Intelligent partner training method and device based on memory mechanism and processor

Abstract

The invention discloses an intelligent partner training method and device based on a memory mechanism and a processor. In the scheme, dialogue content input by a user is obtained; extracting multiple pieces of key memory information from the dialogue content, performing memory grading processing on the multiple pieces of key memory information, and constructing a memory grading information set; on the basis of information in the memory grading information set, historical memory grading information stored in a target storage area corresponding to the dialogue training partner is updated, and a target information set is obtained; extracting target key memory information matched with the dialogue content from a target information set according to the memory level; and generating a reply result by the large language model based on the target key memory information and the dialogue content, and sending the reply result to the user. According to the scheme, by introducing a memory mechanism, the content input into the large language model is subjected to structured compression on the premise of keeping core information in the dialogue partner training process, so that the effectiveness of dialogue partner training is remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an intelligent training method, device and processor based on a memory mechanism. Background Art

[0002] In the insurance industry, agent training typically requires significant investment in manpower, material resources, and time to cultivate an excellent sales agent. Traditional training methods often rely on manual practice or simulated drills. While this approach can improve agents' communication and sales skills to a certain extent, it is inefficient, costly, and difficult to cover diverse customer scenarios and complex conversation situations.

[0003] The rapid development of artificial intelligence technology, especially the significant advancements in natural language understanding and generation achieved by large language models (LLMs), has provided a new solution for insurance agent training. This solution utilizes large models to simulate real-world customer conversations and conduct multiple rounds of interactive training with inexperienced agents, thereby enabling efficient and low-cost training. In this application scenario, large models must dynamically assume different customer roles based on pre-set customer profiles, intent, and conversation status, conducting full-process conversations with agents, including preliminary interviews, product explanations, and objection handling. Agents, in turn, must flexibly apply professional language and sales strategies based on the conversational content to achieve training objectives.

[0004] Traditional methods typically concatenate the entire conversation history as input to a large language model, which then generates corresponding responses based on the entire conversation history. However, when there are many conversation turns, traditional methods often require the model to process a large amount of redundant information (such as greetings, repeated expressions, etc.). This not only consumes excessive computing resources but also significantly reduces the model's response speed, impacting the user interaction experience. To alleviate the performance pressure brought by long context, existing technologies typically adopt a strategy of truncating historical conversation content, retaining only the most recent rounds of historical conversation content as input to the large language model. However, this can result in the loss of core information (such as the customer's initial needs, objections, and emotional changes), thereby affecting the authenticity of simulated customer behavior and the effectiveness of agent training, making it difficult to achieve a high-quality, realistic training experience. Summary of the Invention

[0005] Based on the above problems, the present application provides an intelligent training method, device and processor based on a memory mechanism, the purpose of which is to perform structured compression on the content input into the large language model while retaining the core information in the dialogue training process, so as to significantly improve the effectiveness of the dialogue training.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] In a first aspect, the present application provides an intelligent training method based on a memory mechanism, the method comprising:

[0008] Acquiring the conversation content input by the user; the conversation content includes the communication data of the user during the conversation training process of the target business;

[0009] Extracting multiple key memory information from the conversation content, and performing memory classification processing on the multiple key memory information to construct a memory classification information set; the memory classification information set includes each key memory information and a memory level corresponding to the key memory information, and the memory level is used to reflect the correlation between the key memory information and the conversation training;

[0010] Based on the information in the memory hierarchical information set, the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training is updated to obtain a target information set;

[0011] Extracting target key memory information that matches the conversation content from the target information set according to the memory level;

[0012] A large language model generates a response result based on the target key memory information and the conversation content, and sends the response result to the user.

[0013] In an optional implementation, performing memory hierarchical processing on the plurality of key memory information to construct a memory hierarchical information set includes:

[0014] Based on a preset relevance evaluation rule, evaluating the relevance score between each of the key memory information and the conversation training;

[0015] According to the association scores corresponding to the respective key memory information, corresponding memory levels are assigned to the key memory information to obtain the memory classification information set.

[0016] In an optional implementation, the updating of the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training based on the information in the memory hierarchical information set to obtain the target information set includes:

[0017] storing the information in the memory hierarchical information set into the target storage area, and detecting whether there is contextually contradictory information in all the memory hierarchical information in the target storage area;

[0018] If the context-conflicting information exists in all the memory hierarchical information, correcting the context-conflicting information in all the memory hierarchical information, and constructing the target information set based on the corrected memory hierarchical information;

[0019] If the context-conflicting information does not exist in all the memory hierarchical information, the target information set is constructed based on all the memory hierarchical information.

[0020] In an optional implementation, extracting target key memory information that matches the conversation content from the target information set based on the memory level includes:

[0021] Dividing the target information set into a first set and a second set according to the memory level; the first set includes a plurality of first key memory information, and the second set includes a plurality of second key memory information, the memory level corresponding to each of the first key memory information is greater than a preset level, and the memory level corresponding to each of the second key memory information is less than or equal to the preset level;

[0022] Inputting the information in the second set and the conversation content into a large language model, and extracting a plurality of third key memory information matching the conversation content from the second set based on the contextual semantics of the conversation content by the large language model;

[0023] Integrate the plurality of the first key memory information and the plurality of the third key memory information to generate the target key memory information.

[0024] In an optional implementation, after obtaining the conversation content input by the user, the method further includes:

[0025] Based on the unique identification information of the conversation training, detecting whether there is a target storage area corresponding to the conversation training in a preset database;

[0026] If the target storage area exists in the preset database, directly storing the conversation content in the target storage area;

[0027] If the target storage area does not exist in the preset database, the target storage area is created in the preset database based on the unique identification information, and the conversation content is stored in the target storage area.

[0028] In a second aspect of the present application, there is provided an intelligent training device based on a memory mechanism, the device comprising:

[0029] An acquisition module, configured to acquire the conversation content input by the user; the conversation content includes communication data of the user during the conversation training process of the target business;

[0030] a memory grading module, configured to extract a plurality of key memory information from the conversation content, and perform memory grading processing on the plurality of key memory information to construct a memory grading information set; the memory grading information set includes each key memory information and a memory grade corresponding to the key memory information, the memory grade being used to reflect the degree of association between the key memory information and the conversation training partner;

[0031] An updating module, configured to update the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training based on the information in the memory hierarchical information set, to obtain a target information set;

[0032] an extraction module, configured to extract target key memory information matching the conversation content from the target information set according to the memory level;

[0033] A generation module is used to generate a response result based on the target key memory information and the conversation content by a large language model, and send the response result to the user.

[0034] Optionally, the memory classification module includes:

[0035] An evaluation unit, configured to evaluate the relevance score between each of the key memory information and the conversation training based on a preset relevance evaluation rule;

[0036] The allocation unit is used to allocate corresponding memory levels to the key memory information according to the association scores corresponding to the respective key memory information, so as to obtain the memory classification information set.

[0037] Optionally, the update module includes:

[0038] a detection unit, configured to store the information in the memory hierarchical information set into the target storage area, and detect whether there is contextually contradictory information in all the memory hierarchical information in the target storage area;

[0039] a correction unit, configured to correct the context-conflicting information if any of the context-conflicting information exists in all the memory hierarchical information, to obtain the target information set;

[0040] A construction unit is configured to construct the target information set based on all the memory hierarchical information if the context-conflicting information does not exist in all the memory hierarchical information.

[0041] In a third aspect of the present application, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the above-mentioned intelligent training method based on the memory mechanism is implemented.

[0042] In a fourth aspect of the present application, a processor is provided for running a computer program, which executes the above-mentioned intelligent training method based on the memory mechanism when the computer program is running.

[0043] Compared with the existing technology, this application has the following beneficial effects:

[0044] In the technical solution of the present application, by obtaining the conversation content input by the user, wherein the conversation content includes the communication data of the user in the conversation training process of the target business; then extracting multiple key memory information from the conversation content, and performing memory classification processing on the multiple key memory information to construct a memory classification information set. Since the memory classification information set includes each key memory information and the memory level corresponding to the key memory information, and the memory level is used to reflect the correlation between the key memory information and the conversation training, multiple key memory information is extracted from the conversation content, and their importance is distinguished through classification processing, and redundant content (such as greetings, repeated expressions) is effectively filtered out, and only the core information of the current conversation training process is retained, which greatly shortens the amount of data input to the large language model;

[0045] Then, based on the information in the memory hierarchical information set, the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training is updated to obtain the target information set, which realizes the retention and updating of core information in the dialogue training process through dynamic management, avoiding the problem of core information loss; then, according to the memory level, the target key memory information matching the dialogue content is extracted from the target information set, realizing structured compression of the content input to the large language model, greatly shortening the input content of the large prediction model; thus, the large language model can quickly focus on the core logic of the dialogue based on the compressed core information (i.e., the target key memory information) and the dialogue content, and generate a reply result that is highly consistent with the context, greatly improving the response speed; and then sending the accurate reply result to the user, significantly improving the effectiveness of the dialogue training, solving the problem of difficult to balance computational efficiency and information integrity in traditional methods, and providing a more efficient and intelligent solution for dialogue training in complex business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0047] Figure 1 A flowchart of an intelligent training method based on a memory mechanism provided in an embodiment of the present application;

[0048] Figure 2 A topological diagram of an intelligent training service provided in an embodiment of the present application;

[0049] Figure 3 A flowchart of another intelligent training method based on a memory mechanism provided in an embodiment of the present application;

[0050] Figure 4 A flowchart of a process for updating historical memory hierarchical information provided in an embodiment of the present application;

[0051] Figure 5 A flowchart of a process for extracting target key memory information provided in an embodiment of the present application;

[0052] Figure 6 A schematic structural diagram of an intelligent training device based on a memory mechanism provided in an embodiment of the present application. DETAILED DESCRIPTION

[0053] As described above, traditional methods typically concatenate the entire conversation history as input to a large language model, which then generates corresponding responses based on the entire conversation history. However, when there are many conversation turns, traditional methods often require the model to process a large amount of redundant information (such as greetings, repeated expressions, etc.). This not only consumes excessive computing resources but also significantly reduces the model's response speed, impacting the user interaction experience. To alleviate the performance pressure caused by long context, existing technologies typically adopt a strategy of truncating historical conversation content, retaining only the most recent rounds of historical conversation content as input to the large language model. However, this can result in the loss of core information (such as the customer's initial needs, objections, and emotional changes), thereby affecting the authenticity of simulated customer behavior and the effectiveness of agent training, making it difficult to achieve a high-quality, realistic training experience.

[0054] After research, the inventors proposed an intelligent training method based on a memory mechanism. The solution obtains the conversation content input by the user, wherein the conversation content includes the communication data of the user during the conversation training process of the target business; then extracts multiple key memory information from the conversation content, and performs memory classification processing on the multiple key memory information to construct a memory classification information set. Since the memory classification information set includes each key memory information and the memory level corresponding to the key memory information, and the memory level is used to reflect the correlation between the key memory information and the conversation training, multiple key memory information is extracted from the conversation content, and their importance is distinguished through classification processing, which effectively filters out redundant content (such as greetings, repeated expressions), retains only the core information in the current conversation training process, and greatly reduces the amount of data input to the large language model;

[0055] Then, based on the information in the memory hierarchical information set, the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training is updated to obtain the target information set, which realizes the retention and updating of core information in the dialogue training process through dynamic management, avoiding the problem of core information loss; then, according to the memory level, the target key memory information matching the dialogue content is extracted from the target information set, realizing structured compression of the content input to the large language model, greatly shortening the input content of the large prediction model; thus, the large language model can quickly focus on the core logic of the dialogue based on the compressed core information (i.e., the target key memory information) and the dialogue content, and generate a reply result that is highly consistent with the context, greatly improving the response speed; and then sending the accurate reply result to the user, significantly improving the effectiveness of the dialogue training, solving the problem of difficult to balance computational efficiency and information integrity in traditional methods, and providing a more efficient and intelligent solution for dialogue training in complex business scenarios.

[0056] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0057] Method Example

[0058] The embodiment of the present application provides an embodiment of an intelligent training method based on a memory mechanism. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0059] See also Figure 1 , which is a flow chart of an intelligent training method based on a memory mechanism provided by an embodiment of the present application, such as Figure 1 As shown, the method includes the following steps:

[0060] Step S101: Acquire the conversation content input by the user.

[0061] In an optional embodiment, an intelligent training system based on a memory mechanism can be used as the execution subject of the intelligent training method based on a memory mechanism in the embodiment of the present application. For the convenience of description, the intelligent training system based on a memory mechanism is referred to as a system below. Figure 2 As shown, the system includes a memory initialization module, a memory management module, a memory search module and a memory update module; the memory initialization module is used to initialize the conversation training so that data isolation is achieved between each conversation training service; the memory management module is used to perform multi-dimensional memory management and conflict self-healing on the conversation content; the memory search module is used to search for key memories that match the conversation content; the memory update module is used to update the response results generated by the large language model to a preset database.

[0062] Alternatively, as Figure 3 As shown, the user request is load-balanced to the training service. Since this application manages the context of the large language model (i.e., the conversation content input into the large language model), the above-mentioned intelligent training system based on the memory mechanism is deployed as a separate service between the training service and the large model service, and is used separately as a memory management service.

[0063] In step S101, the conversation content includes the communication data of the user during the conversation training process of the target business. For example, the user enters "Hello, Ms. Zhang, are you free this weekend? I will go to the coffee shop downstairs from your house to introduce you to product xxx."

[0064] In the embodiment of the present application, the conversation content input by the user may be text information or voice information; if the conversation content input by the user is voice information, the system may convert the voice information into corresponding text information.

[0065] In order to achieve data isolation between various conversation training services and avoid resource contention and performance interference, in an embodiment of the present application, after the system obtains the conversation content input by the user, it can detect whether there is a target storage area corresponding to the conversation training in the preset database based on the unique identification information of the conversation training; if the target storage area exists in the preset database, the system directly stores the conversation content in the target storage area; if the target storage area does not exist in the preset database, the system creates a target storage area in the preset database based on the unique identification information, and stores the conversation content in the target storage area.

[0066] In an embodiment of the present application, the unique identification information of the conversation training session may be the task_id of the conversation training session, which is the task number of the entire conversation training session. For example, after obtaining the conversation content input by the user, the system may detect whether there is a target storage area corresponding to the task_id in a preset database (e.g., a structured database) based on the task_id of the conversation content. If there is a target storage area corresponding to the task_id in the preset database, the system may store the conversation content in the target storage area based on the sentence_id of the conversation content; if there is no target storage area corresponding to the task_id in the preset database, the system may perform initialization processing to create a corresponding target storage area for the user's conversation training session based on the task_id of the conversation content input by the user, and store the conversation content in the target storage area.

[0067] Step S102: extract multiple key memory information from the conversation content, perform memory classification processing on the multiple key memory information, and construct a memory classification information set.

[0068] In step S102, the memory classification information set includes each key memory information and the memory level corresponding to the key memory information. The memory level is used to reflect the correlation between the key memory information and the conversation training. The higher the correlation, the higher the memory level, and the lower the correlation, the higher the memory level.

[0069] In an embodiment of the present application, memory grading is used to perform multi-dimensional memory management on user-entered conversation content, grading the user-entered conversation content on a scale of 1-10 in terms of importance. For example, the role information played by the large model is rated at level 10, the objection is rated at level 8, ..., and the small talk is rated at level 1. The system can use the large language model to extract multiple key memory information from the conversation content and perform memory grading on the multiple key memory information based on the corresponding relevance scores of each key memory information, thereby constructing a memory grading information set. For example, if the conversation content "Hello, Miss Zhang, are you free this weekend? I'll go to the coffee shop downstairs to introduce you to product xxx" is input into the large language model, the large language model can extract multiple key memory information from the conversation content {User: Miss Zhang, Intention: product xxx}; then, based on the corresponding relevance scores of each key memory information, the system can perform memory grading on the multiple key memory information, thereby constructing a memory grading information set {"Level: 10, User: Miss Zhang", "Level: 9, Intention: product xxx"}.

[0070] Specifically, the system can evaluate the correlation scores between each key memory information and the dialogue partner based on the preset correlation evaluation rules; according to the correlation scores corresponding to each key memory information, the key memory information is assigned a corresponding memory level to obtain a memory graded information set.

[0071] In an embodiment of the present application, the system can extract the topic of the conversation training of the target business based on the preset relevance evaluation rules, and evaluate the similarity between each key memory information and the topic; then, based on the similarity, determine the correlation score between each key memory information and the conversation training, and based on the correlation score corresponding to each key memory information, assign a corresponding memory level to the key memory information to obtain a memory grade information set. For example, from the conversation content, multiple key memory information {User: Zhang Jie, Intention: xxx product} are extracted, the correlation score of User: Zhang Jie is 99, and the correlation score of Intention: xxx product is 96. The memory levels assigned by the system according to the correlation score are {"Level: 10, User: Zhang Jie", "Level: 9, Intention: xxx product"}.

[0072] It should be noted that multiple key memory information is extracted from the conversation content, and its importance is distinguished through hierarchical processing, which effectively filters out redundant content (such as greetings and repeated expressions). Only the core information in the current conversation training process is retained, which greatly reduces the amount of data input into the large language model.

[0073] Step S103: Based on the information in the memory hierarchical information set, the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training is updated to obtain a target information set.

[0074] In an embodiment of the present application, the system can merge the latest key memory information (i.e., the information in the memory hierarchical information set) with the historical memory hierarchical information and correct the conflicting information to update the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training, thereby achieving the retention and update of core information in the dialogue training process through dynamic management, avoiding the problem of core information loss.

[0075] Specifically, see Figure 4 , which is a flowchart of a process for updating historical memory hierarchical information provided by an embodiment of the present application, and the process includes the following steps:

[0076] Step S1031: store the information in the memory hierarchical information set into a target storage area, and detect whether there is contextually contradictory information in all the memory hierarchical information in the target storage area.

[0077] In the present embodiment, contextually contradictory information refers to information that is clearly inconsistent, such as misidentification of agents, confusing names, inconsistent timelines, logical errors, inconsistent character behavior, etc. After the system stores the information in the memory hierarchy information set into the target storage area, the large language model can automatically detect whether there is contextually contradictory information in all the memory hierarchy information in the target storage area, thereby avoiding redundant or erroneous memory information in all the memory hierarchy information in the target storage area, thereby providing an accurate data foundation for dialogue training.

[0078] Step S1032: If there is context-contradictory information in all the memory hierarchical information, the context-contradictory information in all the memory hierarchical information is corrected, and a target information set is constructed based on the corrected memory hierarchical information.

[0079] In an embodiment of the present application, the system can use a large language model to correct contextually contradictory information in all memory hierarchical information. For example, the large language model extracted from the existing conversation content may play the role of a single customer, but the agent's subsequent chat provides incorrect information, such as married. This information extracted as memory information conflicts with the original memory information, and the system can directly discard the incorrect information. For another example, the user played by the large model is called Li Ting, a female, and the agent can call her Sister Li, Sister Ting, or Ms. Li. The system will directly merge the memory information.

[0080] Step S1033: If there is no context-contradictory information in all the memory hierarchical information, a target information set is constructed based on all the memory hierarchical information.

[0081] It should be noted that by dynamically managing, retaining, and updating the core information in the dialogue training process, the problem of core information loss caused by the existing technology only retaining the historical dialogue content of the most recent rounds as the input of the large language model is solved, thereby providing an accurate data foundation for subsequent dialogue training.

[0082] Step S104: extract target key memory information that matches the conversation content from the target information set according to the memory level.

[0083] In order to achieve structured compression of the content input to the large language model and significantly shorten the input content of the large prediction model, in an embodiment of the present application, the system can extract target key memory information that matches the conversation content from the target information set based on the memory level.

[0084] Specifically, see Figure 5 , which is a flow chart of a process for extracting target key memory information provided by an embodiment of the present application, and the process includes the following steps:

[0085] Step S1041, divide the target information set into a first set and a second set according to the memory level; the first set includes multiple first key memory information, and the second set includes multiple second key memory information, the memory level corresponding to each first key memory information is greater than the preset level, and the memory level corresponding to each second key memory information is less than or equal to the preset level.

[0086] For example, the system may classify key memory information with a memory level greater than level 5 into a first set, and classify key memory information with a memory level less than or equal to level 5 into a second set.

[0087] It should be noted that dividing the target information set into the first set and the second set according to the memory level can achieve accurate context matching based on the correlation between key memory information and dialogue training, thereby providing an accurate data basis for subsequent dialogue training.

[0088] Step S1042: Input the information in the second set and the conversation content into the large language model, and the large language model extracts a plurality of third key memory information matching the conversation content from the second set based on the contextual semantics of the conversation content.

[0089] In an embodiment of the present application, the system can analyze the contextual semantics of the conversation content through the semantic understanding ability of the large language model, and extract multiple third key memory information matching the conversation content from the second set based on the contextual semantics of the conversation content.

[0090] Step S1043 , integrating the plurality of first key memory information and the plurality of third key memory information to generate target key memory information.

[0091] It should be noted that by extracting the target key memory information that matches the conversation content from the target information set based on the memory level and the semantic understanding ability of the large language model, it is possible to significantly shorten the input content of the large prediction model while effectively retaining the core information in the conversation training process, thereby enabling the large language model to quickly focus on the core logic of the conversation and provide accurate data preparation for generating response results that are highly consistent with the context.

[0092] In step S105 , the large language model generates a response result based on the target key memory information and the conversation content, and sends the response result to the user.

[0093] In an embodiment of the present application, the system can generate a reply result based on the target key memory information and the conversation content through a large language model, quickly focus on the core logic of the conversation, and generate a reply result that is highly consistent with the context, greatly improving the response speed; and then send the accurate reply result to the user, significantly improving the effectiveness of the conversation training, solving the problem of balancing computational efficiency and information integrity in traditional methods, and providing a more efficient and intelligent solution for conversation training in complex business scenarios.

[0094] Alternatively, as Figure 2 As shown, after sending the reply result to the user, the system can also update the reply result to the target storage area through the memory update module to provide a data basis for the next round of dialogue practice.

[0095] The intelligent training method based on the memory mechanism provided by the embodiment of the present application realizes the extraction of multiple key memory information from the conversation content, distinguishes its importance through hierarchical processing, effectively filters out redundant content (such as greetings, repeated expressions), and retains only the core information in the current conversation training process, greatly shortening the amount of data input to the large language model; it realizes the retention and update of the core information in the conversation training process through dynamic management, avoiding the problem of core information loss; it realizes the structured compression of the content input to the large language model, greatly shortening the input content of the large prediction model; thus, the large language model can quickly focus on the core logic of the conversation based on the compressed core information (i.e., the target key memory information) and the conversation content, and generate a reply result that is highly consistent with the context, greatly improving the response speed; and then sending the accurate reply result to the user, significantly improving the effectiveness of the conversation training, solving the problem of difficulty in balancing computational efficiency and information integrity in traditional methods, and providing a more efficient and intelligent solution for conversation training in complex business scenarios.

[0096] Device embodiment

[0097] The embodiment of the present application provides an intelligent training device based on a memory mechanism, wherein: Figure 6 A schematic diagram of the structure of an intelligent training device based on a memory mechanism provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the device includes: an acquisition module 11, a memory classification module 12, an update module 13, an extraction module 14 and a generation module 15. Figure 6 You can see the connection relationship between several modules.

[0098] The acquisition module 11 is used to acquire the conversation content input by the user; the conversation content includes the communication data of the user during the conversation training process of the target business;

[0099] A memory grading module 12 is configured to extract multiple key memory information from the conversation content, perform memory grading on the multiple key memory information, and construct a memory grading information set; the memory grading information set includes each key memory information and the memory grade corresponding to the key memory information, and the memory grade is used to reflect the correlation between the key memory information and the conversation training;

[0100] An updating module 13 is configured to update the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training based on the information in the memory hierarchical information set to obtain a target information set;

[0101] An extraction module 14 is used to extract target key memory information that matches the conversation content from the target information set based on the memory level;

[0102] The generation module 15 is used to generate a response result based on the target key memory information and the conversation content by the large language model, and send the response result to the user.

[0103] Optionally, the memory grading module includes: an evaluation unit and an evaluation unit.

[0104] An evaluation unit, configured to evaluate the relevance scores between each key memory information and the conversation training based on preset relevance evaluation rules;

[0105] The allocation unit is used to allocate corresponding memory levels to key memory information according to the association scores corresponding to each key memory information, and obtain a memory classification information set.

[0106] Optionally, the update module includes: a detection unit, an evaluation unit and a construction unit.

[0107] a detection unit, configured to store information in the memory hierarchical information set into a target storage area, and detect whether there is contextually contradictory information in all the memory hierarchical information in the target storage area;

[0108] an evaluation unit, configured to correct the context-contradictory information in all the memory hierarchical information if there is context-contradictory information in all the memory hierarchical information, and construct a target information set based on the corrected memory hierarchical information;

[0109] A construction unit is used to construct a target information set based on all memory hierarchical information if there is no context-contradictory information in all memory hierarchical information.

[0110] Optionally, the extraction module includes: a construction unit, an extraction unit and an integration unit.

[0111] a division unit, configured to divide the target information set into a first set and a second set according to a memory level; the first set includes a plurality of first key memory information, the second set includes a plurality of second key memory information, the memory level corresponding to each first key memory information is greater than a preset level, and the memory level corresponding to each second key memory information is less than or equal to the preset level;

[0112] an extraction unit, configured to input the information in the second set and the conversation content into the large language model, and extract, by the large language model based on the contextual semantics of the conversation content, a plurality of third key memory information that matches the conversation content from the second set;

[0113] The integration unit is used to integrate multiple first key memory information and multiple third key memory information to generate target key memory information.

[0114] Optionally, the intelligent training device based on the memory mechanism further includes: a detection module, a storage module and a creation module.

[0115] A detection module is used to detect whether there is a target storage area corresponding to the dialogue training in the preset database based on the unique identification information of the dialogue training after obtaining the dialogue content input by the user;

[0116] A storage module, configured to directly store the conversation content in the target storage area if the target storage area exists in the preset database;

[0117] The creation module is configured to create a target storage area in the preset database based on the unique identification information if the target storage area does not exist in the preset database, and store the conversation content in the target storage area.

[0118] Storage medium embodiment

[0119] The present application provides a computer-readable storage medium having a program stored thereon. When executed by a processor, the program implements some or all of the steps of the intelligent training method using a memory mechanism described in the aforementioned method embodiment of the present application. The storage medium can be any medium capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0120] Processor Embodiments

[0121] An embodiment of the present application provides a processor for running a program, wherein, when the program is running, some or all of the steps in the intelligent training method of the memory mechanism introduced in the aforementioned method embodiment are executed.

[0122] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components indicated as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0123] The above is only one specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An intelligent training method based on memory mechanism, characterized in that: include: Get the conversation content entered by the user; The conversation content includes communication data of the user during the conversation training process of the target service; Extracting multiple key memory information from the conversation content, and performing memory classification processing on the multiple key memory information to construct a memory classification information set; the memory classification information set includes each key memory information and a memory level corresponding to the key memory information, and the memory level is used to reflect the correlation between the key memory information and the conversation training; Based on the information in the memory hierarchical information set, the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training is updated to obtain a target information set; Extracting target key memory information that matches the conversation content from the target information set according to the memory level; A large language model generates a response result based on the target key memory information and the conversation content, and sends the response result to the user.

2. The method according to claim 1, characterized in that The step of performing memory hierarchical processing on the plurality of key memory information to construct a memory hierarchical information set includes: Based on a preset relevance evaluation rule, evaluating the relevance score between each of the key memory information and the conversation training; According to the association scores corresponding to the respective key memory information, corresponding memory levels are assigned to the key memory information to obtain the memory classification information set.

3. The method according to claim 1, characterized in that The method of updating the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training based on the information in the memory hierarchical information set to obtain the target information set includes: storing the information in the memory hierarchical information set into the target storage area, and detecting whether there is contextually contradictory information in all the memory hierarchical information in the target storage area; If the context-conflicting information exists in all the memory hierarchical information, correcting the context-conflicting information in all the memory hierarchical information, and constructing the target information set based on the corrected memory hierarchical information; If the context-conflicting information does not exist in all the memory hierarchical information, the target information set is constructed based on all the memory hierarchical information.

4. The method according to claim 1, wherein The step of extracting target key memory information matching the conversation content from the target information set according to the memory level includes: Dividing the target information set into a first set and a second set according to the memory level; the first set includes a plurality of first key memory information, and the second set includes a plurality of second key memory information, the memory level corresponding to each of the first key memory information is greater than a preset level, and the memory level corresponding to each of the second key memory information is less than or equal to the preset level; Inputting the information in the second set and the conversation content into a large language model, and extracting a plurality of third key memory information matching the conversation content from the second set based on the contextual semantics of the conversation content by the large language model; Integrate the plurality of the first key memory information and the plurality of the third key memory information to generate the target key memory information.

5. The method according to claim 1, characterized in that After obtaining the conversation content input by the user, the method further includes: Based on the unique identification information of the conversation training, detecting whether there is a target storage area corresponding to the conversation training in a preset database; If the target storage area exists in the preset database, directly storing the conversation content in the target storage area; If the target storage area does not exist in the preset database, the target storage area is created in the preset database based on the unique identification information, and the conversation content is stored in the target storage area.

6. An intelligent training device based on a memory mechanism, characterized in that: include: The acquisition module is used to obtain the conversation content input by the user; The conversation content includes communication data of the user during the conversation training process of the target service; a memory grading module, configured to extract a plurality of key memory information from the conversation content, and perform memory grading processing on the plurality of key memory information to construct a memory grading information set; the memory grading information set includes each key memory information and a memory grade corresponding to the key memory information, the memory grade being used to reflect the degree of association between the key memory information and the conversation training partner; An updating module, configured to update the historical memory hierarchical information stored in the target storage area corresponding to the dialogue training based on the information in the memory hierarchical information set, to obtain a target information set; an extraction module, configured to extract target key memory information matching the conversation content from the target information set according to the memory level; A generation module is used to generate a response result based on the target key memory information and the conversation content by a large language model, and send the response result to the user.

7. The device according to claim 6, characterized in that The memory classification module includes: An evaluation unit, configured to evaluate the relevance score between each of the key memory information and the conversation training based on a preset relevance evaluation rule; The allocation unit is used to allocate corresponding memory levels to the key memory information according to the association scores corresponding to the respective key memory information, so as to obtain the memory classification information set.

8. The device according to claim 6, characterized in that The update module includes: a detection unit, configured to store the information in the memory hierarchical information set into the target storage area, and detect whether there is contextually contradictory information in all the memory hierarchical information in the target storage area; a correction unit, configured to correct the context-conflicting information if any of the context-conflicting information exists in all the memory hierarchical information, to obtain the target information set; A construction unit is configured to construct the target information set based on all the memory hierarchical information if the context-conflicting information does not exist in all the memory hierarchical information.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the intelligent training method based on the memory mechanism as described in any one of claims 1 to 5 is implemented.

10. A processor, characterized in that: Used to run a computer program, which, when running, executes the intelligent training method based on the memory mechanism as described in any one of claims 1 to 5.

Images