Training data construction method and device, equipment and medium

Abstract

The invention provides a training data construction method and device, equipment and a medium, and relates to the technical field of big data analysis and artificial intelligence. The method comprises the following steps: acquiring feature data of each sample in a sample set under multi-dimensional features; aiming at each dimension feature in the multi-dimension features, taking feature data of each sample under each dimension feature as observation data; judging whether the observation data corresponding to each dimension feature accords with normal distribution or not, and constructing a data grading boundary for the observation data corresponding to each dimension feature according to a judgment result; according to the boundary of the data grading, determining the numerical value degree of the feature data of each sample under each dimension feature; and constructing a training sample data set for large language model learning according to the numerical values of the samples and the annotation information. According to the method, the numerical value degree of the observation data of each sample is intelligently determined, and the training sample is constructed according to the numerical value degree of the observation data, so that the learning effect of the large language model is improved.

Description

Technical Field

[0001] The present application relates to the field of big data analysis and artificial intelligence technology, and in particular to a training data construction method, device, equipment and medium. Background Art

[0002] For financial institutions, the analysis of key indicator data is crucial to operations. However, manually analyzing indicator data and giving corresponding processing methods based on the analysis results is not only inefficient, but also easily affected by various human factors.

[0003] Therefore, how to combine big data, artificial intelligence and other methods to automatically realize indicator data analysis and determine the corresponding processing methods based on the analysis results has become an important direction for the development of financial technology. Summary of the invention

[0004] The present application provides a training data construction method, apparatus, device and medium for constructing training samples for learning a large language model, and using the large language model to predict corresponding processing methods based on indicator data.

[0005] In a first aspect, the present application provides a training data construction method, comprising:

[0006] Obtain feature data of each sample in the sample set under multi-dimensional features;

[0007] For each dimensional feature in the multi-dimensional features, taking the feature data of each sample under each dimensional feature as observation data;

[0008] Determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and construct a data classification boundary for the observed data corresponding to each dimensional feature according to the determination result;

[0009] According to the dividing line of the data classification, respectively determine the numerical value of the characteristic data of each sample under each dimensional feature;

[0010] According to the numerical values ​​of each sample under each dimensional feature and the labeling information of each sample, a training sample data set for large language model learning is constructed.

[0011] In a second aspect, the present application provides a training data construction device, comprising:

[0012] A feature data acquisition module is used to obtain feature data of each sample in the sample set under multi-dimensional features;

[0013] An observation data determination module, configured to use, for each dimensional feature in the multi-dimensional features, feature data of each sample under each dimensional feature as observation data;

[0014] A dividing line construction module is used to determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and to construct a dividing line for data classification for the observed data corresponding to each dimensional feature according to the determination result;

[0015] A numerical value determination module, used to determine the numerical value of the characteristic data of each sample under each dimensional feature according to the dividing line of the data classification;

[0016] The training sample data set construction module is used to construct a training sample data set for large language model learning according to the numerical value of each sample under each dimensional feature and the labeling information of each sample.

[0017] In a third aspect, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a training data construction method as described in any one of the embodiments of the present application is implemented.

[0018] In a fourth aspect, the present application further provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the training data construction method as described in any one of the embodiments of the present application.

[0019] In a fifth aspect, the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the training data construction method as described in any one of the embodiments of the present application.

[0020] The training data construction method, device, equipment and medium provided in the present application first determine whether the observed data of the sample under each dimensional feature conforms to the normal distribution, and construct a dividing line for data classification for the observed data based on the judgment result. Then, the numerical level of the characteristic data of the sample under each dimensional feature is determined based on the dividing line. Finally, a training sample data set for learning a large language model is constructed based on the numerical level and annotation information of each sample. In this way, the numerical level of the characteristic data of each sample can be intelligently determined, and then the training sample can be constructed based on the numerical level, realizing the transformation from quantitative analysis to qualitative analysis, thereby improving the learning effect of the large language model. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 A flowchart of a training data construction method provided in an embodiment of the present application;

[0023] Figure 2 Schematic diagram of data classification in the training data construction method provided in the embodiment of the present application Figure 1 ;

[0024] Figure 3 Schematic diagram of data classification in the training data construction method provided in the embodiment of the present application Figure 2 ;

[0025] Figure 4 A flowchart for determining whether observed data conforms to a normal distribution in a training data construction method provided in an embodiment of the present application.

[0026] Figure 5 A flowchart of calculating normal distribution similarity in the training data construction method provided in an embodiment of the present application;

[0027] Figure 6 A schematic diagram of the structure of a training data construction device provided in an embodiment of the present application;

[0028] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0029] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0030] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only the parts related to the present application, rather than all structures, are shown in the accompanying drawings.

[0031] It should be noted that similar numbers and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance. The acquisition, storage, use, processing, etc. of data in the technical solution of this application comply with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned, and they should be regarded as exemplary. The purpose is only to illustrate the feasibility of the implementation of the technical solution of this application, but it does not mean that the applicant has or must use the relevant content of the solution.

[0032] Figure 1 A flowchart of a training data construction method provided in an embodiment of the present application. This embodiment is applicable to analyzing financial indicator data and constructing training sample data for large language model learning based on the analysis results. The method can be executed by a training data construction device, which can be implemented in software and / or hardware, preferably configured in an electronic device, such as a computer device, an intelligent terminal or a server. Figure 1 As shown, the method specifically includes:

[0033] S101, obtaining feature data of each sample in the sample set under multi-dimensional features.

[0034] S102: For each dimensional feature in the multi-dimensional feature, the feature data of each sample under each dimensional feature is used as observation data.

[0035] S103: Determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and construct a data classification boundary for the observed data corresponding to each dimensional feature according to the determination result.

[0036] S104. According to the dividing line of the data classification, determine the numerical value of the characteristic data of each sample in each dimensional feature.

[0037] S105. Construct a training sample data set for large language model learning based on the numerical value of each sample in each dimensional feature and the annotation information of each sample.

[0038] The sample set refers to the user sample set. The observation data can refer to the user's financial indicator data. Financial indicator data is the basis of all financial analysis. These indicators can also be collectively referred to as features. The specific values ​​of the features of multiple dimensions of each user sample are observed to obtain the observation data of each user under each dimension feature. Then, the observation data of each feature of all users are collected to form the observation data set of the user's overall features.

[0039] In the collection stage of observation data, taking into account possible data missing and abnormal problems, the data is cleaned and preliminarily processed in the embodiment of the present application, including eliminating missing data and abnormal data. For example, for user A, it is necessary to observe the values ​​of its 10 features. When only 5 features of the user have data, and the remaining features are not counted and displayed as null values, the data of the user can be abandoned. For another example, for user B, all 10 features have data, but for some features, there are values ​​or data types that are obviously not within their normal value range, the data of the user can also be abandoned to avoid interfering with the overall analysis of other normal data. After the above cleaning and preliminary processing operations are completed, the observation data finally used for further analysis can be obtained.

[0040] First, determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and build the data classification boundary for the observed data corresponding to each dimensional feature based on the judgment results. Then, based on the data classification boundary, determine the numerical value of each sample's feature data under each dimensional feature. Finally, based on the numerical value of each sample under each dimensional feature and the annotation information of each sample, build a training sample data set for large language model learning.

[0041] It should be noted that before constructing the dividing line for data grading, it is necessary to first determine whether the observed data conforms to the normal distribution. This is because data with different distribution characteristics cannot use the same grading mechanism, otherwise the divided categories will be seriously mismatched. For observed data whose data distribution is close to the normal distribution, the normal grading method can be used for grading. However, if the data distribution characteristics obviously do not conform to the normal distribution, the use of the normal grading method will lead to extremely abnormal grading results. For example, the number of samples classified into each grade is seriously unbalanced, and even the vast majority of samples will be classified into extremely high or extremely low extreme areas, which completely fails to achieve the goal of accurate sample differentiation, which is very unfavorable for financial data analysis.

[0042] Therefore, in one embodiment, if the observed data obeys a normal distribution, the normal partitioning method is used to perform data binning, otherwise the data may obey a disorderly distribution, and a non-normal partitioning method should be used to perform data binning, thereby improving the accuracy of data binning.

[0043] Specifically, the classification can be set to 9 levels: extremely high, very high, relatively high, slightly high, medium, slightly low, relatively low, very low, and extremely low. The demarcation standard for each level is not divided by specific values, because the numerical ranges of different dimensional features are different. If fixed numerical boundaries are blindly adopted, the accuracy of the classification will be seriously affected.

[0044] If it conforms to the normal distribution, it is divided into bins according to the normal division method, that is, the dividing line of the data bins is constructed according to the mean and standard deviation of the observed data. The mean μ and standard deviation σ of the observed data of a single feature are calculated, and the dividing lines of each bin are constructed based on μ and σ, for example:

[0045] Q1=μ-2×σ(approximately 2.5% quantile)

[0046] Q2=μ-1.6×σ(approximately 6% quantile)

[0047] Q3=μ-1.2×σ(approximately 12% quantile)

[0048] Q4=μ-0.8×σ(approximately 22% quantile)

[0049] Q5=μ+0.8×σ(approximately 78% quantile)

[0050] Q6=μ+1.2×σ(approximately 88% quantile)

[0051] Q7=μ+1.6×σ(approximately 94% quantile)

[0052] Q8 = μ + 2 × σ (approximately 97.5% quantile)

[0053] When the data falls below Q1, it is extremely low; when it falls between Q1 and Q2, it is very low; when it falls between Q2 and Q3, it is relatively low; when it falls between Q3 and Q4, it is slightly low; when it falls between Q4 and Q5, it is medium; when it falls between Q5 and Q6, it is slightly high; when it falls between Q6 and Q7, it is relatively high; when it falls between Q7 and Q8, it is very high; when it falls above Q8, it is extremely high. Thus, the index can be divided into nine equal parts for level analysis.

[0054] If it does not conform to the normal distribution, it is divided into bins according to the non-normal division method, that is, the data binning boundary is constructed according to the first quartile, the third quartile and the quartile interval. The numerical value of the observed data is determined according to the data binning boundary. Among them, the first quartile is the 25% quartile q 0.25 , the third quartile is the 75% quantile q 0.75 , the interquartile range IQR is the difference between the two, that is, IQR = q 0.75 -q 0.25 On this basis, the dividing lines of each tier are constructed, for example:

[0055] Q1'=q 0.25 -1.5 × IQR

[0056] Q2'=q 0.25 -1×IQR

[0057] Q3'=q 0.25 -0.5 × IQR

[0058] Q4'=q 0.25

[0059] Q5'=q 0.75

[0060] Q6'=q 0.75 +0.5×IQR

[0061] Q7'=q 0.75 +1×IQR

[0062] Q8'=q 0.75 +1.5×IQR

[0063] Similarly, Q1'~Q8' can be divided into nine equal parts for the analysis of the level of indicators. When the data falls below Q1', it is extremely low; between Q1' and Q2', it is very low; between Q2' and Q3', it is relatively low; between Q3' and Q4', it is slightly low; between Q4' and Q5', it is medium; between Q5' and Q6', it is slightly high; between Q6' and Q7', it is relatively high; between Q7' and Q8', it is very high; and above Q8', it is extremely high.

[0064] Figure 2 Schematic diagram of data classification in the training data construction method provided in the embodiment of the present application Figure 1 , showing the classification according to the normal classification method. Figure 3 Schematic diagram of data classification in the training data construction method provided in the embodiment of the present application Figure 2 , showing the situation of grading according to the non-normal grading method. As can be seen from the figure, the grading obtained by the two different methods in the embodiment of the present application has a certain degree of similarity, and the grading effect obtained by the non-normal grading method is close to the grading effect under the normal distribution. Therefore, regardless of whether the observed data conforms to the normal distribution, the grading effect obtained in the embodiment of the present application is more accurate and scientific, which is conducive to the subsequent accurate judgment of the numerical value of the data, and at the same time, it also provides a data basis for the analysis of data change trends.

[0065] At the same time, in order to completely liberate human resources, with the help of the learning ability of AI big models, machines can replace financial practitioners to extract business strategy information from the previous analysis results, which is the future development direction of financial technology. In the embodiment of the present application, large language model technology, such as BERT and GPT, is used to achieve rapid learning of sample data, so as to replace people to perform analysis and summary. However, the learning efficiency of large model technology mainly depends on the deconstruction and information extraction of data results. In fact, AI big models are good at text learning and prediction, but they are not good at numerical calculation and quantitative analysis. Therefore, in the embodiment of the present application, the quantitative analysis of financial indicator data is transformed into qualitative analysis, that is, after constructing the dividing line of the data classification of the observation data, the numerical level of each sample feature data is determined according to the dividing line of the data classification, and then the training sample data set is constructed according to the numerical level and annotation information of each sample, thereby improving the learning effect of the large language model.

[0066] For example, the level classification results of different characteristic index data that have been obtained and the corresponding processing methods (annotation information) are organized into a JSON format file. The data arrangement format is shown in the following table:

[0067] Sample No. Feature 1 Feature 2 Feature 3 …… Feature N Processing Sample 1 Very high Very high Higher …… Lower Method 1 Sample 2 Slightly higher Higher Very low …… Slightly lower Method 3 Sample 3 Slightly higher Very high Slightly lower …… Higher Method 4 Sample 4 Very low Higher Very high …… Slightly higher Method 7 …… …… …… …… …… …… …… Sample M Very low Higher Slightly lower …… Slightly higher Method 3

[0068] Among them, each column in the table corresponds to a feature, with a total of N features; each row corresponds to the classification result of the numerical value of a sample in each dimension feature and the marked processing method of the sample, with a total of M samples. The processing method here is the method summarized by experienced financial practitioners, and its specific form is presented in the form of text processing opinions.

[0069] Then, the data in this format is collected and organized, packaged into a JSON file and input into the AI ​​big language model. The model will take each line of record as a sample, and learn the relationship between the numerical distribution of different levels of each feature and the proposed processing methods by financial practitioners one by one. When the number of samples M is large enough, the big language model will be trained smart enough to accurately obtain what countermeasures should be taken for different feature level combinations from the training samples. On this basis, when the numerical distribution of the features of each dimension of the new target is input, the AI ​​big language model can directly predict the reasonable processing method for the target, thereby realizing artificial intelligence replacement for the business processing of financial practitioners.

[0070] The technical solution of the embodiment of the present application first determines whether the observed data of the sample under each dimensional feature conforms to the normal distribution, and constructs a dividing line for data classification for the observed data based on the judgment result. Then, the numerical level of the characteristic data of the sample under each dimensional feature is determined based on the dividing line. Finally, a training sample data set for learning a large language model is constructed based on the numerical level and annotation information of each sample. In this way, the numerical level of the characteristic data of each sample can be intelligently determined, and then a training sample can be constructed based on the numerical level, realizing the transformation from quantitative analysis to qualitative analysis, thereby improving the learning effect of the large language model.

[0071] Figure 4 A flowchart for determining whether observed data conforms to a normal distribution in a training data construction method provided in an embodiment of the present application.

[0072] S201. Based on the statistical hypothesis verification method, obtain the statistical hypothesis verification value of whether the observed data corresponding to each dimensional feature conforms to the normal distribution.

[0073] S202. Calculate the mean, standard deviation and statistical characteristic values ​​of the observed data, where the statistical characteristic values ​​include skewness, kurtosis, and quantiles corresponding to the mean increased by a set number of times the standard deviation.

[0074] S203: If the statistical characteristic value meets the statistical characteristic judgment standard of the normal distribution, the normal distribution similarity is calculated according to the statistical characteristic value and the statistical hypothesis verification value.

[0075] S204: Determine whether the observed data conforms to the normal distribution according to the normal distribution similarity and a preset threshold.

[0076] In the embodiment of the present application, statistical hypothesis verification and the statistical characteristics of the normal distribution are combined to comprehensively judge whether the observed data conforms to the normal distribution, which has higher accuracy.

[0077] Specifically, the statistical hypothesis verification value can be obtained by the statistical hypothesis verification method, and the statistical hypothesis verification value can be compared with the set threshold to determine whether the observed data conforms to the normal distribution. For example, the observed data is input into the KS (Kolmogorov-Smirnov) verification model to obtain the verification value P value. From the statistical theory, when the p value is less than 0.05, it is considered not to be a normal distribution at the statistical level. However, taking a p value less than 0.05 as the judgment dividing line is only a judgment method at the statistical level, that is, a negative judgment is made when the data is extremely unlike a normal distribution. In fact, even if the p value reaches 0.2 or even 0.3, it is difficult to say how much similarity the observed data has with the normal distribution. Therefore, in order to ensure the accuracy of the data, preferably, a p value less than 0.4 can be used as the judgment basis. When the p value exceeds 0.4, the statistical hypothesis verification can have a relatively large guarantee to ensure the consistency of the data with the normal distribution.

[0078] It is not enough to make a judgment only by statistical hypothesis verification. In the embodiment of the present application, the characteristics of the normal distribution itself are also introduced to assist in verification. The statistical characteristics of the normal distribution are symmetrical distribution about the mean μ, with a skness of 0 and a kurtosis of 4. The corresponding quantile values ​​after the mean increases by 1, 2, and 3 times the standard deviation σ are 84%, 98%, and 99.9% quantiles, that is, q 0.84 ,q 0.98 ,q 0.999 Therefore, we only need to construct the following judgment criteria based on these characteristics, and when all the judgment criteria are established, we can conclude that the data conforms to the normal distribution in terms of statistical characteristics:

[0079] μ+0.8*σ≤q 0.84 ≤μ+1.2*σ

[0080] μ+1.6*σ≤q 0.98 ≤μ+2.4*σ

[0081] μ+2.4*σ≤q 0.999 ≤μ+3.6*σ

[0082] -0.25≤skness≤0.25

[0083] 3.5≤kurtosis≤4.5

[0084] In the embodiment of the present application, the mean μ and standard deviation σ of the observed data of a certain feature are calculated, and then the skness, kurtosis, and the quantile value q corresponding to the mean increasing by 1, 2, and 3 times the standard deviation σ are calculated. 0.84 ,q 0.98 ,q 0.999 , if the skness, kurtosis, and the corresponding quantile values ​​after the mean increases by 1, 2, and 3 times the standard deviation σ all meet the above-mentioned judgment criteria of the normal distribution statistical characteristics, then the normal distribution similarity is further calculated based on the statistical characteristic values ​​and the statistical hypothesis verification values. For example, the normal distribution similarity can be obtained by calculating the product between each statistical characteristic and the statistical hypothesis verification value, and then summing them up. Using the normal distribution similarity to measure whether the observed data conforms to the normal distribution is equivalent to combining the statistical hypothesis testing method with the judgment method of the statistical characteristics of the normal distribution, and combining the p-value calculated by the statistical hypothesis verification method with the q obtained in the normal distribution statistical characteristics method. 0.84 ,q 0.98 ,q 0.999 The combined calculation and analysis of the three factors, skness, and kurtosis, is more accurate than the single use of statistical hypothesis testing method or statistical characteristics of normal distribution method to make judgments.

[0085] The technical solution of this embodiment combines the statistical hypothesis check value with the statistical characteristics of the normal distribution to determine whether the observed data conforms to the normal distribution. Therefore, by combining the statistical hypothesis check and the judgment method of the statistical characteristics of the normal distribution, it is accurately determined whether the observed data conforms to the normal distribution, and then different demarcation methods are used to construct the demarcation line for data grading according to the judgment result. While realizing intelligence, it can accurately quantify the numerical value of the observed data, make the data analysis results clear at a glance, and fully display the subtle changes in the data, thereby improving the accuracy of data grading.

[0086] In one embodiment, the Figure 5 The process shown calculates the normal distribution similarity. Figure 5 Flow chart of calculating the normal distribution similarity in the training data construction method provided in the embodiment of the present application. This embodiment is further optimized on the basis of any of the above embodiments. Figure 5 As shown, the process of calculating the normal distribution similarity includes:

[0087] S301. Calculate the deviation of each statistical characteristic value according to a statistical characteristic judgment standard.

[0088] S302: Multiply the deviations of the statistical characteristic values ​​by their respective weight coefficients, and sum the multiplication results to obtain a sum value.

[0089] S303: Calculate the product of the sum value and the statistical hypothesis check value to obtain the normal distribution similarity.

[0090] Among them, the deviation degree indicates the degree of deviation between the statistical characteristic value calculated based on the observed data and the statistical characteristic judgment standard that conforms to the normal distribution, which is used to measure the gap between the two. On this basis, a weight coefficient is set for each statistical characteristic value, and the deviation degree of each statistical characteristic value is multiplied by each weight coefficient, and the multiplication results are summed to obtain the sum value, and then the product of the sum value and the statistical hypothesis verification value is calculated to obtain the normal distribution similarity.

[0091] In one implementation, the normal distribution similarity lik is calculated according to the following formula:

[0092]

[0093] Among them, t represents the current time, skness t , kurtosis t ,q 0.84,t ,q 0.98,t ,q 0.999,tThey represent the skewness, kurtosis, 84% quantile, 98% quantile, and 99.9% quantile of the observed data at the current time t, respectively. μ and σ represent the mean and standard deviation of the observed data at the current time t, respectively. t represents the statistical hypothesis verification value obtained based on the observed data at the current time t, α 1,t , α 2,t , α 3,t , α 4,t , α 5,t They respectively represent the weight coefficients of the statistical characteristic values ​​at the current time t.

[0094] Specifically, the observation data of a feature is usually not static. For example, if the observation data of a new user is collected next month or the original indicators of an old user change, the observation data set of the feature will change. Assuming that the observation data set is updated once a month, then with a monthly cycle, the observation data set of each month can be obtained.

[0095] In the embodiment of the present application, the weight coefficient at the current time t is calculated based on the weight coefficient at the previous time t-1. That is, the weight coefficient α at the current time t is 1,t , α 2,t , α 3,t , α 4,t , α 5,t It is constructed based on the changes in skewness, kurtosis, and quantile values ​​at time t and time t-1 to achieve intelligent dynamic fine-tuning of the weight coefficient and further improve the accuracy of the calculation.

[0096] In one implementation, the weight coefficient at the current time t is calculated according to the following formula:

[0097]

[0098]

[0099] Wherein, sgn(x) is a function for adjusting the sign, when x≥0, sgn(x)=1, when x<0, sgn(x)=﹣1.

[0100] From the above formula, we can see that the intelligent dynamic fine-tuning of the weight coefficient has two characteristics: First, the weight coefficient α at the current time t 1,t , α 2,t , α 3,t , α 4,t , α 5,t Will be affected by the weight coefficient α at the previous time t-1 1,t-1 , α 2,t-1 , α 3,t-1 , α 4,t-1 , α 5,t-1The influence of this effect decreases exponentially; secondly, the weight function (α i,t , i=1,…,5) has a limited range of change and cannot exceed the weight function of the previous cycle (α i,t-1 ,i=1,…,5) within a fluctuation range of plus or minus 30%, thus ensuring that the fluctuation of the weight coefficient is not too large. The combination of the above two points can better ensure the accurate characterization of the normal distribution similarity.

[0101] After calculating the normal distribution similarity lik, the observed data is judged whether it conforms to the normal distribution according to the normal distribution similarity and the preset threshold. For example, when 0≤lik<0.1, it means that the similarity is low, and it is necessary to add artificial expert experience to further judge whether the observed data conforms to the normal distribution; when 0.1≤lik<0.4, it means that the similarity is high and no expert experience is needed, but the shape of the feature data distribution histogram can also be used to check whether it conforms to the normal distribution; when lik≥0.4, it means that the similarity is very high and no auxiliary judgment is needed, and it can be directly determined that the observed data conforms to the normal distribution.

[0102] In summary, according to the method of the embodiment of the present application, the normal distribution of the observed data of any feature can be directly judged, and then the grading is determined based on the judgment result whether the grading is divided by the normal method. The grading results can also be displayed in the form of a bar chart, and the proportion of the number of samples at each grading level is fed back, and the median, mean and other statistical characteristics of each grading of the observed data are listed at the same time, so that decision makers can judge the information reflected by the data and make corresponding business deployments.

[0103] It should be additionally explained that no matter which grading method is used, a good grading system must ensure that its grading results present a bell-shaped distribution characteristic of a normal distribution with less at both ends and more in the middle. Therefore, when using the method of the embodiment of the present application to grade the data, it is also necessary to use feedback on the grading results to further correct the grading model. For example, the KS test method can still be used to input the grading level results as data samples into the KS test module, and compare the differences between the grading level data and the normal distribution. Similarly, whether the p value is greater than 0.3 is used as the dividing point. If the p value is too small, it means that the grading effect is not good. At this time, it is necessary to modify and adjust the grading model, mainly for the boundaries Q1,..., Q8 of the 9 gradings of extremely high, very high, relatively high, slightly high, medium, slightly low, relatively low, very low, and extremely low. The adjustment method can refer to the intelligent adjustment weight coefficient α described above. 1,t , α 2,t , α 3,t , α 4,t , α 5,tThe correction idea adopts similar correction logic, and determines whether each boundary is rising or falling based on the past dividing line and the size of the changes in the past data.

[0104] In summary, the embodiments of the present application have the following beneficial effects:

[0105] 1. We have created an analytical quantification system based on data indicator statistics, and built a set of intelligent data evaluation methods on this system to achieve intelligent management of high and low grade analysis of characteristic values, which can fully display the subtle changes in data;

[0106] 2. Combine artificial intelligence technology to build a new learning strategy that the AI ​​large language model can adapt to, build a training sample data set based on the numerical level of sample observation data, and design the corresponding information storage format to achieve artificial intelligence's reinforcement learning control of data analysis key points and characteristics, and demonstrate the establishment of an electronic data analyst to help financial business managers achieve unmanned financial business management.

[0107] Figure 6 A schematic diagram of the structure of a training data construction device provided in an embodiment of the present application. Figure 6 As shown, the training data construction device 40 includes:

[0108] The feature data acquisition module 410 is used to acquire feature data of each sample in the sample set under multi-dimensional features;

[0109] An observation data determination module 420 is used to use the feature data of each sample under each dimensional feature as observation data for each dimensional feature in the multi-dimensional feature;

[0110] The dividing line construction module 430 is used to determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and to construct a dividing line for data classification for the observed data corresponding to each dimensional feature according to the determination result;

[0111] A numerical value determination module 440 is used to determine the numerical value of the characteristic data of each sample under each dimensional feature according to the dividing line of the data classification;

[0112] The training sample data set construction module 450 is used to construct a training sample data set for large language model learning according to the numerical values ​​of each sample under each dimensional feature and the labeling information of each sample.

[0113] In some embodiments, the demarcation line construction module 430 includes:

[0114] A statistical hypothesis verification value acquisition unit, used to obtain a statistical hypothesis verification value of whether the observed data corresponding to each dimensional feature conforms to a normal distribution based on a statistical hypothesis verification method;

[0115] A statistical characteristic calculation unit, used to calculate the mean, standard deviation and statistical characteristic values ​​of the observed data, wherein the statistical characteristic values ​​include skewness, kurtosis, and quantiles corresponding to the mean value increased by a set number of times the standard deviation;

[0116] a normal distribution similarity calculation unit, configured to calculate the normal distribution similarity according to the statistical characteristic value and the statistical hypothesis verification value if the statistical characteristic value meets the statistical characteristic judgment standard of the normal distribution;

[0117] A judging unit is used to judge whether the observed data conforms to the normal distribution according to the normal distribution similarity and a preset threshold.

[0118] In some embodiments, the normal distribution similarity calculation unit includes:

[0119] a deviation calculation subunit, used to calculate the deviation of each of the statistical characteristic values ​​according to the statistical characteristic judgment standard;

[0120] A sum calculation subunit, used for multiplying the deviation of each of the statistical characteristic values ​​by the respective weight coefficients, and summing the multiplication results to obtain a sum value;

[0121] The normal distribution similarity calculation subunit is used to calculate the product of the sum value and the statistical hypothesis verification value to obtain the normal distribution similarity.

[0122] In some embodiments, the normal distribution similarity lik is calculated according to the following formula:

[0123]

[0124] Among them, t represents the current time, skness t , kurtosis t ,q 0.84,t ,q 0.98,t ,q 0.999,t They represent the skewness, kurtosis, 84% quantile, 98% quantile, and 99.9% quantile of the observed data at the current time t, respectively. μ and σ represent the mean and standard deviation of the observed data at the current time t, respectively. t represents the statistical hypothesis verification value obtained based on the observed data at the current time t, α 1,t , α 2,t , α 3,t , α 4,t , α 5,tThey respectively represent the weight coefficients of the statistical characteristic values ​​at the current time t.

[0125] In some embodiments, the weight coefficient at the current time t is calculated based on the weight coefficient at the previous time t-1.

[0126] In some embodiments, the weight coefficient at the current time t is calculated according to the following formula:

[0127]

[0128] Wherein, sgn(x) is a function for adjusting the sign, when x≥0, sgn(x)=1, when x<0, sgn(x)=﹣1.

[0129] The training data construction device provided in the embodiment of the present application can be used to execute the technical solution of the training data construction method in the above embodiment. Its implementation principle and technical effect are similar and will not be repeated here.

[0130] It should be noted that it should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the demarcation line construction module 430 can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a processing element of the above device. The function of the above demarcation line construction module 430. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in the processor element or instructions in the form of software.

[0131] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 7 As shown, the electronic device may include: a transceiver 121 , a processor 122 , and a memory 123 .

[0132] The processor 122 executes the computer execution instructions stored in the memory, so that the processor 122 executes the scheme in the above embodiment. The processor 122 can be a general-purpose processor, including a central processing unit CPU, a network processor (network processor, NP), etc.; it can also be a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0133] The memory 123 is connected to the processor 122 via a system bus and completes communication between them. The memory 123 is used to store computer program instructions.

[0134] The transceiver 121 may be used to obtain tasks to be executed and configuration information of the tasks to be executed.

[0135] The system bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The transceiver is used to realize the communication between the database access device and other computers (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory.

[0136] The electronic device provided in the embodiment of the present application may be the terminal device of the above embodiment.

[0137] An embodiment of the present application also provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the technical solution of the training data construction method of the above embodiment.

[0138] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When at least one processor executes the computer program, the technical solution of the training data construction method in the above embodiment can be implemented.

[0139] In the process of implementation, the computer program product can be written in one or more programming languages ​​or a combination thereof to perform the computer program code of the present application, and the programming language includes an object-oriented programming language, such as Java, Smalltalk, C++, and also includes a conventional procedural programming language, such as "C" language or similar programming language. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).

[0140] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A training data construction method, characterized in that: include: Obtain feature data of each sample in the sample set under multi-dimensional features; For each dimensional feature in the multi-dimensional features, taking the feature data of each sample under each dimensional feature as observation data; Determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and construct a data classification boundary for the observed data corresponding to each dimensional feature according to the determination result; According to the dividing line of the data classification, respectively determine the numerical value of the characteristic data of each sample under each dimensional feature; According to the numerical values ​​of each sample under each dimensional feature and the labeling information of each sample, a training sample data set for large language model learning is constructed.

2. The method according to claim 1, characterized in that The determining whether the observed data corresponding to each dimensional feature conforms to a normal distribution includes: Based on the statistical hypothesis verification method, obtain the statistical hypothesis verification value of whether the observed data corresponding to each dimensional feature conforms to the normal distribution; Calculate the mean, standard deviation and statistical characteristic values ​​of the observed data, wherein the statistical characteristic values ​​include skewness, kurtosis, and quantiles corresponding to the mean value increased by a set number of times the standard deviation; If the statistical characteristic value meets the statistical characteristic judgment standard of the normal distribution, then calculating the normal distribution similarity according to the statistical characteristic value and the statistical hypothesis verification value; Whether the observed data conforms to the normal distribution is determined according to the normal distribution similarity and a preset threshold.

3. The method according to claim 2, characterized in that The calculating the normal distribution similarity according to the statistical characteristic value and the statistical hypothesis check value comprises: Calculating the deviation of each of the statistical characteristic values ​​according to the statistical characteristic judgment standard; Multiplying the deviations of the statistical characteristic values ​​by their respective weight coefficients, and summing the multiplication results to obtain a sum value; The product of the sum value and the statistical hypothesis check value is calculated to obtain the normal distribution similarity.

4. The method according to claim 2 or 3, characterized in that: The normal distribution similarity lik is calculated according to the following formula: Among them, t represents the current time, skness t , kurtosis t ,q 0.84,t ,q 0.98,t ,q 0.999,t They represent the skewness, kurtosis, 84% quantile, 98% quantile, and 99.9% quantile of the observed data at the current time t, respectively. μ and σ represent the mean and standard deviation of the observed data at the current time t, respectively. t represents the statistical hypothesis verification value obtained based on the observed data at the current time t, α 1,t , α 2,t , α 3,t , α 4,t , α 5,t They respectively represent the weight coefficients of the statistical characteristic values ​​at the current time t.

5. The method according to claim 4, characterized in that The weight coefficient at the current time t is calculated based on the weight coefficient at the previous time t-1.

6. The method according to claim 4, characterized in that The weight coefficient at the current time t is calculated according to the following formula: Wherein, sgn(x) is a function for adjusting the sign, when x≥0, sgn(x)=1, when x<0, sgn(x)=﹣1.

7. A training data construction device, characterized in that: include: A feature data acquisition module is used to obtain feature data of each sample in the sample set under multi-dimensional features; An observation data determination module, configured to use, for each dimensional feature in the multi-dimensional features, feature data of each sample under each dimensional feature as observation data; A dividing line construction module is used to determine whether the observed data corresponding to each dimensional feature conforms to the normal distribution, and to construct a dividing line for data classification for the observed data corresponding to each dimensional feature according to the determination result; A numerical value determination module, used to determine the numerical value of the characteristic data of each sample under each dimensional feature according to the dividing line of the data classification; The training sample data set construction module is used to construct a training sample data set for large language model learning according to the numerical value of each sample under each dimensional feature and the labeling information of each sample.

8. The device according to claim 7, characterized in that The demarcation line construction module includes: A statistical hypothesis verification value acquisition unit, used to obtain a statistical hypothesis verification value of whether the observed data corresponding to each dimensional feature conforms to a normal distribution based on a statistical hypothesis verification method; A statistical characteristic calculation unit, used to calculate the mean, standard deviation and statistical characteristic values ​​of the observed data, wherein the statistical characteristic values ​​include skewness, kurtosis, and quantiles corresponding to the mean value increased by a set number of times the standard deviation; a normal distribution similarity calculation unit, configured to calculate the normal distribution similarity according to the statistical characteristic value and the statistical hypothesis verification value if the statistical characteristic value meets the statistical characteristic judgment standard of the normal distribution; A judging unit is used to judge whether the observed data conforms to the normal distribution according to the normal distribution similarity and a preset threshold.

9. The device according to claim 8, characterized in that The normal distribution similarity calculation unit comprises: a deviation calculation subunit, used to calculate the deviation of each of the statistical characteristic values ​​according to the statistical characteristic judgment standard; A sum calculation subunit, used for multiplying the deviation of each of the statistical characteristic values ​​by the respective weight coefficients, and summing the multiplication results to obtain a sum value; The normal distribution similarity calculation subunit is used to calculate the product of the sum value and the statistical hypothesis verification value to obtain the normal distribution similarity.

10. The device according to claim 8 or 9, characterized in that The normal distribution similarity lik is calculated according to the following formula: Among them, t represents the current time, skness t , kurtosis t ,q 0.84,t ,q 0.98,t ,q 0.999,t They represent the skewness, kurtosis, 84% quantile, 98% quantile, and 99.9% quantile of the observed data at the current time t, respectively. μ and σ represent the mean and standard deviation of the observed data at the current time t, respectively. t represents the statistical hypothesis verification value obtained based on the observed data at the current time t, α 1,t , α 2,t , α 3,t , α 4,t , α 5,t They respectively represent the weight coefficients of the statistical characteristic values ​​at the current time t.

11. The device according to claim 10, characterized in that The weight coefficient at the current time t is calculated based on the weight coefficient at the previous time t-1.

12. The device according to claim 10, characterized in that The weight coefficient at the current time t is calculated according to the following formula: Wherein, sgn(x) is a function for adjusting the sign, when x≥0, sgn(x)=1, when x<0, sgn(x)=﹣1.

13. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the training data construction method according to any one of claims 1 to 6.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the training data construction method according to any one of claims 1 to 6.

15. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the training data construction method according to any one of claims 1 to 6.

Images