Interaction method for finely adjusting model process and result

Abstract

The invention relates to the technical field of AI models, and discloses an interaction method for finely adjusting a model process and a result, which comprises the following steps: when a model receives a dialogue content sent by a user and generates a result, if a calculation process is judged to exist, displaying a current adjustable index list in the calculation process through an interface, displaying other adjustable indexes related to the current adjustable index to the user through a first recommendation mechanism; when the model generates a dialogue result, the result is finely adjusted, and the content of the related dialogue result is displayed to the user through a second recommendation mechanism in the fine adjustment; if it is judged that the pattern exists, a pattern configuration interface is generated, and the pattern configuration interface is used for adjusting the pattern. Through a recommendation mechanism and user behavior analysis, the model adjustment is more efficient, intelligent and accurate, the difficulty and cognitive burden of user adjustment are reduced, and the personalized customization ability and interaction experience are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of AI models, and more particularly, to an interaction method for fine-tuning the model process and results. Background Art

[0002] In the field of natural conversations of large AI models, currently, in-depth communication with the model is gradually achieved through conversations combined with question recommendations to obtain the desired results. After different users ask the same question, they may have different background knowledge, needs, and preferences. Through fine-tuning, the results can be better customized to be closer to the requirements of specific users, avoiding the inapplicability brought by the "one-size-fits-all" solution. However, if you want to perform batch fine-tuning on the process and results, there is currently no good interaction solution.

[0003] Specifically, there is a lack of relevant recommendations during fine-tuning, that is, users need to have a good understanding of the content, as well as a clear understanding of the adjusted information and the information after adjustment. Obviously, many users do not have such capabilities. At the same time, when there is a large amount of content to be adjusted, users may have omissions, confusion, etc. In addition, when the conclusion is based on the data uploaded by the user in the background and the user modifies some questions, the retrieved data should also be modified accordingly. Currently, modifying through natural language is obviously not accurate and convenient enough. Summary of the Invention

[0004] In view of this, the present invention proposes an interaction method for fine-tuning the model process and results, including:

[0005] When the model receives the conversation content sent by the user and generates a result, if it is determined that there is a calculation process, the current adjustable index list during the calculation process is displayed through the interface, and other adjustable indexes related to the current adjustable index are displayed to the user through the first recommendation mechanism;

[0006] When the model generates a conversation result, fine-tuning is performed on the result, and the relevant content of the conversation result is displayed to the user through the second recommendation mechanism during the fine-tuning;

[0007] If it is determined that there is a pattern, a pattern configuration interface is generated, and the pattern configuration interface is used for adjusting the pattern;

[0008] The first recommendation mechanism includes: collecting the historical conversation data of the user and extracting the preference features of the user;

[0009] Calculating the similarity with other users according to the preference features, and when the similarity between other users and the current user is greater than or equal to a pre-set similarity threshold, inputting the other users into the similar user set;

[0010] Extract the historical conversation data of the current user and each user in the set of similar users, calculate the recommended weight of each adjustable indicator according to the historical conversation data, arrange the recommended weights in descending order and select the top K adjustable indicators to generate the list of adjustable indicators;

[0011] The second recommendation mechanism generates content related to the conversation result based on the selection result of the current user in the adjustable list.

[0012] Further, the historical conversation data includes: user question records, user adjustment records, and user feedback records.

[0013] Further, when extracting the preference features of the user, it includes:

[0014] Classify each user question record through word segmentation and a pre-trained text classification model, count the occurrence times of each category and calculate the problem category distribution vector Q vec , where Q vec =(q1,q2,...,q n ), q n is the ratio of the number of questions in the nth category to the total number of questions;

[0015] Statistically analyze each adjustable indicator in the user adjustment record, calculate the frequency of each adjustable indicator and construct the adjustment behavior vector A vec , where A vec =(a1,a2,...,a m ), a m is the ratio of the number of adjustments of the mth adjustable indicator to the total number of adjustments;

[0016] According to the user feedback record, statistically analyze the proportion of each feedback type to the total number of feedbacks, and construct the feedback behavior vector F vec , where F vec =(f1,f2,f3), f k is the ratio of the number of times of the kth feedback type to the total number of feedbacks, k = 1, 2, 3; f1 represents the feedback type of adoption, f2 represents the feedback type of modification, and f3 represents the feedback type of rejection.

[0017] Further, when extracting the preference features of the user, it also includes:

[0018] Use the L2 normalization or Min-Max normalization method to map the values of the problem category distribution vector, the adjustment behavior vector, and the feedback behavior vector to the unit length range in turn to obtain the user preference vector U f , U f =[Q vec ∥A vec ∥F vec ;

[0019] After associating the user preference vector with the user ID, a user feature vector table is generated and stored in the database.

[0020] Further, when calculating the similarity with other users according to the preference features and inputting other users into the similar user set, it includes:

[0021] Collect the preference vectors of all users, and calculate the similarity through the following relationship:

[0022]

[0023] where S i is the similarity between the current user and the i-th user, U f is the preference vector of the current user, and U fi is the preference vector of the i-th user;

[0024] Compare the similarity with the similarity threshold. When the similarity is greater than or equal to the similarity threshold, continue to screen according to the question category distribution vector, adjustment behavior vector, and feedback behavior vector of the current user and the i-th user. When 0.8Q vec <Q vec,i <1.2Q vec 、0.8A vec <A vec,i <1.2A vec 、0.8F vec <F vec,i <1.2F vec is satisfied, input the user ID and the preference vector of the i-th user into the similar user set; Q vec,i is the question category distribution vector of the i-th user, A vec,i is the adjustment behavior vector of the i-th user, and F vec,i is the feedback behavior vector of the i-th user.

[0025] Further, extract the historical conversation data of the current user and each user in the similar user set, and calculate the recommended weight of each adjustable index according to the historical conversation data, including:

[0026] Obtain the residence time from when the model gives the result each time to when the user starts a new conversation again, and the occurrence times of each adjustable index in each conversation, and assign a first weight to each adjustable index according to the residence time and the occurrence times;

[0027] The first weight is calculated through the following relationship:

[0028] W1 = α·W T +β·W F ;

[0029] Among them, W1 represents the first weight, and W T is the residence time weight, W F is the occurrence quantity weight, α is the residence time sensitivity coefficient, β is the occurrence quantity sensitivity coefficient, and the value ranges of α and β are (0, 1].

[0030] Furthermore, the residence time weight and the occurrence quantity weight are calculated respectively by the following formulas:

[0031] W T = log(1 + T1);

[0032]

[0033] Among them, T1 is the residence time, C F，L is the occurrence quantity of the L-th adjustable index, and C is the occurrence quantity of all adjustable indexes.

[0034] Furthermore, extracting the historical conversation data of the current user and each user in the similar user set, and calculating the recommendation weight of each adjustable index according to the historical conversation data further includes:

[0035] Obtaining the residence time and the occurrence quantity of each similar user in the similar user set, obtaining the a m value corresponding to each adjustable index among the similar users, and calculating the second weight;

[0036] The second weight satisfies the following relationship:

[0037] W2 = y · (α · W T’ + β · W F’ + γ · a m );

[0038] Among them, y is the similar user sensitivity coefficient, the value range of y is (0, 1), W T’ is the residence time weight of the similar user, W F’ is the occurrence quantity weight of the similar user, γ is the adjustment times sensitivity coefficient, and the value range of β is (0, 1].

[0039] Furthermore, taking the sum of the first weight and the second weight as the recommendation weight of each adjustable index, and sorting each adjustable index according to the recommendation weight to obtain the adjustable index list containing K adjustable indexes.

[0040] Furthermore, the second recommendation mechanism is to generate a click entry for displaying the comparison content of the adjustable index before and after adjustment according to the adjustable index list.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] Through the first recommendation mechanism, other adjustable indicators related to the currently adjustable indicator are provided to the user, enabling the user to more comprehensively understand and operate the calculation process of the model, and avoiding missing key adjustment items. Through the second recommendation mechanism, relevant conversation results are recommended based on the user's adjustment behavior, reducing the user's understanding cost of the effect after adjustment, and improving the accuracy and efficiency of adjustment. The solution combines the user's historical interaction data, automatically extracts preference features, and recommends suitable adjustment items accordingly, enabling the user to complete the optimization adjustment without deeply understanding the model calculation process. By automatically generating the comparison content before and after adjustment, the impact of the adjustment is intuitively displayed, reducing the trouble that the user cannot understand the adjustment result due to lack of professional knowledge. By calculating the similarity between users and optimizing the recommendation in combination with the historical data of similar users, the pertinence and effectiveness of the recommendation are improved. Multi-dimensional factors such as residence time, adjustment frequency, and feedback behavior are used to calculate the recommendation weight, making the recommended adjustable indicators more in line with the user's needs, and enhancing the consistency and rationality of adjustment. The solution dynamically optimizes the recommendation in a data-driven manner, enabling the user to obtain adjustment options that more conform to their own habits and needs during the interaction process, and improving the interaction experience. Visual adjustment is supported through the pattern configuration interface, avoiding the understanding deviation caused by unclear text description for the user, and enhancing the intuitiveness and convenience of adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered as limiting the present invention. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0044] Figure 1 It is a flowchart of an interaction method for fine-tuning the model process and results provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the drawings and in combination with the embodiments.

[0046] Refer to Figure 1As shown in the figure, an embodiment of the present invention provides an interaction method for fine-tuning the model process and results, including:

[0047] S1: When the model receives the conversation content sent by the user and generates a result, if it is determined that there is a calculation process, the current adjustable index list in the calculation process is displayed through the interface, and other adjustable indexes related to the current adjustable index are displayed to the user through the first recommendation mechanism;

[0048] S2: When the model generates a conversation result, fine-tune the result, and the fine-tuning displays the content of the relevant conversation result to the user through the second recommendation mechanism;

[0049] S3: If it is determined that there is a pattern, a pattern configuration interface is generated, and the pattern configuration interface is used to adjust the pattern;

[0050] The first recommendation mechanism includes: collecting the historical conversation data of the user and extracting the preference features of the user;

[0051] Calculate the similarity with other users according to the preference features. When the similarity between other users and the current user is greater than or equal to the preset similarity threshold, input the other users into the similar user set;

[0052] Extract the historical conversation data of the current user and each user in the similar user set, calculate the recommendation weight of each adjustable index according to the historical conversation data, arrange the recommendation weights in descending order and select the top K adjustable indexes to generate an adjustable index list;

[0053] The second recommendation mechanism is to generate content related to the conversation result based on the selection result of the current user in the adjustable list.

[0054] It should be noted that by collecting the historical conversation data of the user and extracting the preference features of the user, a personalized adjustable index list can be generated for the specific needs and interests of each user. This refined recommendation mechanism enables users to make adjustments more efficiently during the conversation, thus improving the user experience.

[0055] By calculating the recommendation weight according to the similarity with similar users, dynamically adjusting and optimizing the recommended indexes to ensure that the content most matching the user's needs is displayed. The similarity judgment between different users and the data mining based on historical behaviors can make the recommendation results closer to the actual needs of users.

[0056] The second recommendation mechanism enables users to clearly understand the effects brought by the adjustment by generating relevant content of the conversation result and the comparison before and after the adjustment. Provide a click entry to view the comparison content, which improves the transparency of the adjustment process and the understandability of the operation.

[0057] Through the pattern configuration interface, users can modify the pattern to better adapt to the actual application scenario. This function provides users with a more intuitive and convenient interaction experience.

[0058] By combining the first recommendation mechanism and the second recommendation mechanism, relevant adjustment suggestions can be provided for the user's real-time needs, and the decision-making efficiency can be improved through real-time feedback. During the refined adjustment process, users can quickly find the key adjustment points, avoiding the interference of irrelevant or low-priority indicators, thereby improving the accuracy of the adjustment.

[0059] In some embodiments of the present application, the historical conversation data includes: user question records, user adjustment records, and user feedback records.

[0060] In some embodiments of the present application, when extracting the user's preference features, it includes:

[0061] Classify each user question record through word segmentation and a pre-trained text classification model, count the number of occurrences of each category, and calculate the problem category distribution vector Q vec , where Q vec =(q1,q2,...,q n ), q n is the ratio of the number of questions in the nth category to the total number of questions;

[0062] Statistically analyze each adjustable indicator in the user adjustment record, calculate the frequency of each adjustable indicator, and construct an adjustment behavior vector A vec , where A vec =(a1,a2,...,a m ), a m is the ratio of the number of adjustments of the mth adjustable indicator to the total number of adjustments;

[0063] According to the user feedback record, statistically analyze the proportion of each feedback type in the total number of feedbacks, and construct a feedback behavior vector F vec , where F vec =(f1,f2,f3), f k is the ratio of the number of times of the kth feedback type to the total number of feedbacks, k = 1, 2, 3; f1 represents the feedback type of adoption, f2 represents the feedback type of modification, and f3 represents the feedback type of rejection.

[0064] It should be noted that the role of the historical conversation data: user question records, user adjustment records, and user feedback records constitute the complete data source for the interaction between the user and the model. These data are used to extract the user's preference features and provide data support for subsequent recommendation and adjustment optimization.

[0065] Calculation method of the problem category distribution vector: By classifying the user's question through word segmentation and text classification model, the user's demand pattern can be modeled, thereby identifying the main problem categories that the user is concerned about.

[0066] Calculating the proportion of each type of problem (i.e., the ratio of the number of problems in the nth category to the total number of problems) can help the model understand the user's main concerns and optimize personalized recommendations.

[0067] Adjust the construction logic of the behavior vector: Count the adjustment frequencies of the user for different adjustable indicators, and calculate the proportion of each adjustable indicator (i.e., the ratio of the number of adjustments of the mth adjustable indicator to the total number of adjustments), which is used to characterize the user's adjustment preference.

[0068] This vector can help determine which adjustment items the user is more concerned about and increase the weights of these indicators in subsequent recommendations.

[0069] Meaning of the feedback behavior vector: By counting the user's feedback records (including adoption, modification, rejection), the user's recognition of the model's recommended content can be quantified.

[0070] The construction of the feedback behavior vector helps to identify which adjustment items the user truly recognizes and optimize the quality of subsequent recommendations.

[0071] In some embodiments of the present application, when extracting the user's preference features, it further includes:

[0072] Using L2 normalization or Min-Max normalization method to map the values of the problem category distribution vector, adjustment behavior vector, and feedback behavior vector to the unit length range in sequence to obtain the user preference vector U f , U f =[Q vec ∥A vec ∥F vec ;

[0073] After associating the user preference vector with the user ID, generate a user feature vector table and store it in the database.

[0074] It should be noted that when extracting the user's preference features, it is first necessary to analyze the user's historical data. This includes the user's question records, adjustment records, and feedback records. By processing these data, the user's personalized preferences can be extracted. To ensure the balance between different data, data normalization is usually performed.

[0075] The purpose of normalization is to map different types of data into a unified range, so that various types of data have the same impact on the final calculation. Without normalization, some data may dominate the calculation due to its large numerical range, resulting in unfair effects. After normalization, the contribution of each data item becomes more equitable.

[0076] In the process of normalization, there are two commonly used methods. One is to normalize the length of the data to 1, which can ensure that the directionality of each data item is reflected and not interfered by its original numerical size. The other method is to scale the data into a fixed range, usually from 0 to 1, which can make the values of all data items fall within the same interval for easy comparison and analysis.

[0077] Through such normalization processing, a user preference vector can be obtained. This preference vector can accurately reflect the specific preferences of users when asking questions, making adjustments, and providing feedback. For example, some users may be more inclined to adjust certain parameters, while some users may pay more attention to the selection of question categories.

[0078] Finally, associate the user's preference vector with their user ID to form a complete user feature vector table and store it in the database. In this way, when making recommendations in the future, the historical preference information of the user can be obtained by querying this table, thereby achieving more accurate personalized recommendations. Through continuous accumulation and update, the user's preference vector can become more refined, further improving the quality and accuracy of recommendations.

[0079] In some embodiments of the present application, when calculating the similarity with other users according to the preference characteristics and inputting other users into the similar user set, it includes:

[0080] Collect the preference vectors of all users and calculate the similarity through the following relationship:

[0081]

[0082] where S i is the similarity between the current user and the i-th user, U f is the preference vector of the current user, and U fi is the preference vector of the i-th user;

[0083] Compare the similarity with the similarity threshold. When the similarity is greater than or equal to the similarity threshold, continue to screen according to the question category distribution vector, adjustment behavior vector, and feedback behavior vector of the current user and the i-th user. When 0.8Q vec <Q vec,i <1.2Q vec 、0.8A vec <A vec,i<1.2A vec 、0.8F vec <F vec,i <1.2F vec When <1.2A, 0.8F < F < 1.2F, input the user ID and preference vector of the i-th user into the similar user set; Q vec,i is the question category distribution vector of the i-th user, A vec,i is the adjustment behavior vector of the i-th user, F vec,i is the feedback behavior vector of the i-th user.

[0084] It should be noted that according to the preference characteristics of users, it is first necessary to calculate the similarity between the current user and other users. To achieve this, the preference vectors of all users are collected and used to calculate the similarity. The calculation of similarity is completed by comparing the preference vector of the current user with the preference vectors of each other user.

[0085] The basic idea of similarity calculation is to judge whether they have similar behaviors and needs by measuring the similarity between the preference vectors of two users. This calculation is usually based on the distance or angle difference between vectors. For example, methods such as the Euclidean distance and cosine similarity between vectors can be used to measure this similarity.

[0086] Once the similarity between each pair of users is calculated, it is compared with a preset similarity threshold. Only when the similarity is greater than or equal to this threshold, the current user will be regarded as similar to other users and enter the similar user set. On this basis, in order to further improve the accuracy of similarity, according to the specific preference data of each user, screening is carried out, especially focusing on three key features: question category distribution, adjustment behavior, and feedback behavior.

[0087] During the screening process, compare the differences between the current user and the i-th user in these three aspects. If these differences are within a certain range (such as 0.8 < < 1.2), it is considered that the preferences of these two users in a specific field are very close. After meeting this condition, the user ID and preference vector of the i-th user will be added to the similar user set.

[0088] The ultimate goal of this process is to ensure that other users with preferences similar to the current user can be found, so as to provide more personalized and accurate recommendations for the current user. By continuously updating the similar user set, the recommendation effect can be continuously optimized and the user experience can be improved.

[0089] In some embodiments of the present application, extract the historical conversation data of the current user and each user in the similar user set, and calculate the recommendation weight of each adjustable index according to the historical conversation data, including:

[0090] Obtain the stay time from when the model gives a result each time until the user starts a new conversation again, as well as the occurrence count of each adjustable indicator in each conversation, and assign a first weight to each adjustable indicator according to the stay time and the occurrence count;

[0091] The first weight is calculated through the following relationship:

[0092] W1 = α·W T + β·W F ;

[0093] Where, W1 represents the first weight, W T is the stay time weight, W F is the occurrence count weight, α is the stay time sensitivity coefficient, β is the occurrence count sensitivity coefficient, and the value ranges of α and β are (0, 1].

[0094] In some embodiments of the present application, the stay time weight and the occurrence count weight are calculated through the following formulas respectively:

[0095] W T = log(1 + T1);

[0096]

[0097] Where, T1 is the stay time, C F，L is the occurrence count of the L-th adjustable indicator, and C is the occurrence count of all adjustable indicators.

[0098] It should be noted that the historical conversation data of the current user and each user in the similar user set is extracted to calculate the recommendation weight of each adjustable indicator. Specifically, first record the stay time from when the model gives a result each time until the user starts a new conversation again, as well as the occurrence count of each adjustable indicator in each conversation. This information provides the basic data for calculating the recommendation weight.

[0099] Next, assign a first weight to each adjustable indicator according to the stay time and the occurrence count. The calculation method of the first weight considers two main factors: the stay time weight and the occurrence count weight. The stay time weight reflects the length of time the user stays on a certain indicator, while the occurrence count weight represents the frequency of the indicator appearing in the user's conversation. By combining these two dimensions, an appropriate weight can be assigned to each adjustable indicator.

[0100] During the calculation process, the stay time and the occurrence count each have their own sensitivity coefficients, which are used to adjust the degree of their influence on the recommendation weight. The value range of the sensitivity coefficient is from 0 to 1, that is, the contribution of the stay time and the occurrence count to the recommendation weight can be flexibly adjusted to more precisely match the actual needs and preferences of the user.

[0101] It should be noted that this calculation method can not only consider the actual interaction data of users, but also further optimize the accuracy and personalization level of recommendations through the adjustment of sensitivity coefficients. In this way, the recommended order of adjustable indicators can be dynamically adjusted according to the user's behavior history, making the recommended results more in line with the actual needs and expectations of users.

[0102] In some embodiments of the present application, when extracting the historical conversation data of the current user and each user in the set of similar users, and calculating the recommendation weight of each adjustable indicator according to the historical conversation data, it further includes:

[0103] Obtain the residence time and appearance quantity of each similar user in the set of similar users, obtain the a m value corresponding to each adjustable indicator among the similar users, and calculate the second weight;

[0104] The second weight satisfies the following relationship:

[0105] W2 = y · (α · W T’ + β · W F’ + γ · a m );

[0106] Wherein, y is the sensitivity coefficient of similar users, the value range of y is (0, 1), W T’ is the residence time weight of similar users, W F’ is the appearance quantity weight of similar users, γ is the sensitivity coefficient of adjustment times, and the value range of β is (0, 1].

[0107] It should be noted that the calculation methods of W F’ and W F are the same, and the calculation methods of W T’ and W T are the same. The difference is that W F’ refers to the appearance quantity of a specific adjustable indicator in each conversation, while a m refers to the number of times the specific adjustable indicator has been adjusted.

[0108] In addition, the residence time weight reflects the duration of user stay on each adjustable indicator. The calculation method of this weight is based on the ratio of the residence time of the user on a certain adjustable indicator to the residence time of other indicators. Specifically, an indicator with a higher residence time weight means that the user stays on this indicator for a relatively longer time, which may indicate that the user has a higher degree of attention to this indicator.

[0109] The appearance quantity weight is calculated based on the frequency of each adjustable indicator appearing in the user's conversation. The appearance quantity weight is obtained by comparing the number of times a certain indicator appears in all conversations with the number of times other indicators appear. An indicator with a higher frequency will obtain a higher weight, indicating its importance in the conversation.

[0110] In this way, not only can we focus on the user's dwell time on a certain metric, but also combine the frequency of occurrence of this metric to comprehensively evaluate the importance of each adjustable metric to the user, so as to more accurately assign recommendation weights to each adjustable metric.

[0111] In some embodiments of the present application, the sum of the first weight and the second weight is used as the recommendation weight for each adjustable metric, and each adjustable metric is sorted according to the recommendation weight to obtain an adjustable metric list containing K adjustable metrics.

[0112] In some embodiments of the present application, the second recommendation mechanism is to generate a click entry for displaying the comparison content of the adjustable metric before and after adjustment according to the adjustable metric list.

[0113] It should be noted that the first weight and the second weight are added to obtain the recommendation weight for each adjustable metric. This recommendation weight reflects the comprehensive evaluation based on the user's historical behavior, similar user data, and other relevant factors. Then, the adjustable metrics are sorted according to these recommendation weights, and the top K adjustable metrics will be selected into the final adjustable metric list. This process ensures that in the refined adjustment process, the user can see the most relevant and worthy-adjusting metrics, thereby improving the efficiency and accuracy of the adjustment.

[0114] The second recommendation mechanism is based on the sorted adjustable metric list to generate a display of the comparison of the effects of each adjustable metric before and after adjustment. Specifically, a "click entry" will be provided for each adjustable metric. When the user clicks, they can view the changes of the metric before and after adjustment to help the user more clearly understand the actual impact of the adjustment. Such a display method can not only provide intuitive feedback, but also enhance the user's adjustment experience and help the user make more accurate adjustment decisions.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the specific implementation manners of the present invention can still be modified or equivalently replaced, and any modification or equivalent replacement without departing from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. An interactive method for fine-tuning the model process and results, characterized in that, Including: When the model receives the conversation content sent by the user and generates a result, if it is determined that there is a calculation process, the list of currently adjustable indicators in the calculation process is displayed through the interface, and other adjustable indicators related to the currently adjustable indicators are displayed to the user through the first recommendation mechanism; When the model generates a conversation result, it performs a refined adjustment on the result, and the refined adjustment displays the content of the relevant conversation result to the user through the second recommendation mechanism; If it is determined that there is a pattern, a pattern configuration interface is generated, and the pattern configuration interface is used for adjusting the pattern; The first recommendation mechanism includes: collecting the user's historical conversation data and extracting the user's preference features; Calculating the similarity with other users according to the preference features, and when the similarity between other users and the current user is greater than or equal to a pre-set similarity threshold, inputting the other users into the similar user set; Extracting the historical conversation data of the current user and each user in the similar user set, calculating the recommendation weight of each adjustable indicator according to the historical conversation data, arranging the recommendation weights in descending order and selecting the top K adjustable indicators to generate the list of adjustable indicators; The second recommendation mechanism is to generate content related to the conversation result based on the selection result of the current user in the adjustable list.

2. The interactive method for fine-tuning the model process and results according to claim 1, wherein The historical conversation data includes: user question records, user adjustment records, and user feedback records.

3. The interactive method for fine-tuning the model process and results according to claim 2, characterized in that, When extracting the user's preference features, it includes: Classify each of the user question records through word segmentation and a pre-trained text classification model, count the occurrences of each category, and calculate the question category distribution vector Q vec , where Q vec =(q1, q2,..., q n ), q n is the ratio of the number of questions in the nth category to the total number of questions; Statistically analyze each adjustable indicator in the user adjustment record, calculate the frequency of each adjustable indicator, and construct an adjustment behavior vector A vec , where A vec = (a1, a2,..., a m ), and a m is the ratio of the adjustment times of the m-th adjustable indicator to the total adjustment times; According to the user feedback records, calculate the proportion of each feedback type in the total number of feedbacks, and construct the feedback behavior vector F vec , where F vec =(f1, f2, f3), f k is the ratio of the number of the k-th feedback type to the total number of feedbacks, k = 1, 2, 3; f1 represents the feedback type of adoption, f2 represents the feedback type of modification, and f3 represents the feedback type of rejection.

4. The interactive method for fine-tuning the model process and results according to claim 3, wherein When extracting the user's preference features, it also includes: The numerical values of the problem category distribution vector, the adjusted behavior vector, and the feedback behavior vector are sequentially mapped within the unit length range by using the L2 normalization or the Min-Max normalization method to obtain the user preference vector U f , U f =[Q vec ∥A vec ∥F vec ; After associating the user preference vector with the user ID, generating a user feature vector table and storing it in the database.

5. The interactive method for fine-tuning the model process and results according to claim 4, characterized in that, When calculating the similarity with other users according to the preference features and inputting the other users into the similar user set, it includes: Collecting the preference vectors of all users and calculating the similarity through the following relationship: Among them, S i is the similarity between the current user and the i-th user, U f is the preference vector of the current user, U fi is the preference vector of the i-th user; Compare the similarity with a similarity threshold. When the similarity is greater than or equal to the similarity threshold, continue the screening based on the question category distribution vector, adjustment behavior vector, and feedback behavior vector of the current user and the i-th user. When 0.8Q vec <Q vec,i <1.2Q vec 、0.8A vec <A vec,i <1.2A vec 、0.8F vec <F vec,i <1.2F vec is satisfied, input the user ID of the i-th user and the preference vector into the set of similar users; Q vec,i is the question category distribution vector of the i-th user, A vec,i is the adjustment behavior vector of the i-th user, and F vec,i is the feedback behavior vector of the i-th user.

6. The interactive method for fine-tuning the model process and results according to claim 5, characterized in that, Extracting the historical conversation data of the current user and each user in the similar user set, and calculating the recommendation weight of each adjustable indicator according to the historical conversation data, including: Obtaining the stay time from when the model gives a result until the user starts a new conversation again each time, and the number of occurrences of each adjustable indicator in each conversation, and assigning a first weight to each adjustable indicator according to the stay time and the number of occurrences; The first weight is calculated through the following relationship: W1 = α·W T + β·W F ; Among them, W1 represents the first weight, W T is the residence time weight, W F is the occurrence quantity weight, α is the residence time sensitivity coefficient, β is the occurrence quantity sensitivity coefficient, and the value ranges of α and β are (0, 1].

7. The interactive method for fine-tuning the model process and results according to claim 6, characterized in that The stay time weight and the occurrence number weight are calculated respectively by the following formulas: W T = log(1 + T1); Among them, T1 is the residence time, C F，L is the occurrence number of the L-th adjustable index, and C is the occurrence number of all adjustable indexes.

8. The interactive method for fine-tuning the model process and results according to claim 7, wherein Extracting the historical conversation data of the current user and each user in the similar user set, and calculating the recommendation weight of each adjustable indicator according to the historical conversation data, and also including: Obtain the residence time and occurrence quantity of each similar user in the set of similar users, and obtain the a m value corresponding to each adjustable index among the similar users, and calculate the second weight; The second weight satisfies the following relationship: W2 = y·(α·W T’ + β·W F’ + γ·a m ); Among them, y is the sensitive coefficient of similar users, and the value range of y is (0, 1), W T’ is the residence time weight of similar users, W F’ is the appearance quantity weight of similar users, γ is the sensitive coefficient of adjustment times, and the value range of β is (0, 1].

9. The interactive method for fine-tuning the model process and results according to claim 8, wherein Taking the sum of the first weight and the second weight as the recommendation weight of each adjustable indicator, sorting each adjustable indicator according to the recommendation weight, and obtaining the list of adjustable indicators containing K adjustable indicators.

10. The interactive method for fine-tuning the model process and results according to claim 9, characterized in that, The second recommendation mechanism is to generate a click entry for displaying the comparison content of the adjustable indicator before and after adjustment according to the list of adjustable indicators.

Images