message data set includes multiple reply message data sets of predetermined message types, including: the predetermined message types include at least consultation type, query type and complaint type. The first user action string is analyzed by a user message type discrimination model trained based on big data, and the first message type output by the user message type discrimination model is obtained. Further, the first reply message data set corresponding to the first message type is matched in the multiple reply message data sets, and the multiple reply message data sets are constructed using historical reply data, and the reply data of the consultation type, query type and complaint type in the historical reply data are obtained, and the reply message data sets are constructed respectively, and multiple reply message data sets are merged to obtain multiple reply message data sets. Among them, the first reply message data set is used to traverse and filter the first user action.

[0035] The method provided in the embodiment of the present application also includes: A second user message is extracted from the user message data set, where the second user message has an identifier of a second message type.

[0036] The second user message is read and extracted based on the predetermined reading and extraction plan to obtain a second user action string.

[0037] Based on the support vector machine principle, supervised learning and verification are performed on the training data set composed of the second user action string and the second message type to obtain the user message type discrimination model.

[0038] Based on big data, a user message data set is obtained, in which historical user messages and corresponding message type identification data are recorded, and the message type identification data categories include consultation messages, query messages and complaint messages. A second user message is extracted from the user message data set, and the second user message is a random message in the user message data set, and the second user message has an identification of the second message type. The second user message is read and extracted through a predetermined reading and extraction plan to obtain a second user action string, and the second user message in the user message data set is continuously obtained, and the corresponding second user action string is obtained. The second user data obtained each time is stored in the training data set, and the data acquisition is repeated until the data volume meets the training requirements, and the training data set construction is completed. Finally, the support vector machine is supervised and tested based on the training data set, wherein the identification data of the message type is the supervision data, and the training is completed until the accuracy of the model output after the test reaches the preset accuracy, and the user message type discrimination model is obtained.

[0039] A first candidate reply is generated according to a first target reply template in the first target reply template set and in combination with the first user action.

[0040] A user satisfaction prediction function is introduced to predict the user satisfaction of the first candidate reply to obtain a first predicted satisfaction.

[0041] The first candidate reply set formed based on the first candidate reply is sorted by reply priority according to the first predicted satisfaction level to obtain a first candidate reply list.

[0042] Extract the target candidate reply from the first candidate reply list and perform an intelligent reply to the first user message.

[0043] According to the first target reply template in the first target reply template set, the first candidate reply is generated in combination with the first user action. That is, according to the first target reply template, the non-template part is filled with content in combination with the first user action, so as to generate the first candidate reply. When filling the template, the corresponding words and sentences are obtained and filled in the first user action based on the part of speech corresponding to the standard non-template part in the template and the proportion of common words, so as to complete the filling of the non-template part in the template and generate the candidate reply. Further, the user satisfaction prediction function is introduced to predict the user satisfaction of the first candidate reply, and the satisfaction prediction result of the prediction function is obtained to obtain the first predicted satisfaction. By obtaining the first satisfaction support, the satisfaction of the reply message is predicted based on the reply messages with user satisfaction identifiers for multiple times, and then the first predicted satisfaction is generated to complete the sorting and screening of the reply messages in the optimal candidate reply set, so as to facilitate the subsequent acquisition of the best reply data. Further, according to the first predicted satisfaction, the first candidate reply set formed based on the first candidate reply is sorted in reply priority. When sorting the reply priority, the first candidate reply set formed by the first candidate reply is sorted from large to small according to the predicted satisfaction, so as to obtain the first candidate reply list. Finally, the target candidate reply in the first candidate reply list is extracted to intelligently reply to the first user message, and the target candidate reply is the candidate reply with the highest satisfaction in the first candidate reply list. The technical problems of low intelligence of message reply, low efficiency of reply method, and poor reply consistency and accuracy in the prior art are solved. Efficient and accurate reply to user messages is achieved, the efficiency and quality of message intelligent reply processing are improved, and the user's message reply experience is improved.

[0044] The method provided in the embodiment of the present application also includes: The first reply message set includes a plurality of reply message data having a reply template identifier.

[0045] The plurality of reply message data having the reply template identifier are matched and screened based on the first target reply template to obtain first target message data, wherein the first target message data includes multiple reply messages having the user satisfaction identifier.

[0046] The user satisfaction prediction function combines the multiple reply messages with the user satisfaction identifier to obtain the first predicted satisfaction.

[0047] The first reply message set includes multiple reply message data with reply template identifiers, that is, the reply message data in the first reply message set all have corresponding reply template identifiers, and the template identifiers may be the same or different. The multiple reply message data with reply template identifiers are matched and screened based on the first target reply template, and the reply message data corresponding to the first target reply template in the historical reply message data are obtained to obtain the first target message data. The first target message data contains multiple reply messages with user satisfaction identifiers, that is, each reply message in the first target message data corresponds to reply data containing multiple user satisfaction identifiers. Finally, a satisfaction prediction calculation is performed based on the user satisfaction prediction function combined with the multiple reply messages with user satisfaction identifiers as the data basis for predicting satisfaction, to obtain the first predicted satisfaction.

[0048] The expression of the user satisfaction prediction function is as follows: ; in, It refers to the reply message among the multiple reply messages with user satisfaction indicators. Reply to the first candidate The first satisfactory support degree of , Refers to a specific reply message. Reply message to the first candidate, is one of the multiple reply messages with user satisfaction indicators, where: for Reply message data and Reply message difference, actually corresponds to 1 minus Reply message data and The similarity parameter of the reply message is obtained by calculating the cosine similarity. Reply to messages, The reply message is processed by word segmentation and stop word removal, and then the processed text is subjected to text feature extraction through TF-IDF to obtain vectorized messages. Finally, the cosine similarity between the two message vectors is calculated by the cosine similarity calculation method to obtain the similarity calculation result. Reply message data and The higher the similarity parameter of the reply message, the higher the corresponding first satisfaction support parameter. k refers to the amplitude of the user satisfaction prediction function, and , Refers to the attenuation factor. The attenuation factor is set based on actual conditions, and .

[0049] The method provided in the embodiment of the present application also includes: Determine whether the average of the satisfactory support obtained by calculating the first satisfactory support meets the support limit.

[0050] If it is matched, the reply message matched in the multiple reply messages with user satisfaction identifiers is The corresponding user satisfaction is used as the first predicted satisfaction.

[0051] Determine whether the average of the satisfied support obtained by calculating the first satisfied support meets the support limit. The support limit is the preset minimum satisfied support. When the first satisfied support is less than the limit, the corresponding first satisfied support does not meet the support limit. The evaluation parameters of the reply message data in the first target message data have a low evaluation support for the current candidate reply, and the output result of the candidate reply is poor. Conversely, when the first satisfied support is greater than or equal to the limit, the corresponding first satisfied support meets the support limit. The evaluation parameters of the reply message data in the first target message data have a high evaluation support for the current candidate reply, and the output result of the candidate reply is better. If the first satisfied support meets the support limit, the reply message matched in the multiple reply messages with user satisfaction identifiers will be The corresponding user satisfaction is used as the first predicted satisfaction. If it does not meet the requirements, a manual reply instruction is issued, and a manual reply is activated based on the manual reply instruction. By obtaining the first satisfaction support, the reply message satisfaction is predicted based on multiple reply messages with user satisfaction identifiers, and then the first predicted satisfaction is generated to complete the sorting and screening of the reply messages in the optimal candidate reply set, thereby obtaining the best reply data.

[0052] The method provided in the embodiment of the present application also includes: Based on the first user action string, templates in the first target reply template set are added or modified.

[0053] The first target reply template set is added or modified according to the predetermined update frequency. When the predetermined update frequency is reached, the corresponding new template or updated template is retrieved through big data based on the acquired first user action string, and the template of the first target reply template set is added and the template of the original template is adjusted based on the new template or updated template. The new template is the reply template corresponding to the newly added first user action string, and the updated template is the template in which the reply template corresponding to the first user action string of the first target reply template set has been modified. The first target reply template set is updated according to the predetermined update frequency to ensure that the information stored in the first target reply template set is the latest, thereby improving the accuracy of the reply.

[0054] In the above, refer to Figure 1 The message intelligent reply processing method based on the semantic analysis large model according to the embodiment of the present invention is described in detail. Figure 2 A message intelligent reply processing device based on a semantic analysis large model according to an embodiment of the present invention is described.

[0055] The message intelligent reply processing device based on the semantic analysis big model according to the embodiment of the present invention solves the technical problems in the prior art that the message reply is low in intelligence, the reply method is inefficient, and the reply consistency and accuracy are poor. It achieves efficient and accurate replies to user messages, improves the efficiency and quality of message intelligent reply processing, and improves the user's message reply experience. The message intelligent reply processing device based on the semantic analysis big model includes: an action string acquisition module 11, a message set acquisition module 12, a template set acquisition module 13, a candidate reply acquisition module 14, a satisfaction prediction module 15, a reply list acquisition module 16, and an intelligent reply module 17.

[0056] The action string acquisition module 11 is used to read the predetermined reading and extraction plan, and perform reading and extraction on the first user message based on the predetermined reading and extraction plan to obtain the first user action string.

[0057] The message set acquisition module 12 is used to traverse and filter the first user action in the first user action string in the reply message data set to obtain a first reply message set.

[0058] The template set acquisition module 13 is used to remove duplicates from the first reply template set obtained by analyzing the first reply message set to obtain a first target reply template set.

[0059] The candidate reply acquisition module 14 is used to generate a first candidate reply according to the first target reply template in the first target reply template set and in combination with the first user action.

[0060] The satisfaction prediction module 15 is used to introduce a user satisfaction prediction function to predict the user satisfaction of the first candidate reply to obtain a first predicted satisfaction.

[0061] The reply list acquisition module 16 is used to sort the reply priorities of the first candidate reply set formed based on the first candidate reply according to the first predicted satisfaction level to obtain a first candidate reply list.

[0062] The intelligent reply module 17 is used to extract the target candidate reply from the first candidate reply list and intelligently reply to the first user message.

[0063] The specific configuration of the message set acquisition module 12 will be described in detail below. The message set acquisition module 12 may further include: the reply message data set includes multiple reply message data sets of predetermined message types, including: analyzing the first user action string through a user message type discrimination model trained based on big data to obtain a first message type. A first reply message data set matching the first message type in the multiple reply message data sets, wherein the first reply message data set is used to traverse and filter the first user action. Wherein, the predetermined message types include at least consultation, inquiry, and complaint types.

[0064] The specific configuration of the message set acquisition module 12 will be described in detail below. The message set acquisition module 12 may further include: extracting a second user message from the user message data set, wherein the second user message has an identifier of a second message type. Performing reading and extraction on the second user message based on the predetermined reading and extraction plan to obtain a second user action string. Performing supervised learning and testing on the training data set composed of the second user action string and the second message type based on the support vector machine principle to obtain the user message type discrimination model.

[0065] The specific configuration of the satisfaction prediction module 15 will be described in detail below. The satisfaction prediction module 15 may further include: the first reply message set includes a plurality of reply message data with a reply template identifier. The plurality of reply message data with a reply template identifier are matched and screened based on the first target reply template to obtain first target message data, the first target message data including multiple reply messages with a user satisfaction identifier. The user satisfaction prediction function combines the multiple reply messages with a user satisfaction identifier to obtain the first predicted satisfaction.

[0066] The expression of the user satisfaction prediction function is as follows: ; in, It refers to the reply message among the multiple reply messages with user satisfaction indicators. Reply to the first candidate The first satisfaction support of , k refers to the magnitude of the user satisfaction prediction function, and , is the attenuation factor, and .

[0067] The specific configuration of the satisfaction prediction module 15 will be described in detail below. The satisfaction prediction module 15 may further include: determining whether the average of the satisfaction support obtained by calculating the first satisfaction support meets the support limit. If so, matching the reply message with the user satisfaction identifier in the multiple reply messages. The corresponding user satisfaction is used as the first predicted satisfaction.

[0068] The specific configuration of the satisfaction prediction module 15 will be described in detail below. The satisfaction prediction module 15 may further include: if not in compliance, issuing a manual reply instruction, and activating a manual reply based on the manual reply instruction.

[0069] The specific configuration of the intelligent reply module 17 will be described in detail below. The intelligent reply module 17 may further include: adding or modifying the templates in the first target reply template set based on the first user action string.

[0070] The message intelligent reply processing device based on the semantic analysis big model provided in the embodiment of the present invention can execute the message intelligent reply processing method based on the semantic analysis big model provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0071] Although the present application makes various references to certain modules in the apparatus according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or the server, and the various units and modules included are only divided according to functional logic, but are not limited to the above divisions, as long as the corresponding functions can be realized. In addition, the specific names of the functional units are only for the convenience of distinguishing from each other, and are not used to limit the scope of protection of the present invention.

[0072] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A message intelligent reply processing method based on a semantic analysis large model, characterized in that: include: Reading a predetermined reading and extraction plan, and performing reading and extraction on the first user message based on the predetermined reading and extraction plan to obtain a first user action string; Traversing and screening the first user action in the first user action string in the reply message data set to obtain a first reply message set; De-duplicating a first reply template set obtained by analyzing the first reply message set to obtain a first target reply template set; generating a first candidate reply according to a first target reply template in the first target reply template set and in combination with the first user action; Introducing a user satisfaction prediction function to predict user satisfaction for the first candidate response to obtain a first predicted satisfaction; According to the first predicted satisfaction level, a first candidate reply set formed based on the first candidate reply is sorted by reply priority to obtain a first candidate reply list; Extract the target candidate reply from the first candidate reply list and intelligently reply to the first user message.

2. The message intelligent reply processing method based on the semantic analysis large model according to claim 1 is characterized in that: The reply message data set includes multiple reply message data sets of predetermined message types, including: Analyze the first user action string by using a user message type discrimination model trained based on big data to obtain a first message type; a first reply message data set matching the first message type in the multiple reply message data sets, wherein the first reply message data set is used for traversal and screening of the first user action; The predetermined message types include at least consultation, inquiry and complaint.

3. The message intelligent reply processing method based on the semantic analysis large model according to claim 2 is characterized in that: include: Extracting a second user message from the user message data set, where the second user message has an identifier of a second message type; Performing reading extraction on the second user message based on the predetermined reading extraction plan to obtain a second user action string; Based on the support vector machine principle, supervised learning and verification are performed on the training data set composed of the second user action string and the second message type to obtain the user message type discrimination model.

4. According to claim 1, the message intelligent reply processing method based on the semantic analysis large model is characterized in that: include: The first reply message set includes a plurality of reply message data having a reply template identifier; Matching and screening the plurality of reply message data having the reply template identifier based on the first target reply template to obtain first target message data, wherein the first target message data includes a plurality of reply messages having the user satisfaction identifier; The user satisfaction prediction function combines the multiple reply messages with the user satisfaction identifier to obtain the first predicted satisfaction.

5. According to claim 4, the message intelligent reply processing method based on the semantic analysis large model is characterized in that: The expression of the user satisfaction prediction function is as follows: ; in, It refers to the reply message among the multiple reply messages with user satisfaction indicators. Reply to the first candidate The first satisfaction support of , k refers to the magnitude of the user satisfaction prediction function, and , is the attenuation factor, and .

6. The message intelligent reply processing method based on the semantic analysis large model according to claim 5 is characterized in that: include: Determining whether the average of the satisfactory support obtained by calculating the first satisfactory support meets the support limit; If it is consistent, the reply message matched in the multiple reply messages with user satisfaction identifiers is The corresponding user satisfaction is used as the first predicted satisfaction.

7. The message intelligent reply processing method based on the semantic analysis big model according to claim 6 is characterized in that: If not, a manual reply instruction is issued, and a manual reply is activated based on the manual reply instruction.

8. The message intelligent reply processing method based on the semantic analysis big model according to claim 7 is characterized in that: Based on the first user action string, templates in the first target reply template set are added or modified.

9. A message intelligent reply processing device based on a semantic analysis large model, characterized in that: include: An action string acquisition module, used to read a predetermined reading and extraction plan, and perform reading and extraction on the first user message based on the predetermined reading and extraction plan to obtain a first user action string; A message set acquisition module, configured to traverse and filter the first user action in the first user action string in the reply message data set to obtain a first reply message set; A template set acquisition module, configured to remove duplicates from a first reply template set obtained by analyzing the first reply message set, to obtain a first target reply template set; a candidate reply acquisition module, configured to generate a first candidate reply according to a first target reply template in the first target reply template set and in combination with the first user action; A satisfaction prediction module, used for introducing a user satisfaction prediction function to predict the user satisfaction of the first candidate reply to obtain a first predicted satisfaction; A reply list acquisition module, configured to sort the reply priorities of a first candidate reply set formed based on the first candidate reply according to the first predicted satisfaction level, to obtain a first candidate reply list; The intelligent reply module is used to extract the target candidate reply from the first candidate reply list and intelligently reply to the first user message.