Large model-based simulation exercise method, system and computer device

Abstract

The application relates to the technical field of artificial intelligence, and particularly discloses a simulation training method and system based on a large model and computer equipment, which comprises the following steps: obtaining training configuration information; repeatedly performing a simulation training step, outputting corresponding reply information based on a large model according to the training configuration information and the dialogue information input by a trainer, and completing the simulation dialogue with the trainer until the simulation dialogue is completed; the simulation training step comprises the following steps: obtaining single dialogue information input by the trainer in the current round; determining a current task stage by analyzing the single dialogue information through the large model according to training dialogue constraint information; generating stage-by-stage trainer role constraint information according to the current task stage and the training dialogue constraint information; and outputting single reply information based on the large model according to historical dialogue records, the stage-by-stage trainer role constraint information and the training configuration information. A strong control mechanism and a constraint mechanism are introduced in the simulation training based on the role playing of the large model, so that the large model can carry out the simulation training according to preset information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a simulation training method and system based on a large model and a computer device. BACKGROUND

[0002] Simulation training is widely used in different scenarios such as shopping guide training, sales training, and employee training. Through the process of simulation training, users can practice the situations they may encounter in the corresponding scenarios in real time, which helps users improve their response level during actual operation. Currently, simulation training systems based on large models usually simply use large models to perform role-playing simulation trainers, but the roles played by large models have many uncontrollable problems, which makes the simulation training effect poor and the trainees do not get good training. SUMMARY

[0003] Therefore, it is necessary to provide a simulation training method, system and computer device based on a large model in view of the many uncontrollable problems of the roles played by the large model.

[0004] A simulation training method based on a large model includes obtaining training configuration information; the training configuration information includes trainee role information, trainer role information, training dialogue control information, and training dialogue constraint information; repeatedly performing a simulation training step, outputting corresponding reply information based on a large model according to the training configuration information and the dialogue information input by the trainee, until completing the simulation dialogue with the trainee; the simulation training step includes obtaining single dialogue information input by the trainee in the current round; determining the current task phase by analyzing the single dialogue information through the large model according to the training dialogue constraint information; generating phase-based trainer role constraint information according to the current task phase and the training dialogue constraint information; outputting single reply information based on the large model according to the historical dialogue record, the phase-based trainer role constraint information, and the training configuration information.

[0005] In one embodiment, after obtaining the single dialogue information input by the trainee in the current round, the method further includes analyzing whether the single dialogue information has an abnormal situation through the large model according to the trainee role information and / or the training dialogue control information.

[0006] In one of the embodiments, the rehearsal dialogue control information comprises a rehearsal dialogue time upper limit and a rehearsal dialogue round upper limit, the abnormal situation comprises a dialogue time abnormality, and the analysis of whether the single dialogue information has the abnormal situation by the large model according to the rehearser role information and / or the rehearsal dialogue control information comprises a comparison of a dialogue round in the current simulation rehearsal with the rehearsal dialogue round upper limit and / or a comparison of a dialogue time in the current simulation rehearsal with the rehearsal dialogue time upper limit; when the comparison result is that the dialogue round in the current simulation rehearsal is greater than the rehearsal dialogue round upper limit and / or the dialogue time in the current simulation rehearsal is greater than the rehearsal dialogue time upper limit, it is judged that there is a dialogue time abnormality and corresponding prompt feedback information is generated; when the comparison result is that the dialogue round in the current simulation rehearsal is less than or equal to the rehearsal dialogue round upper limit and the dialogue time in the current simulation rehearsal is less than or equal to the rehearsal dialogue time upper limit, it is judged that there is no dialogue time abnormality.

[0007] In one of the embodiments, the abnormal situation comprises inappropriate language, and the analysis of whether the single dialogue information has the abnormal situation by the large model according to the rehearser role information and / or the rehearsal dialogue control information comprises semantic detection of the single dialogue information by a large language model to determine whether there is inappropriate language in the single dialogue information; when there is inappropriate language in the single dialogue information, corresponding prompt feedback information is generated; and when there is no inappropriate language in the single dialogue information, it is determined that the language in the single dialogue information is normal.

[0008] In one of the embodiments, the abnormal situation comprises malicious deception, and the analysis of whether the single dialogue information has the abnormal situation by the large model according to the rehearser role information and / or the rehearsal dialogue control information comprises a determination of whether there is a malicious deception intention in the single dialogue information by an intention recognition model according to the rehearser role information and the rehearsal dialogue control information; when there is a malicious deception intention in the single dialogue information, corresponding prompt feedback information is generated; and when there is no malicious deception intention in the single dialogue information, it is determined that the intention in the single dialogue information is normal.

[0009] In one of the embodiments, after the output of the single reply information based on the large model according to the historical dialogue record, the stage-by-stage accompanying trainer role constraint information and the accompanying trainer role information, the method further comprises saving the single dialogue information and the single reply information of the current round; and generating new historical dialogue record by combining the single dialogue information and the single reply information of the current round with the single dialogue information and the single reply information of previous rounds.

[0010] A large model-based simulation training system, comprising a training configuration module configured to obtain training configuration information; the training configuration information comprises trainer role information, training partner role information, training dialogue control information, and training dialogue constraint information; a training dialogue exception processing module connected to the training configuration module, configured to obtain single dialogue information input by the trainer, the trainer role information, and the training dialogue control information, and further configured to analyze whether the single dialogue information has an abnormal situation based on the trainer role information and / or the training dialogue control information through the large model; a training dialogue constraint module connected to the training configuration module and the training dialogue exception processing module, configured to obtain the single dialogue information and the training dialogue constraint information, and further configured to determine a current task stage based on the single dialogue information and the training dialogue constraint information through the large model, and generate stage-based training partner role constraint information based on the current task stage and the training dialogue constraint information; and a training partner module connected to the training configuration module and the training dialogue constraint module, configured to obtain historical dialogue records, the training partner role information, the single dialogue information, and the stage-based training partner role constraint information, and further configured to output single reply information based on the large model based on the historical dialogue records, the stage-based training partner role constraint information, and the training partner role information.

[0011] In one embodiment, the large model-based simulation training system further comprises a training historical dialogue record module connected to the training partner module, configured to save the single dialogue information and the single reply information of the current round, and generate new historical dialogue records by combining the single dialogue information and the single reply information of the current round with the single dialogue information and the single reply information of previous rounds.

[0012] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the large model-based simulation training method of any one of the above embodiments when executing the computer program.

[0013] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the large model-based simulation training method of any one of the above embodiments.

[0014] The simulation training method based on the large model obtains training configuration information configured by a user, and outputs corresponding reply information for dialogue information input by a trainer based on the training configuration information by using a large model until a simulation dialogue with the trainer is completed. During the simulation training, according to training dialogue constraint information, a current task stage is determined by analyzing single dialogue information by using the large model, and corresponding stage training role constraint information is generated. The stage training role constraint information is used to constrain the role playing state of the large model, and single reply information for single dialogue information is output based on the large model according to historical dialogue records, the stage training role constraint information and the training configuration information. By introducing a strong control mechanism and a constraint mechanism in the simulation training based on the role playing of the large model, it can be ensured that the large model carries out the simulation training according to the preset training information, ensures the stability of the simulation training dialogue, and improves the simulation training effect. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present specification or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present specification, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 The application environment schematic diagram of the simulation training method based on the large model in one of the embodiments of the present application;

[0017] Figure 2 The flowchart of the simulation training method based on the large model in one of the embodiments of the present application;

[0018] Figure 3 The flowchart of the simulation training method based on the large model in another embodiment of the present application;

[0019] Figure 4 The flowchart of the simulation training method based on the large model in one of the embodiments of the present application;

[0020] Figure 5 The flowchart of the simulation training method based on the large model in another embodiment of the present application;

[0021] Figure 6 The flowchart of the simulation training method based on the large model in another embodiment of the present application;

[0022] Figure 7 The flowchart of the simulation training method based on the large model in another embodiment of the present application;

[0023] Figure 8A structural schematic diagram of a large model-based simulation and practice system in one of the embodiments of the present application;

[0024] Figure 9 A structural schematic diagram of a large model-based simulation and practice system in another embodiment of the present application;

[0025] Figure 10 An internal structural diagram of a computer device in one of the embodiments of the present application. DETAILED DESCRIPTION

[0026] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided so that the disclosure of the present application can be more thoroughly and completely understood.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description of the present application herein is for the purpose of describing the specific embodiments only and is not intended to be limiting of the present application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0028] Figure 1 An application environment schematic diagram of a large model-based simulation and practice method in one of the embodiments of the present application, the large model-based simulation and practice method provided by the embodiments of the present application can be applied in the application environment as shown. Figure 1 The terminal 102 communicates with the server 104 through the network. The data storage system can store the data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart televisions, smart air conditioners, smart vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be realized by an independent server or a server cluster composed of multiple servers.

[0029] It should be further noted that the related information (including but not limited to user input information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties.

[0030] There are various factors that cause large models to be uncontrollable in current simulation systems. For example, in the case of multiple rounds of dialogue with a role played by a large model, the role played by the large model becomes unstable and is more likely to deviate from the role setting as the dialogue continues. Secondly, due to the lack of constraints, the role played by the large model lacks chapter and is usually not subject to the set action framework compared to human sparring. In addition, due to the lack of control, the role played by the large model is too cooperative and will cooperate with the user input regardless of the information, and even agree to some unreasonable requirements.

[0031] To solve one or more of the above problems, the embodiments of the present application provide a large model-based simulation method, system and computer device. Figure 2 For the flowchart of the large model-based simulation method in one of the embodiments of the present application, in one of the embodiments, as shown in Figure 2 , a large model-based simulation method is provided. Taking the application environment in Figure 1 as an example for illustration, the method can include the following steps S100 to S200.

[0032] Step S100: Obtain simulation configuration information; the simulation configuration information includes role information of a simulator, role information of a sparring partner, simulation dialogue control information, and simulation dialogue constraint information.

[0033] Before carrying out the dialogue simulation, the user can first input the simulation configuration information to set the related information of this dialogue. The simulation configuration information input by the user is received and stored. In this embodiment, the simulation configuration information can include but is not limited to the role information of the simulator, the role information of the sparring partner, the simulation dialogue control information, and the simulation dialogue constraint information.

[0034] The role information of the simulator can refer to the role information to be played by the simulator (e.g., the user). The role information of the simulator can include but is not limited to the role of the simulator, the theme of the simulator, and the background of the simulator. For example, taking the dialogue scene between a pharmaceutical representative and a doctor as an example, the role of the simulator can be a pharmaceutical representative, and the role of the sparring partner can be a doctor. The large model can construct the global role image of the simulator based on the role information of the simulator, and at the same time, the simulator can be more clear about his task and background during the simulation. In addition, the large model can also assist the simulator to determine whether the input content deviates from the role setting of the simulator or deviates from the theme of the simulator based on the role information of the simulator.

[0035] The role information of the accompanying trainer can refer to the role information to be played by the accompanying trainer (e.g., a large model). The role information of the accompanying trainer can include but is not limited to the role of the accompanying trainer, the image of the accompanying trainer, the theme of the accompanying trainer, and the background of the accompanying trainer. The large model can construct a global role image of the accompanying trainer based on the role information of the accompanying trainer, so that the large model explicitly plays the role of the accompanying trainer, so that the positioning of the large model for the role of the accompanying trainer is more accurate and more targeted, and the training effect is improved.

[0036] The rehearsal dialogue control information can refer to a related control mechanism capable of controlling the entire rehearsal process. The rehearsal dialogue control information can include but is not limited to the dialogue target, the upper limit of the rehearsal dialogue time, and the upper limit of the rehearsal dialogue round. For example, taking the dialogue scene between a pharmaceutical representative and a doctor as an example, the dialogue target can be that the pharmaceutical representative promotes a new drug to the doctor. The rehearsal dialogue control information can be used as a control basis for the rehearsal dialogue time, and the rehearsal dialogue control information can also be used as a basis for detecting whether the dialogue of the accompanying trainer deviates from the target.

[0037] The rehearsal dialogue constraint information can refer to a related constraint mechanism capable of constraining the accompanying trainer played by the large model. The rehearsal dialogue constraint information can include accompanying trainer constraint conditions, which can be one or more. The rehearsal dialogue constraint information can include one or more accompanying trainer constraint conditions at different stages, and each stage of the accompanying trainer constraint condition can include but is not limited to the task trigger condition, the task passing condition, and the role requirement of the accompanying trainer played by the large model.

[0038] Based on the accompanying trainer constraint condition at each stage, an execution framework of the entire rehearsal at different stages can be constructed. For example, taking the dialogue scene between a pharmaceutical representative and a doctor as an example, assuming that the accompanying trainer needs to perform the dialogue according to three different stages in this rehearsal, the rehearsal dialogue constraint information at the three different stages can include three different stages of rehearsal dialogue tasks. The task of the first stage can be that "the pharmaceutical representative introduces a new drug to you. Your current role state is indifferent, but you can be persuaded, and you will continue to further understand only when the pharmaceutical representative mentions a certain discount"; the task of the second stage can be that "you propose that the old drug can be fully reimbursed, but the new drug cannot. Your current role state is that you cannot accept the new drug, but you can be persuaded, and you will continue to further understand only when the pharmaceutical representative explains that the new drug has a certain special effect"; and the task of the third stage can be that "you propose how to persuade the patient to accept self-financing. Your current role state is that you can be persuaded, but you will not disclose the conditions for persuasion. Only when the pharmaceutical representative explains to you that the new drug can achieve the effect of the old drug and prevent damage to a certain drug, and the overall cost is more cost-effective". It can be seen that based on the rehearsal dialogue constraint information, an execution framework composed of three different stages can be used as a stage-by-stage role constraint basis for the large model playing the role of the doctor.

[0039] Step S200: repeatedly performing the simulation rehearsal step, outputting corresponding reply information based on the large model according to the rehearsal configuration information based on the dialogue information input by the rehearser, until the simulation dialogue with the rehearser is completed.

[0040] The large model can play the role of the rehearsal partner according to the rehearsal configuration information and generate reply information. In this embodiment, the large model can be a general large model or a fine-tuned model, but is not limited thereto, as long as it can complete role playing and answer the information input by the rehearser according to the rehearsal configuration information. The simulation rehearsal step is repeated according to the information input by the rehearser to output reply information to the rehearser until the simulation dialogue is completed.

[0041] The simulation rehearsal step can include the following steps S210 to S240.

[0042] Step S210: obtaining single dialogue information input by the rehearser in the current round.

[0043] The simulation dialogue with the rehearser can include multiple dialogue rounds, and each time corresponding single reply information is given to the single dialogue information input by the rehearser in the current round.

[0044] Step S220: determining the current task stage by analyzing the single dialogue information through the large model according to the rehearsal dialogue constraint information.

[0045] Step S230: generating stage-by-stage rehearsal partner role constraint information according to the current task stage and the rehearsal dialogue constraint information.

[0046] The large model can detect the dialogue content of the rehearser according to the stage-by-stage task in the rehearsal dialogue constraint information, determine which task stage the current dialogue round is in by judging which stage task has been completed, and generate corresponding stage-by-stage role constraint information. Whether the task in a certain stage in the rehearsal dialogue constraint information is completed is determined by the task passing condition in the stage. In this embodiment, the large model can be a general large model or a fine-tuned model, but is not limited thereto, as long as it can complete the detection and generation of stage-by-stage role constraint information.

[0047] The stage-by-stage role constraint information can refer to information that can be used to constrain the large model to rehearse according to the execution framework of the current stage. Taking the simulation rehearsal scene of a pharmaceutical representative selling drugs to a doctor as an example, the stage-by-stage role constraint information is described. The rehearser (i.e., the user) plays the role of a pharmaceutical representative, and the rehearsal partner (i.e., the large model) plays the role of a doctor.

[0048] When the task of the first stage is "a medical representative introduces a new drug to you. Your current role state is to be indifferent, but you can be persuaded to continue further in-depth understanding only when the medical representative mentions a certain discount", the stage role constraint information of the first stage can be: the condition for being persuaded is a certain discount, but the condition for being persuaded cannot be disclosed actively.

[0049] When the task of the second stage is "you propose that the old drug can be fully reimbursed, but the new drug cannot. Your current role state is that you cannot accept the new drug, but you can be persuaded to continue further in-depth understanding only when the medical representative mentions that the new drug has a certain special effect", the stage role constraint information of the second stage can be: do not in-depth understanding of the new drug before the medical representative mentions that the new drug has a certain special effect.

[0050] When the task of the second stage is "you propose how to persuade the patient to accept the self-financing problem. Your current role state is that you can be persuaded, but you will not disclose the condition for being persuaded actively. Only when the medical representative explains to you that the new drug can achieve the effect of the old drug, can prevent the damage of a certain drug, and the overall cost is actually more affordable", the stage role constraint information of the third stage can be: the condition for being persuaded is that the new drug can achieve the effect of the old drug, can prevent the damage of a certain drug, and the overall cost is actually more affordable, but the condition for being persuaded cannot be disclosed actively.

[0051] After receiving the single dialogue information input by the trainer in the current round, the training dialogue constraint information and the current dialogue content of the trainer are input into the large model for detection and processing, and the large model generates corresponding detection results and stage role constraint information. The large model can determine whether the dialogue content of the trainer has completed the task of the stage in which the previous dialogue is located according to the training dialogue constraint information. If not, the role constraint information of the current stage is generated according to the training dialogue constraint information. If it has been completed, the role constraint information of the next stage is generated according to the training dialogue constraint information.

[0052] Based on the stage role constraint information, the large model can be forced to follow the set execution framework during the training process, rather than responding randomly as the trainer played by the large model, greatly improving the effectiveness of the training. In addition, through the stage constraint information, the role of the trainer can be limited to the current stage, so that the large model can obtain more accurate role positioning information. Even in the case of multiple rounds of dialogue, the large model can still maintain accurate performance, and will not appear due to too much information and insufficient attention, resulting in poor training effect. The quality of the trainer is greatly improved, thereby improving the training effect.

[0053] Step S240: outputting single reply information based on the large model according to the historical dialogue record, the stage-by-stage trainer role constraint information, and the rehearsal configuration information.

[0054] The historical dialogue record, the stage-by-stage trainer role constraint information, and the rehearsal configuration information are obtained, and the large model is simulated based on the obtained information. The large model can clearly know all global information about the trainer to be played, such as the background of the trainer to be played and the current training theme, according to the preset trainer role information, so that the large model can perform the trainer according to the role setting.

[0055] The historical dialogue record can refer to the dialogue situation in the previous round of the current simulated dialogue, and can include single dialogue information input by the trainer in the previous round and single reply information replied by the large model playing the trainer in the previous round. Based on the historical dialogue record, the large model can take the historical dialogue record in the current rehearsal dialogue as context information, so that the large model can accurately understand the progress and content of the current round of dialogue without additional memory of the dialogue progress and content, thereby forming a single round of dialogue.

[0056] According to the stage-by-stage trainer role constraint information, the large model can be ensured to perform the role playing in stages according to the set execution framework. Further, the trainer role played by the large model can be further constrained to a more accurate stage, so as to perform more accurate playing to generate more accurate replies.

[0057] The simulation and rehearsal method based on the large model provided in the present application can realize personalized rehearsal through personalized role configuration, play different roles in different scenarios, and thus can perform targeted rehearsal. Through the preset stage-by-stage task, the controllability of role playing can be enhanced, so that the large model can play according to the preset stage-by-stage task, that is, perform training according to the execution framework, and the effect of training is improved. Through the stage-by-stage role constraint information in the preset stage-by-stage task, the accuracy of role playing is enhanced, so that the large model can more accurately play the trainer, and the quality of the trainer is improved. In addition, by limiting the dialogue between the trainer and the trainer played by the large model to a single round, the possibility of the large model deviating from the theme due to multiple rounds of dialogue can be reduced, and the stability of the rehearsal is improved.

[0058] Figure 3 For the flowchart of the simulation and rehearsal method based on the large model in another embodiment of the present application, in one embodiment, after obtaining the single dialogue information input by the trainer in the current round, the simulation and rehearsal method based on the large model can further include the following step S250.

[0059] Step S250: analyzing whether the single dialogue information has an abnormal situation through the large model according to the trainer role information and / or the rehearsal dialogue control information.

[0060] After obtaining the single dialogue information input by the current round of the trainer, the content of the single dialogue information input by the trainer can be analyzed for abnormality according to the trainer role information and / or the dialogue control information of the rehearsal. Further, corresponding correction prompts or replies can be generated according to the abnormality to feedback and correct the trainer or to alert and remind. In the embodiment, the large model can be a general large model or a fine-tuned model, but is not limited thereto, as long as it can complete abnormality detection and generate corresponding correction or reply information.

[0061] In one of the embodiments, the dialogue control information of the rehearsal can include an upper limit of the dialogue time of the rehearsal and an upper limit of the dialogue round of the rehearsal. The abnormality can include dialogue time abnormality, that is, when the dialogue time between the trainer and the partner is too long and / or the dialogue round is too many, it is determined that the dialogue of the rehearsal has dialogue time abnormality.

[0062] Figure 4 For the flowchart of determining the abnormality in one of the embodiments of the present application, whether the single dialogue information is abnormal can include the following steps S251 to S255 according to the trainer role information and / or the dialogue control information of the rehearsal through the large model.

[0063] Step S251: comparing the dialogue round in the current simulation rehearsal with the upper limit of the dialogue round of the rehearsal, and / or comparing the dialogue time in the current simulation rehearsal with the upper limit of the dialogue time of the rehearsal.

[0064] Step S253: when the comparison result is that the dialogue round in the current simulation rehearsal is greater than the upper limit of the dialogue round of the rehearsal, and / or the dialogue time in the current simulation rehearsal is greater than the upper limit of the dialogue time of the rehearsal, it is determined that the dialogue time is abnormal and corresponding prompt feedback information is generated.

[0065] Step S255: when the comparison result is that the dialogue round in the current simulation rehearsal is less than or equal to the upper limit of the dialogue round of the rehearsal, and the dialogue time in the current simulation rehearsal is less than or equal to the upper limit of the dialogue time of the rehearsal, it is determined that the dialogue time is normal.

[0066] Based on the large model, the dialogue control mechanism of the rehearsal is constructed, and whether the dialogue time is too long and / or the dialogue times is too many is determined by comparing the dialogue round in the current simulation rehearsal with the upper limit of the dialogue round of the rehearsal, and / or comparing the dialogue time in the current simulation rehearsal with the upper limit of the dialogue time of the rehearsal. When the dialogue round in the current simulation rehearsal is greater than the upper limit of the dialogue round of the rehearsal, and / or the dialogue time in the current simulation rehearsal is greater than the upper limit of the dialogue time of the rehearsal, it is determined that the dialogue of the current rehearsal has dialogue time abnormality and corresponding prompt feedback information is generated. For example, the prompt feedback information can be "the dialogue time has reached the upper limit" or "the dialogue times has reached the upper limit".

[0067] When the comparison result is that the number of dialogue turns in the current simulation is less than or equal to the upper limit of the number of dialogue turns in the simulation, and the dialogue duration in the current simulation is less than or equal to the upper limit of the dialogue duration in the simulation, it can be determined that the current simulation dialogue is normal in terms of time and the number of dialogues, and other processing or analysis operations can be continued according to the single dialogue information input by the current turn of the trainer.

[0068] Figure 5 For the flowchart of judging abnormal situations in another embodiment of the present application, in one embodiment, the abnormal situation can include inappropriate language. The inappropriate language can refer to the existence of malicious content, such as profanity, in the content input by the trainer. According to the trainer role information and / or the simulation dialogue control information, whether the single dialogue information contains an abnormal situation can be determined by the large model as follows: steps S257 to S261.

[0069] Step S257: Perform semantic detection on the single dialogue information using a large language model to determine whether inappropriate language exists in the single dialogue information.

[0070] Step S259: When inappropriate language exists in the single dialogue information, generate corresponding prompt feedback information.

[0071] Step S261: When inappropriate language does not exist in the single dialogue information, determine that the language of the single dialogue information is normal.

[0072] The trained large language model is used to determine whether inappropriate language exists in the content input by the trainer, for example, to determine whether content that attempts to attack the trainer played by the large model, such as profanity, exists in the content input by the trainer. When inappropriate language exists in the single dialogue information, the large model can generate corresponding prompt feedback information. For example, when profanity is detected, the prompt feedback information can be "I hope to have a harmonious conversation with you, please continue to answer in a peaceful language".

[0073] Figure 6 For the flowchart of judging abnormal situations in another embodiment of the present application, in one embodiment, the abnormal situation can include malicious deception. For example, the content output by the trainer is content that attempts to manipulate the trainer played by the large model, such as content that exchanges the roles of the trainer played by the large model and the trainer; or the content output by the trainer is content that maliciously deviates from the topic. According to the trainer role information and / or the simulation dialogue control information, whether the single dialogue information contains an abnormal situation can be determined by the large model as follows: steps S263 to S267.

[0074] Step S263: According to the trainer role information and the simulation dialogue control information, determine whether the single dialogue information contains a malicious deception intent by an intent recognition model.

[0075] Step S265: When the single-turn dialogue information exists the malicious teasing intention, corresponding prompt feedback information is generated.

[0076] Step S267: When the single-turn dialogue information does not exist the malicious teasing intention, it is determined that the intention of the single-turn dialogue information is normal.

[0077] According to the role information of the trainer, the background of the current trainer, the theme of the current simulation, and other information about the trainer can be clearly known, so that the intention recognition model can accurately detect whether the content input by the trainer exists the intention of malicious teasing of the trainer played by the large model. When the single-turn dialogue information exists the malicious teasing intention, corresponding prompt feedback information is generated. For example, when the malicious teasing intention is detected, the prompt feedback information can be "the content input by you is irrelevant to the current simulation theme, please input again".

[0078] The simulation and simulation method based on the large model provided in the present application can eliminate the possibility of the trainer maliciously teasing the trainer played by the large model before replying to the trainer through the strong control mechanism of abnormal situation detection, thereby improving the effectiveness of the simulation.

[0079] Figure 7 For the flowchart of the simulation and simulation method based on the large model in another embodiment of the present application, in one of the embodiments, after outputting the single-turn reply information based on the large model according to the historical dialogue record, the stage-by-stage trainer role constraint information and the trainer role information, the simulation and simulation method based on the large model can further include the following steps S270 to S280.

[0080] Step S270: Save the single-turn dialogue information and the single-turn reply information of the current round.

[0081] Step S280: Combine the single-turn dialogue information and the single-turn reply information of the current round with the single-turn dialogue information and the single-turn reply information of the previous round to generate a new historical dialogue record.

[0082] Save the single-turn dialogue information input by the trainer in the current round and the single-turn reply information output by the trainer played by the large model, and combine the single-turn dialogue information input by the trainer in the previous round and the single-turn reply information output by the trainer played by the large model to generate a new historical dialogue record. When the trainer inputs new single-turn dialogue information in the next dialogue round, the large model can determine the context information based on the new historical dialogue record. That is, the large model does not need to additionally remember the dialogue progress and content, so that each dialogue with the trainer played by the large model becomes a single-turn dialogue.

[0083] Since large models usually do not have long-term memory like humans, they rely more on short-term context information to generate answers. If the conversation spans a long time or involves complex plots, the model may have difficulty maintaining consistent role performance. Therefore, in the embodiments of the present application, all conversation records between the actor and the practice partner are cached to form the historical conversation records of this practice. After receiving the new input conversation content of the actor, the large model is input with the historical conversation records integrated from all previous conversation records and the current conversation content of the actor, and the historical conversation records are used as context information to enable the large model to accurately understand the current conversation progress and conversation content, so as to better answer the current input sentence of the actor. By making every conversation between the actor and the practice partner played by the large model become a single round of conversation, the problem caused by multi-round conversation can be effectively solved, such as deviating from the role setting.

[0084] It should be understood that although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or steps or stages in other steps.

[0085] Based on the same inventive concept, the embodiments of the present application also provide a large model-based simulation practice system for implementing the large model-based simulation practice method described above. The implementation scheme for solving problems provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more large model-based simulation practice system embodiments provided below can refer to the limitations of the large model-based simulation practice method described above, which will not be repeated here.

[0086] Each module in the above large model-based simulation practice system can be realized by software, hardware and their combinations in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0087] Figure 8FIG. 1 is a schematic diagram of a large model-based simulation training system according to an embodiment of the present application. In an embodiment, the large model-based simulation training system can include a training configuration module 100, a training dialogue exception processing module 200, a training dialogue constraint module 300, and a training partner module 400.

[0088] The training configuration module 100 can be configured to obtain training configuration information. A user can set the training configuration information through the training configuration module 100. The training configuration information can include, but is not limited to, trainer role information, training partner role information, training dialogue control information, and training dialogue constraint information. The training configuration module 100 can receive and save the above training configuration information.

[0089] The training dialogue exception processing module 200 can be connected to the training configuration module 100. The training dialogue exception processing module 200 can be configured to obtain single dialogue information input by a trainer, trainer role information, and training dialogue control information. The training dialogue exception processing module 200 can also be configured to analyze whether the single dialogue information has an abnormal situation based on the trainer role information and / or the training dialogue control information through a large model.

[0090] In a possible implementation, the interaction between the training dialogue exception processing module 200 and the large model can be achieved using a prompt (a technology based on artificial intelligence instructions), but is not limited thereto. The training dialogue exception processing module 200 can obtain trainer role information and training dialogue control information from the training configuration module 100, and obtain single dialogue information input by a trainer and construct a training dialogue exception processing mechanism based on a large model.

[0091] The training dialogue constraint module 300 can be connected to the training configuration module 100 and the training dialogue exception processing module 200, respectively. The training dialogue constraint module 300 can obtain training dialogue constraint information from the training configuration module 100 and single dialogue information from the training dialogue exception processing module 200. The training dialogue constraint module 300 can be configured to determine a current task stage by analyzing the single dialogue information based on the training dialogue constraint information through a large model, and generate stage-based training partner role constraint information according to the current task stage and the training dialogue constraint information. The training dialogue constraint module 300 can be configured to analyze which stage the current dialogue round is in, and generate stage-based training partner role constraint information corresponding to the stage, and send the single dialogue information input by the trainer and the stage-based training partner role constraint information to the training partner module 400.

[0092] In a possible implementation, the rehearsal dialogue constraint module 300 can be constructed based on a large model, and the interaction of the rehearsal dialogue constraint module 300 with the large model can use the prompt mode, but is not limited thereto. The large model detects the single dialogue information input by the rehearser according to the phased task in the rehearsal dialogue constraint information, and generates corresponding phased role constraint information. After the rehearsal dialogue constraint module 300 receives the single dialogue information input by the rehearser, the rehearsal dialogue constraint information and the single dialogue information are input to the large model for detection and processing, and the large model generates corresponding detection results and phased role constraint information.

[0093] The large model determines whether the current dialogue information meets the task passing condition of a certain phase in the rehearsal dialogue constraint information, to determine whether the task of the phase is completed. If not, the phased role constraint information of the current phase is generated according to the rehearsal dialogue constraint information, and if yes, the phased role constraint information of the next phase is generated according to the rehearsal dialogue constraint information. The phased role constraint information can make the large model perform according to the set execution framework during the rehearsal, rather than randomly perform as the rehearser played by the large model, thereby greatly improving the effectiveness of the rehearsal. On the other hand, by using the phased constraint information, the rehearser role played by the large model is limited to the current phase, so that the large model can obtain more accurate role positioning information, and even in the case of multiple rounds of dialogue, the rehearser can still maintain accurate performance, and the rehearser effect caused by too much information and insufficient attention can be avoided. The quality of the rehearser is greatly improved, thereby improving the rehearsing effect.

[0094] The rehearser module 400 can be connected with the rehearsal configuration module 100 and the rehearsal dialogue constraint module 300 respectively. The rehearser module 400 can be configured to obtain the historical dialogue record, the rehearser role information, the single dialogue information, and the phased rehearser role constraint information. The rehearser module 400 can obtain the rehearser role information from the rehearsal configuration module 100, obtain the single dialogue information and the phased rehearser role constraint information from the rehearsal dialogue constraint module 300, and obtain the historical dialogue record from a storage unit configured to store the dialogue situation of the previous round. The rehearser module 400 can also be configured to output the single reply information based on the large model according to the historical dialogue record, the phased rehearser role constraint information, and the rehearser role information.

[0095] In a possible implementation, the simulation trainer module 400 can be constructed based on a large model. The interaction between the rehearsal constraint module 400 and the large model can use the prompt mode, but is not limited thereto. The large model can play the role of the trainer according to the role information of the trainer, and generate reply information. The trainer module 400 inputs the global trainer role information obtained from the rehearsal configuration module 100 to the large model, so that the trainer played by the large model has a global role positioning. The trainer module 400 inputs the stage role constraint information received from the rehearsal dialogue constraint module 300 to the large model, so that the trainer played by the large model has a more accurate stage role positioning. The trainer module 400 inputs the historical dialogue record of the current rehearsal to the large model, so that the large model can master the progress and content of the entire dialogue. Thus, the large model can generate an accurate reply according to the single dialogue information input by the rehearser. The trainer module 400 sends the reply information to the rehearser. The trainer played by the large model constantly updates the role played according to the stage role constraint information, so as to more accurately complete the role playing, further improve the quality of the trainer, and improve the rehearsal effect.

[0096] In one of the embodiments, the rehearsal dialogue exception handling module 200 can detect whether the rehearser has the intention to maliciously play with the trainer played by the large model in the dialogue process, and timely corrects the feedback to the rehearser. The rehearsal dialogue exception handling module 200 can identify the number of dialogue rounds and the dialogue time length in the current rehearsal simulation according to the rehearsal dialogue control information, and generate a corresponding prompt feedback to the rehearser when the number of dialogue rounds is too many or the dialogue time length is too long.

[0097] In one of the embodiments, the rehearsal dialogue exception handling module 200 can use semantic detection of the large language model to identify whether the single dialogue information input by the rehearser has inappropriate language such as dirty words, and generate a corresponding correction or warning prompt feedback to the rehearser when the single dialogue information has inappropriate language.

[0098] In one of the embodiments, the rehearsal dialogue exception handling module 200 can also input the trainer role information to the large model through the prompt, so that the large model can clearly know the background of the current rehearser, the theme of the current rehearsal, and other information about the rehearser, and accurately detect whether the rehearser has the intention to maliciously play with the trainer played by the large model through the intention recognition model and the semantic detection model. When the large model detects that the rehearser has malicious intention, it generates a corresponding correction prompt content feedback to the rehearser.

[0099] Figure 9For another embodiment of the structure schematic diagram of the large model-based simulation practice system of the present application, in one embodiment, the large model-based simulation practice system can further include a practice history dialogue record module 500. The practice history dialogue record module 500 can be connected with the practice partner module 400. The practice history dialogue record module 500 can be used to save the single dialogue information and single reply information of the current round, and generate new history dialogue records in combination with the single dialogue information and single reply information of the current round and the single dialogue information and single reply information of the previous round. The practice history dialogue record module 500 can save the input content of the current round of the practice partner and the reply content of the practice partner, and also save the input content of all the practice partners in the previous round of the current dialogue practice and the reply content of the practice partner, so as to generate the practice dialogue history records based on the input content of the practice partners in the current round and the previous round and the reply content of the practice partner. That is, in each dialogue, the practice partner module 400 can obtain the previous history dialogue records from the practice history dialogue record module 500.

[0100] By recording the practice dialogue history, the context information can be provided to the practice partner played by the large model, so that each dialogue between the practice partner played by the large model and the practice partner becomes a single round dialogue. By caching all the dialogue records of the practice partner and the practice partner, i.e., the dialogue history of the current practice, when receiving the new dialogue input content of the practice partner, all the previous dialogue records and the current dialogue content of the practice partner are input to the large model together, the history dialogue records are used as context information, the large model can accurately understand the current dialogue progress and dialogue content, and answer the current input question of the practice partner, so that each dialogue with the large model becomes a single round, thereby effectively solving the problem caused by multiple rounds of dialogue.

[0101] The above-mentioned various modules of the large model-based simulation practice system can be realized by software, hardware and combinations thereof in whole or in part. The above-mentioned various modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above-mentioned various modules.

[0102] In one embodiment, Figure 10 For the internal structure diagram of the computer device in one embodiment of the present application, a computer device is provided, which can be a server, and the internal structure diagram thereof can be as shown in Figure 10As shown in the figure. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is used to store data. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement a large model-based simulation exercise method.

[0103] Those skilled in the art can understand that, Figure 10 The skilled in the art can understand that,

[0104] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. The non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. The volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0105] In the description of the present specification, the description of the terms "some embodiments", "other embodiments", "ideal embodiments", etc. means that the specific features, structures, materials or characteristics described in combination with the embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above-mentioned terms does not necessarily refer to the same embodiment or example.

[0106] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above-mentioned embodiments are not described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present specification.

[0107] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A large model-based simulation training method, characterized in that, The method comprises: obtaining rehearsal configuration information; the rehearsal configuration information comprises rehearsal role information, accompanying rehearsal role information, rehearsal dialogue control information, and rehearsal dialogue constraint information; the rehearsal dialogue constraint information comprises accompanying rehearsal constraint conditions of multiple different stages, and each stage's accompanying rehearsal constraint condition comprises a task trigger condition, a task passing condition, and a role requirement for an accompanying rehearsal role played by a large model; an execution framework of different stages in the rehearsal is constructed based on the accompanying rehearsal constraint condition of each stage; repeating the simulated rehearsal step to output corresponding reply information based on the large model according to the rehearsal configuration information until the simulated dialogue with the rehearsal is completed; the simulated rehearsal step comprises: obtaining single dialogue information input by the rehearsal in the current round; analyzing, by the large model, whether the single dialogue information has an abnormal situation according to the rehearsal role information and / or the rehearsal dialogue control information; determining a current task stage by analyzing the single dialogue information by the large model according to the rehearsal dialogue constraint information; generating stage accompanying rehearsal role constraint information according to the current task stage and the rehearsal dialogue constraint information, wherein the stage accompanying rehearsal role constraint information is information for constraining the large model to rehearse according to the execution framework of the current stage; outputting single reply information based on the large model according to the historical dialogue record, the stage accompanying rehearsal role constraint information, and the rehearsal configuration information.

2. The large model-based simulation rehearsal method of claim 1, wherein, The rehearsal dialogue control information comprises a rehearsal dialogue time upper limit and a rehearsal dialogue round upper limit, and the abnormal situation comprises a dialogue time abnormality; the step of analyzing, by the large model, whether the single dialogue information has an abnormal situation according to the rehearsal role information and / or the rehearsal dialogue control information comprises: comparing the dialogue round in the current simulated rehearsal with the rehearsal dialogue round upper limit, and / or comparing the dialogue time in the current simulated rehearsal with the rehearsal dialogue time upper limit; when the comparison result is that the dialogue round in the current simulated rehearsal is greater than the rehearsal dialogue round upper limit, and / or the dialogue time in the current simulated rehearsal is greater than the rehearsal dialogue time upper limit, judging that the dialogue time is abnormal and generating corresponding prompt feedback information; when the comparison result is that the dialogue round in the current simulated rehearsal is less than or equal to the rehearsal dialogue round upper limit, and / or the dialogue time in the current simulated rehearsal is less than or equal to the rehearsal dialogue time upper limit, judging that the dialogue time is normal.

3. The large model-based simulation rehearsal method of claim 1, wherein, The abnormal situation comprises inappropriate language, and the step of analyzing, by the large model, whether the single dialogue information has an abnormal situation according to the rehearsal role information and / or the rehearsal dialogue control information comprises: detecting the semantic of the single dialogue information by using a large language model to determine whether there is inappropriate language in the single dialogue information; when there is inappropriate language in the single dialogue information, generating corresponding prompt feedback information; when there is no inappropriate language in the single dialogue information, determining that the language of the single dialogue information is normal.

4. The large model-based simulation rehearsal method of claim 1, wherein, The abnormal situation includes malicious deception, and the analyzing, by the large model, whether the single dialogue information has an abnormal situation according to the role information of the trainer and / or the dialogue control information of the rehearsal includes: According to the role information of the trainer and the dialogue control information of the rehearsal, determining whether the single dialogue information has a malicious deception intention through an intention recognition model; When the single dialogue information has a malicious deception intention, generating corresponding prompt feedback information; When the single dialogue information does not have a malicious deception intention, determining that the intention of the single dialogue information is normal.

5. The large model-based simulation rehearsal method of claim 1, wherein, After outputting the single reply information based on the large model according to the historical dialogue record, the stage trainer role constraint information and the trainer role information, the method further includes: Saving the single dialogue information and the single reply information of the current round; Combining the single dialogue information and the single reply information of the current round with the single dialogue information and the single reply information of previous rounds to generate new historical dialogue records.

6. A large model-based simulation training system, characterized by, It includes: A rehearsal configuration module is configured to obtain rehearsal configuration information; the rehearsal configuration information includes trainer role information, trainer role information, rehearsal dialogue control information, rehearsal dialogue constraint information; the rehearsal dialogue constraint information includes trainer constraint conditions of multiple stages, and each stage of the trainer constraint condition includes a task trigger condition, a task passing condition, a role requirement of a trainer played by a large model, and an execution framework of different stages in the rehearsal is constructed based on the trainer constraint condition of each stage; A rehearsal dialogue exception processing module is connected with the rehearsal configuration module and is configured to obtain single dialogue information input by the trainer, the trainer role information and the rehearsal dialogue control information, and is further configured to analyze whether the single dialogue information has an abnormal situation according to the trainer role information and / or the rehearsal dialogue control information through the large model; A rehearsal dialogue constraint module is connected with the rehearsal configuration module and the rehearsal dialogue exception processing module, and is configured to obtain the single dialogue information and the rehearsal dialogue constraint information, and is further configured to analyze the single dialogue information to determine the current task stage according to the rehearsal dialogue constraint information through the large model, generate stage trainer role constraint information according to the current task stage and the rehearsal dialogue constraint information, and the stage trainer role constraint information is information for constraining the large model to rehearse according to the execution framework of the current stage; A trainer module is connected with the rehearsal configuration module and the rehearsal dialogue constraint module, and is configured to obtain historical dialogue records, trainer role information, single dialogue information and stage trainer role constraint information, and is further configured to output single reply information based on the large model according to the historical dialogue records, the stage trainer role constraint information and the trainer role information.

7. The large model-based simulation rehearsal system of claim 6, wherein, The large model-based simulation rehearsal system further includes: The drill history dialogue record module is connected with the accompanying drill module, and is used for saving the single dialogue information and the single reply information of the current round, and generating new history dialogue records by combining the single dialogue information and the single reply information of the current round with the single dialogue information and the single reply information of previous rounds. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The computer program is executed by the processor to implement the steps of the large model-based simulation drill method in any one of claims 1 to 5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the large model-based simulation drill method in any one of claims 1 to 5.

Images