An intelligent interaction system applied to AI training

Abstract

The application relates to the technical field of AI interaction, in particular to an intelligent interaction system applied to AI training, which comprises a storage module used for storing model training data, a learning module used for training an AI model based on the stored model training data, an interaction module used for providing an interaction platform for users and the AI model, a recording module used for recording the interaction content and the interaction times of a plurality of users with the AI model and storing the interaction content as model training data, an auxiliary module used for periodically dividing the interactions of the users and determining the interaction times and the corresponding interaction content of the users in the corresponding period, a response module used for judging whether the training of the AI model meets the standard and judging the reason why the standard is not met based on the response activity, and a control module used for modifying the corresponding parameters. The application effectively realizes real-time monitoring and accurate adjustment of the training of the AI model according to the interaction conditions of the users, and effectively improves the training efficiency of the AI model.

Description

Technical Field

[0001] This invention relates to the field of AI interaction technology, and in particular to an intelligent interaction system for AI training. Background Technology

[0002] AI training refers to the process of inputting data into an AI model, allowing the model to learn data features and patterns, and adjusting its internal parameters based on the learning results, thereby gradually improving its ability to complete specific tasks. An AI model is a computer program or data framework that uses algorithms and data to simulate human intelligent behavior. Its core is to use mathematical logic and data patterns to achieve the ability to process specific tasks. Its essence is to use a set of rules to enable computers to autonomously complete complex tasks without programming them one by one. There are many types of AI models. Chat interaction models are technical frameworks used to realize natural language interaction between humans and computers. They can understand users' text information, analyze their intentions and needs, and generate responses that are context-appropriate and logically coherent. Intelligent interaction systems applied to AI training are tools that assist AI models in the training process and are of great significance to improving the interactive capabilities of AI models. With the continuous improvement of living standards, research on chat interaction models has important research significance for the improvement of human living standards.

[0003] Chinese Patent Publication No. CN116774817A discloses an AI interactive control method and system for making AI-matched gestures and postures conform to the speaker's posture and unique speaking rhythm, resulting in greater diversity and naturalness, thus improving the interactive experience during human-computer interaction. The method includes: acquiring a dialogue video stream of the speaker; inputting the dialogue video stream into a target feature extraction model for feature extraction to determine the speaker's text information, rhythm information, identity information, and body posture information; determining AI feedback information based on the text information, and determining a target gesture group from a feature mapping database based on the AI ​​feedback information, rhythm information, identity information, and body posture information; inputting the target gesture group into a style fusion model for feature fusion to generate interactive gestures that conform to the speaker's posture and dialogue style, wherein the style fusion model is trained using several gesture samples; and displaying the AI ​​feedback information and interactive gestures to the speaker.

[0004] Therefore, the above solution determines AI feedback information through the dialogue video stream, identifies target gesture groups that match the speaker's style, generates interactive gestures, and displays the AI ​​feedback information and interactive gestures to the speaker. It matches gesture postures based on the speaker's text, rhythm, identity, and body language information, ensuring that the AI-matched gestures match the speaker's posture and unique speaking rhythm. However, this solution cannot monitor and precisely adjust the AI ​​model's training in real time based on user interaction, thus failing to guarantee the training efficiency of the AI ​​model. Summary of the Invention

[0005] To address this issue, the present invention provides an intelligent interactive system for AI training, which overcomes the problem in the prior art that the training of AI models cannot be monitored and precisely adjusted in real time according to the user's interaction, resulting in low training efficiency of AI models.

[0006] To achieve the above objectives, the present invention provides an intelligent interactive system for AI training, comprising:

[0007] The storage module is used to store the model training data required for training AI models;

[0008] A learning module, which is connected to the storage module, is used to train the AI ​​model based on the stored model training data;

[0009] The interaction module provides an interaction platform for users and AI models.

[0010] The recording module is connected to the storage module and the interaction module respectively, and is used to record the interaction content between several users and the AI ​​model, and to record the number of interactions of each user; the recording module is also used to store the interaction content and mark it as the model training data;

[0011] An auxiliary module, connected to the recording module, is used to periodically divide the interactions of each user and determine the number of interactions and the corresponding interaction content of each user within the corresponding period after the division.

[0012] A response module, which is connected to the auxiliary module, is used to determine whether the training of the AI ​​model meets the standard based on the response activity, and, if the training of the AI ​​model is determined to be non-compliant, to determine the reason why the training of the AI ​​model is non-compliant based on the interaction similarity.

[0013] The response activity is the ratio of the number of actual registered users whose interaction frequency is greater than the second interaction frequency within a single division period to the total number of actual registered users.

[0014] The interaction similarity is the average of the actual interaction similarities of each actual registered user, and the actual interaction similarity of a single actual registered user is the ratio of the number of similar interaction features appearing in the user's interaction content to the number of interactions of the actual registered user.

[0015] The control module, which is connected to the storage module, the learning module and the response module respectively, is used to correct the total number of training data of the model based on the interaction similarity difference, reduce the context weight parameter based on the word ratio, correct the cosine similarity based on the interaction similarity ratio, correct the detail weight based on the query ratio, and increase the maximum length based on the difficulty of the query words.

[0016] Wherein, the interaction similarity difference is the difference between the interaction similarity and the preset interaction similarity;

[0017] The word ratio is the ratio of the number of identical interactive words to the number of interactions.

[0018] The interaction similarity ratio is the ratio of the preset interaction similarity to the interaction similarity.

[0019] The cosine similarity is a semantic analysis parameter used to determine whether there is a relationship between interactive content;

[0020] The query ratio is the ratio of the total number of query terms in the interaction content of each user to the number of interactions.

[0021] Furthermore, the response module is also used to determine whether a user's registration meets the standard based on the interaction frequency, and to remove the family member registered user from the total number of registered users if the registered user is determined to be a family member registered user, or to retain the corresponding registered user in the total number of registered users if the registered user is determined to be an active registered user.

[0022] Furthermore, the response module is also used to determine whether the training of the AI ​​model meets the standard based on the response activity after the removal of the family registration users and the retention of the active registration users, and to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity according to the determination result.

[0023] Furthermore, the response module is also used to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity, and, if the reason is that the training accuracy of the AI ​​model does not meet the standard, to adjust the total number of training data of the model based on the interaction similarity difference, or, if the reason is that the comprehensiveness of the AI ​​model's response content does not meet the standard, to adjust the detail weight based on the query ratio.

[0024] Furthermore, the control module is also used to increase the total number of model training data based on the interaction similarity difference, and the increase in the total number of model training data is proportional to the interaction similarity difference.

[0025] Furthermore, the control module is also used to reduce the context weight parameter based on the word ratio after the total number of training data for the model has increased, and the reduction of the context weight parameter is proportional to the word ratio.

[0026] Furthermore, the response module is also used to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity after the context weight parameter is reduced, and to correct the cosine similarity based on the interaction similarity ratio according to the determination result, or to correct the detail weight based on the query ratio.

[0027] Furthermore, the control module is also used to increase the cosine similarity based on the interaction similarity ratio, and the increase in cosine similarity is inversely proportional to the interaction similarity ratio.

[0028] Furthermore, the control module is also used to increase the detail weight based on the query ratio, and the increase in the detail weight is proportional to the query ratio.

[0029] Furthermore, the control module is also used to increase the maximum length based on the difficulty of the query words after the detail weight is increased, and the increase in the maximum length is proportional to the difficulty of the query words.

[0030] Compared with existing technologies, the beneficial effects of this invention are as follows: By setting up a response module and a control module, the response module determines whether a user's registration meets the standards based on the interaction frequency. After removing family registration users and retaining active registration users, it determines whether the training of the AI ​​model meets the standards based on the response activity. This allows for a quick and accurate determination of whether the AI ​​model's training meets the standards. If the AI ​​model's training does not meet the standards, the reason for this is determined based on the interaction similarity. This effectively achieves real-time monitoring of the AI ​​model's training based on user interaction. The control module corrects the corresponding parameters, effectively improving the training efficiency of the AI ​​model while precisely adjusting the AI ​​model's training based on user interaction.

[0031] Furthermore, the response module of this invention determines whether a user's registration meets the standards based on the interaction frequency. Effectively determining whether a registered user is a family member or an active user is beneficial for determining whether the subsequent training of the AI ​​model meets the standards. This further improves the training efficiency of the AI ​​model while enabling real-time monitoring of the AI ​​model training based on user interaction.

[0032] Furthermore, the response module of this invention is also used to determine whether the training of the AI ​​model meets the standard based on the response activity after removing family-registered users and retaining active registered users. It can timely and accurately determine whether the training of the AI ​​model meets the standard, avoiding misjudgment. While further realizing real-time monitoring of the training of the AI ​​model based on user interaction, it further improves the training efficiency of the AI ​​model.

[0033] Furthermore, the response module of this invention is also used to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity, which effectively avoids misjudgment between the total number of training data to be corrected and the detailed weights to be corrected. While further realizing real-time monitoring of the training of the AI ​​model based on the user's interaction, it further improves the training efficiency of the AI ​​model.

[0034] Furthermore, the control module of this invention is also used to increase the total number of model training data based on the interaction similarity difference, which effectively avoids the situation where the training of the AI ​​model does not meet the standard due to the total number of model training data not meeting the standard. While further realizing the precise adjustment of the training of the AI ​​model according to the user's interaction, it further improves the training efficiency of the AI ​​model.

[0035] Furthermore, the control module of this invention is also used to reduce the context weight parameter based on the word ratio after the total number of model training data has increased. This effectively reduces the context weight parameter and maintains the synergy between the context weight parameter and the user's interaction. This further improves the training efficiency of the AI ​​model while enabling precise adjustment of the AI ​​model training based on the user's interaction.

[0036] Furthermore, the response module of this invention is also used to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity after the context weight parameter is reduced. It can further determine the reason for non-compliance, avoiding misjudgment between the need to correct cosine similarity and the need to correct detail weight. While further realizing real-time monitoring of the training of the AI ​​model based on the user's interaction, it further improves the training efficiency of the AI ​​model.

[0037] Furthermore, the control module of this invention is also used to increase the cosine similarity based on the interaction similarity ratio, which effectively avoids the situation where the training of the AI ​​model does not meet the standard due to the cosine similarity not meeting the standard. While further realizing the precise adjustment of the training of the AI ​​model according to the user's interaction, it further improves the training efficiency of the AI ​​model.

[0038] Furthermore, the control module of this invention is also used to increase the detail weights based on the query ratio. Increasing the detail weights can effectively avoid the situation where the training of the AI ​​model does not meet the standard due to the detail weights not meeting the standard. While further realizing the precise adjustment of the training of the AI ​​model according to the user's interaction, it further improves the training efficiency of the AI ​​model.

[0039] Furthermore, the control module of this invention is also used to increase the maximum length based on the difficulty of the query words after the detail weights are increased. This effectively maintains the coordination between the maximum length and the difficulty of the query words, and effectively avoids the situation where the training of the AI ​​model is judged to be non-compliant due to the maximum length not meeting the standard. This further realizes the precise adjustment of the training of the AI ​​model based on the user's interaction, while further improving the training efficiency of the AI ​​model. Attached Figure Description

[0040] Figure 1 This is a structural block diagram of an intelligent interactive system applied to AI training according to an embodiment of the present invention;

[0041] Figure 2 This is a flowchart illustrating the workflow of an intelligent interactive system applied to AI training according to an embodiment of the present invention.

[0042] Figure 3 This is a flowchart illustrating how to determine whether the training of an AI model conforms to the standard and the reasons why the training of an AI model does not conform to the standard, according to an embodiment of the present invention.

[0043] Figure 4 This is a flowchart illustrating the reasons why the training of an AI model does not meet the standards, as described in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0045] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0046] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0047] Please see Figure 1The diagram shown is a structural block diagram of an intelligent interactive system for AI training according to an embodiment of the present invention. The intelligent interactive system for AI training according to this embodiment includes a storage module, a learning module, an interaction module, a recording module, an auxiliary module, a response module, and a control module; wherein,

[0048] The storage module is used to store the model training data required for training the AI ​​model;

[0049] The learning module is connected to the storage module and is used to train the AI ​​model based on the stored model training data;

[0050] The interaction module is used to provide an interaction platform for users and AI models;

[0051] The recording module is connected to the storage module and the interaction module respectively, and is used to record the interaction content between several users and the AI ​​model, as well as the number of interactions of each user; the recording module is also used to store the interaction content and mark it as the model training data;

[0052] The auxiliary module is connected to the recording module and is used to periodically divide the interactions of each user and determine the number of interactions and the corresponding interaction content of each user in the corresponding period after the division.

[0053] The response module is connected to the auxiliary module and is used to determine whether the training of the AI ​​model meets the standard based on the response activity, and, if the training of the AI ​​model does not meet the standard, to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity.

[0054] The control module is connected to the storage module, the learning module and the response module respectively, and is used to correct the total number of training data of the model based on the interaction similarity difference, reduce the context weight parameter based on the word ratio, correct the cosine similarity based on the interaction similarity ratio, correct the detail weight based on the query ratio, and increase the maximum length based on the difficulty of the query words.

[0055] Specifically, when the auxiliary module periodically divides the interactions of each user, the division period is 72 hours. That is, the current day is selected as the division day, and the 36 hours before the division day and the division day are selected as a single division period. After the division day ends, the next division period begins, that is, the next 72 hours after the division day ends is the next division period.

[0056] Please see Figure 2As shown, it is a flowchart of the intelligent interactive system applied to AI training according to an embodiment of the present invention. When the intelligent interactive system for AI training described in this embodiment of the invention is running, the storage module stores the model training data required for training the AI ​​model, the learning module trains the AI ​​model based on the stored model training data, the interaction module provides the interaction platform for the user and the AI ​​model, the recording module records the interaction content between several users and the AI ​​model and records the number of interactions for each user, the recording module is also used to store the interaction content and mark it as model training data, the auxiliary module periodically divides the interactions of each user and determines the number of interactions and corresponding interaction content for each user in the corresponding period after the division, the response module determines whether the training of the AI ​​model meets the standard based on the response activity, the response module determines the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity if it determines that the training of the AI ​​model does not meet the standard, the control module corrects the total number of model training data based on the interaction similarity difference, reduces the context weight parameter based on the word ratio, or corrects the cosine similarity based on the interaction similarity ratio, the control module is also used to correct the detail weight based on the query ratio, and increases the maximum length based on the difficulty of the query words.

[0057] Please continue reading. Figure 3 The diagram shows a flowchart illustrating how an AI model's training conforms to standards and the reasons why training might not meet those standards, according to an embodiment of the present invention. The response module of this invention is also used to determine whether a user's registration conforms to standards based on the interaction frequency.

[0058] If the interaction frequency is less than or equal to the preset interaction frequency F set in the response module, the response module determines that the registered user is a family registration user and the user's registration does not meet the standard. The family registration user is removed from the total number of registered users. In this embodiment, the preset interaction frequency F = 1 time / 5 months.

[0059] If the interaction frequency is greater than the preset interaction frequency F, the response module determines that the registered user is an active registered user and the user's registration meets the standard, and the corresponding active registered user is still retained in the total number of registered users.

[0060] Specifically, some users helped users who lacked the ability to use the service to register accounts that showed no usage trend. These users were recorded as family registration users. Users whose interaction frequency was greater than the preset interaction frequency were recorded as accounts with usage trend. These users were recorded as active registration users.

[0061] The interaction frequency of a single registered user is the number of interactions made by the user using the intelligent interaction platform from the time of registration to the completion of the current division period, that is, the total number of interaction messages sent by the user after registration. The ratio of the number of interactions to the number of registration days is calculated and recorded as the interaction frequency. The number of registration days is the total number of days from the date of registration to the date of completion of the current division period. Based on the interaction frequency, it is possible to accurately determine whether a registered user is an active registered user or a family registered user.

[0062] The preset interaction frequency is set to 1 time / 5 months. If the interaction frequency is less than or equal to the preset interaction frequency, it indicates that the user does not tend to use the intelligent interaction platform after registration. In order not to affect the system's judgment accuracy for training the AI ​​model, such family registration users are removed from the total number of registered users. In addition, if the interaction frequency is greater than the preset interaction frequency, it indicates that the registered user tends to use the intelligent interaction platform after registration. The registered user is judged as an active registered user. 1 time / 5 months is a reasonable value in real-world applications.

[0063] The total number of registered users after removing the family-registered users and retaining the active registered users is marked as the actual total number of registered users.

[0064] Please see Figure 3 The diagram shows a flowchart illustrating how an AI model's training conforms to standards and the reasons why training might not meet those standards, according to an embodiment of the present invention. The response module described in this embodiment is used to determine whether the AI ​​model's training conforms to standards based on the response activity level after the removal of family-registered users and the retention of active registered users.

[0065] If the response activity is greater than the preset response activity T set in the response module, the response module determines that the training of the AI ​​model meets the standard, completes the determination of whether the training of the AI ​​model meets the standard for the corresponding division period, and determines whether the training of the AI ​​model in the next division period meets the standard. In this embodiment, the preset response activity T = 97.2%;

[0066] If the response activity is less than or equal to the preset response activity T, the response module determines that the training of the AI ​​model does not meet the standard, and determines the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity.

[0067] Specifically, within a single division period, the actual registered users whose interaction frequency is greater than the secondary interaction frequency G are identified and recorded as the number of interactive users. The ratio of the number of interactive users to the total number of actual registered users is calculated, and the obtained ratio is recorded as the response activity. When the secondary interaction frequency is 1 time / 36h, it can be used as the critical point between user activity and inactivity in real life. Therefore, in this embodiment of the invention, the secondary interaction frequency G = 1 time / 36h.

[0068] If the response activity level is greater than the preset response activity level, it indicates that the user's usage rate of the intelligent interaction platform meets the platform's expectations. The allowable error is 3 people when the total number of actual registered users is 100, and the allowable error is 2.8 million people when the total number of actual registered users is 100 million. Therefore, the preset response activity level is set to 97.2%. Furthermore, when it is determined that the training of the AI ​​model does not meet the standard, if the response activity level is less than or equal to the preset response activity level, it indicates that some users have not frequently used the intelligent interaction platform after registration. This is determined to be because the training of the AI ​​model does not meet the user's satisfaction. Therefore, it is determined that the training of the AI ​​model does not meet the standard within this division period, and the reason for the training of the AI ​​model not meeting the standard is determined based on the interaction similarity.

[0069] The interaction similarity is the average of the actual interaction similarities of each of the actual registered users. The actual interaction similarity is the average of the ratio of the number of similar interaction features appearing in the interaction content to the number of interactions. For a single actual registered user, the actual interaction similarity is the ratio of the number of similar interaction features appearing in the user's interaction content to the number of interactions of that actual registered user. Through interaction similarity, the number of times a user asks about the same thing and the different ways of asking about the same thing can be accurately analyzed. The reason why the training of the AI ​​model does not meet the standard can be determined through interaction similarity.

[0070] The number of similar interaction features refers to the number of identical interaction words appearing in the interaction content of the actual registered users and the number of interaction content with a cosine similarity greater than a preset cosine similarity E. Cosine similarity is determined by semantic analysis between interaction content. If the cosine similarity of the language vectors of two pieces of content is greater than the preset cosine similarity, it indicates that there is a relationship between the interaction content. Interaction content with a relationship is considered to have similar interaction features. In this embodiment of the invention, the preset cosine similarity E = 0.65.

[0071] Please continue reading. Figure 3 As shown, the response module of this invention is also used to determine the reason why the training of the AI ​​model does not meet the standard based on the interaction similarity:

[0072] If the interaction similarity is greater than the preset interaction similarity R set in the response module, the response module determines that the training of the AI ​​model does not meet the standard because the training accuracy of the AI ​​model does not meet the standard, and corrects the total number of training data of the model based on the interaction similarity difference. In this embodiment, the preset interaction similarity R = 38%, and the interaction similarity difference is the difference between the interaction similarity and the preset interaction similarity.

[0073] If the interaction similarity is less than or equal to the preset interaction similarity R, the response module determines that the AI ​​model training does not meet the standard because the comprehensiveness of the AI ​​model's response content does not meet the standard, and adjusts the detail weights based on the query ratio.

[0074] Specifically, if the ratio of the number of similar interaction features appearing in the interaction content of a single actual registered user to the number of interactions of that actual registered user is greater than 38%, it means that the number of similar interaction features appearing in 100 interaction contents is greater than 38. This indicates that the user's query repetition frequency does not meet the standard. The reason is that the accuracy of the response content of the trained AI model does not meet the standard. 38% is a value obtained through actual application, that is, the training accuracy of the AI ​​model does not meet the standard. If the interaction similarity is less than or equal to 38%, it means that the user's query repetition rate meets the standard. The interaction similarity at this time may be due to the user's query habits or the user's understanding of a certain word not meeting expectations. This indicates that the comprehensiveness of the response content of the trained AI model does not meet the standard.

[0075] When the accuracy of the response content of the trained AI model does not meet the standard, it is considered that the total amount of training data used to train the AI ​​model does not meet the standard, resulting in inaccurate response content. The total amount needs to be increased. The interaction similarity difference can effectively indicate the degree to which the interaction similarity is greater than the preset interaction similarity. The greater the degree to which it is greater than the preset interaction similarity, the greater the degree to which the accuracy of the AI ​​model's response content does not meet the standard. Therefore, the greater the increase in the total amount of training data, the more important it is to increase the detail weights.

[0076] The query ratio is the ratio of the total number of query words with meanings in the interaction content of each user, as determined by semantic analysis, to the total number of interactions. The query ratio can accurately indicate the frequency of the number of query words in the interaction content, and the detail weights can be adjusted based on the query ratio.

[0077] Please continue reading. Figure 3 As shown, the control module of the present invention is also used to increase the total number of model training data based on the interaction similarity difference:

[0078] If the interaction similarity difference is greater than the second preset interaction similarity difference △L2 set in the response module, the control module increases the total number of model training data to 1.87 times the initial total number of model training data. In this embodiment, the second preset interaction similarity difference △L2 = 49%.

[0079] If the interaction similarity difference is less than or equal to the second preset interaction similarity difference △L2 and greater than the first preset interaction similarity difference △L1 set in the response module, the control module increases the total number of model training data to 1.67 times the initial total number of model training data. In this embodiment, the first preset interaction similarity difference △L1 = 28%.

[0080] If the interaction similarity difference is less than or equal to the first preset interaction similarity difference △L1, the control module increases the total number of model training data to 1.29 times the initial total number of model training data;

[0081] Specifically, the preset interaction similarity is 38%, and the interaction similarity difference is the difference between the interaction similarity and the preset interaction similarity. Therefore, the value of the interaction similarity difference is ∈ [0, 62%]. Within this range, two values ​​that reasonably divide the interval are selected to represent the increasing trend of the interaction similarity difference. The two values ​​selected in practice are 28% and 49%. The initial total number of training samples for the medium-sized chat model of this invention is 0.8 billion. Considering the function and model of the model, the maximum total number of training samples is set to 1.5 billion to solve the problem of training accuracy not meeting the standard. Therefore, the multiplier of the increase in the total number of training samples is ∈ [1, 1.875]. Three values ​​are selected within the range of the multiplier of the increase in the total number of training samples to represent the increasing trend of the multiplier of the total number of training samples. The three selected values ​​are 1.29, 1.67 and 1.87. As mentioned in the above analysis, the increase in the total number of training samples is proportional to the interaction similarity difference.

[0082] Please continue reading. Figure 3 As shown, the control module of this invention is also used to reduce the context weight parameter based on the word ratio after the total number of model training data has been reduced:

[0083] If the word ratio is greater than the second preset word ratio H2 set in the response module, the control module reduces the context weight parameter to 0.58 times the initial context weight parameter, wherein, in this embodiment, the second preset word ratio H2 = 0.76;

[0084] If the word ratio is less than or equal to the second preset word ratio H2 and greater than the first preset word ratio H1 set in the response module, the control module reduces the context weight parameter to 0.76 times the initial context weight parameter, wherein, in this embodiment, the first preset word ratio H1 = 0.33;

[0085] If the word ratio is less than or equal to the first preset word ratio H1, the control module reduces the context weight parameter to 0.92 times the initial context weight parameter;

[0086] Specifically, when the training accuracy of the AI ​​model is determined to be unsatisfactory, the training data is increased. After the training data is increased, in order to avoid the AI ​​model's response content being devoid of substantive content, the AI ​​model's output response content must be forced to include the interactive keywords from the interactive content.

[0087] The word ratio is the ratio of the number of identical interactive words to the number of interactions. A larger word ratio indicates that more identical interactive words appear in the interactive content, and fewer interactive keywords are included in the response content. Therefore, it is necessary to reduce the context weight parameter to effectively improve the accuracy of the response content. The word ratio is ∈ [0, 1]. Two values ​​are selected within this range to divide the trend of word ratio changes. The two values ​​selected in practice are 0.33 and 0.76. The context weight parameter is set to 0.23 during use, and the minimum context weight parameter is set to 0.1. Therefore, the context weight parameter is ∈ [0.1, 0.23]. Thus, the reduction factor of the context weight parameter is ∈ [0.43, 1]. Three values ​​are selected within this range to represent the decreasing trend of the context weight parameter. The three selected values ​​are 0.58, 0.76, and 0.92.

[0088] Please see Figure 4 The flowchart shown illustrates the reasons why the AI ​​model's training does not meet the standards. The response module described in this embodiment of the invention is further used to determine the reasons why the AI ​​model's training does not meet the standards based on the interaction similarity after the context weight parameter is reduced.

[0089] If the interaction similarity is greater than the preset interaction similarity R set in the response module, the response module determines that the training of the AI ​​model does not meet the standard because the training accuracy of the AI ​​model does not meet the standard, and corrects the cosine similarity based on the interaction similarity ratio, wherein the interaction similarity ratio is the ratio of the preset interaction similarity to the interaction similarity.

[0090] If the interaction similarity is less than or equal to the preset interaction similarity R, the response module determines that the AI ​​model's training does not meet the standard because the comprehensiveness of the AI ​​model's response content does not meet the standard, and adjusts the detail weight based on the query ratio.

[0091] Specifically, after the context parameter is reduced, the interaction similarity is re-determined based on the user's language habits and semantic goals within the current segmentation period. If the interaction similarity is still greater than the preset interaction similarity, it indicates that the cosine similarity needs to be corrected to reasonably include the number of interaction contents with a cosine similarity greater than the preset cosine similarity E. If the interaction similarity is less than or equal to the preset interaction similarity, it indicates that the comprehensiveness of the AI ​​model's response content does not meet the standard and the detail weights need to be corrected.

[0092] Please continue reading. Figure 4 As shown, the control module in this embodiment of the invention is further used to increase the cosine similarity based on the interaction similarity ratio:

[0093] If the interaction similarity ratio is greater than the second preset interaction similarity ratio S2 set in the response module, the control module increases the cosine similarity to 1.21 times the initial cosine similarity, wherein, in this embodiment, the second preset interaction similarity ratio S2 = 79%;

[0094] If the interaction similarity ratio is less than or equal to the second preset interaction similarity ratio S2 and greater than the first preset interaction similarity ratio S1 set in the response module, the control module increases the cosine similarity to 1.34 times the initial cosine similarity, wherein, in this embodiment, the first preset interaction similarity ratio S1 = 56%;

[0095] If the interaction similarity ratio is less than or equal to the first preset interaction similarity ratio S1, the control module increases the cosine similarity to 1.48 times the initial cosine similarity.

[0096] Specifically, in this embodiment of the invention, the preset cosine similarity is 0.65, and the maximum cosine similarity is 1. Therefore, the value range of the cosine similarity increase factor is ∈ [1, 1.54]. Three values ​​are selected within this range to represent the trend of the cosine similarity increase factor. The three selected values ​​are 1.21, 1.34, and 1.48. The preset interaction similarity is 38%, and the value range of interaction similarity is ∈ (38%, 100%). The value range of the interaction similarity factor, i.e., the interaction similarity ratio, is ∈ (38%, 100%). Two values ​​are selected within this range to divide the interval. The two values ​​selected in practice are 56% and 79%. The larger the interaction similarity ratio, the smaller the interaction similarity and the smaller the trend of the cosine similarity increase factor.

[0097] Please continue reading. Figure 4 As shown, the control module in this embodiment of the invention is further configured to increase the detail weight based on the query ratio:

[0098] If the query ratio is greater than the second preset query ratio X2 set in the response module, the control module determines to increase the detail weight to 1.72 times the initial detail weight, wherein, in this embodiment, the second preset query ratio X2 = 69%;

[0099] If the query ratio is less than or equal to the second preset query ratio X2 and greater than the first preset query ratio X1 set in the response module, the control module determines to increase the detail weight to 1.51 times the initial detail weight, wherein, in this embodiment, the first preset query ratio X1 = 43%;

[0100] If the query ratio is less than or equal to the first preset query ratio X1, the control module determines to increase the detail weight to 1.25 times the initial detail weight;

[0101] Specifically, the query ratio ranges from [0, 100%]. However, based on actual results, the query ratio range is concentrated in the middle of the range. Therefore, two values ​​are selected within this range to divide the range. The two selected values ​​are 43% and 69%. The initial detail weight is 0.52, and the maximum detail weight is 0.9. Therefore, the range of the detail weight multiplier increase is from [1, 1.73]. Three values ​​are selected within this range to represent the increasing trend of the detail weight multiplier. The three selected values ​​are 1.25, 1.51, and 1.72. The larger the query ratio, the more total query terms there are, and the greater the increase in the detail weight multiplier.

[0102] Please continue reading. Figure 4 As shown, the control module in this embodiment of the invention is further configured to increase the maximum length based on the difficulty of the query words after the detail weights have been increased:

[0103] If the query term is a simple query term, the control module determines to increase the maximum length to 1.31 times the initial maximum length;

[0104] If the query term is a medium-length query term, the control module determines to increase the maximum length to 1.55 times the initial maximum length;

[0105] If the query word is a difficult query word, the control module determines to increase the maximum length to 1.79 times the initial maximum length;

[0106] Specifically, the initial maximum length is 8K tokens, and the maximum value of the maximum length is set to 15K tokens to meet the training requirements of the AI ​​model in this embodiment of the invention. The range of the maximum length increase factor is ∈ [1, 1.875]. Three values ​​are selected from this range to represent the increase factor of the maximum length. The three selected values ​​are 1.31, 1.55 and 1.79. The difficulty level of the query words can be divided into three levels according to the number of core elements contained in their concepts. The core elements are less than or equal to 3 and are set as easy query words. The core elements are greater than 3 and less than 5 and are set as medium query words. The core elements are greater than or equal to 5 and are set as difficult query words. The higher the difficulty level of the query word, the greater the increase factor of the maximum length.

[0107] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent interaction system applied to AI training, characterized in that, The application relates to an AI model training system, which comprises the following modules: a storage module for storing model training data required for training an AI model; a learning module connected with the storage module for training an AI model based on the stored model training data; an interaction module for providing an interaction platform for users and the AI model; a recording module connected with the storage module and the interaction module respectively for recording interaction content of a plurality of users with the AI model and recording the interaction times of the users; the recording module is also used for storing the interaction content and marking the interaction content as the model training data; an auxiliary module connected with the recording module for periodically dividing the interactions of the users and determining the interaction times of the users in the corresponding periods after the division and the corresponding interaction content; a response module connected with the auxiliary module for determining whether the training of the AI model meets the standard based on response activity and determining the reason why the training of the AI model does not meet the standard based on interaction similarity when it is determined that the training of the AI model does not meet the standard; wherein the response activity is the ratio of the actual registered users with interaction frequency greater than twice the interaction frequency to the total number of the actual registered users in a single division period; the interaction similarity is the average value of the actual interaction similarities of the actual registered users, and the actual interaction similarity of a single actual registered user is the ratio of the number of similar interaction features in the interaction content of the user to the interaction times of the actual registered user; a control module connected with the storage module, the learning module and the response module respectively for correcting the total number of model training data based on the interaction similarity difference value, reducing the context weight parameter based on the word ratio value, correcting the cosine similarity based on the interaction similarity ratio value, correcting the detail weight based on the inquiry ratio and increasing the maximum length based on the difficulty of the inquiry vocabulary; wherein the interaction similarity difference value is the difference between the interaction similarity and a preset interaction similarity; the word ratio value is the ratio of the number of the same interaction words to the interaction times; the interaction similarity ratio value is the ratio of the preset interaction similarity to the interaction similarity; the cosine similarity is a semantic analysis parameter for determining whether there is an associated relationship between the interaction content; the inquiry ratio is the ratio of the total number of inquiry words in the interaction content of the user to the interaction times. 2.The intelligent interaction system applied to AI training of claim 1, wherein, The response module is also used for determining whether the registration of the user meets the standard based on the interaction frequency and removing the family registered user from the total number of registered users or retaining the corresponding registered user in the total number of registered users when it is determined that the registered user is an active registered user. 3.The smart interactive system for AI training of claim 2, wherein, The response module is also used for determining whether the training of the AI model meets the standard based on the response activity after the family registered user is removed and the active registered user is retained and determining the reason why the training of the AI model does not meet the standard based on the interaction similarity according to the determination result. 4.The smart interaction system applied to AI training of claim 3, wherein, The response module is further configured to determine a reason for the AI model not meeting the standard based on the interaction similarity, and based on the interaction similarity difference to correct the total number of model training data if the reason is that the training accuracy of the AI model does not meet the standard, or based on the inquiry ratio to correct the detail weight if the reason is that the comprehensiveness of the response content of the AI model does not meet the standard. 5.The smart interactive system for AI training of claim 4, wherein, The control module is further configured to increase the total number of model training data based on the interaction similarity difference, and the increase of the total number of model training data is proportional to the interaction similarity difference. 6.The smart interactive system for AI training of claim 5, wherein, The control module is further configured to decrease the context weight parameter based on the word ratio after the total number of model training data is increased, and the decrease of the context weight parameter is proportional to the word ratio. 7.The smart interactive system for AI training of claim 6, wherein, The response module is further configured to determine a reason for the AI model not meeting the standard based on the interaction similarity after the context weight parameter is decreased, and correct the cosine similarity based on the interaction similarity ratio or correct the detail weight based on the inquiry ratio according to the determination result. 8.The smart interactive system for AI training of claim 7, wherein, The control module is further configured to increase the cosine similarity based on the interaction similarity ratio, and the increase of the cosine similarity is inversely proportional to the interaction similarity ratio. 9.The smart interactive system for AI training of claim 7, wherein, The control module is further configured to increase the detail weight based on the inquiry ratio, and the increase of the detail weight is proportional to the inquiry ratio. 10.The smart interaction system applied to AI training of claim 9, wherein, The control module is further configured to increase the maximum length based on the difficulty of the inquiry vocabulary after the detail weight is increased, and the increase of the maximum length is proportional to the difficulty of the inquiry vocabulary.

Images