Method and system for optimizing a data association model

Abstract

According to one aspect of the present application, a method for optimizing a data association model is provided, the data association model comprising a first sub-model and a second sub-model, the method comprising: determining a candidate sample output result based on at least candidate data associated with data in a second multi-source heterogeneous data set determined from a first multi-source heterogeneous data set using the first sub-model, determining optimization data of the first sub-model based on at least the candidate sample output result using the second sub-model; optimizing parameters of the first sub-model using the optimization data of the first sub-model to obtain an optimized first sub-model; constructing a fine-tuning data set based on at least user feedback, the user feedback being used to correct an association result generated by the optimized first sub-model based on data to be associated; and performing a collaborative optimization of parameters of the optimized first sub-model and parameters of the second sub-model based on at least the fine-tuning data set to optimize the data association model.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically to methods and systems for optimizing data association models, computer-readable storage media for implementing the above methods, and computer program products. Background Technology

[0002] In the company's daily operations, establishing relationships between different datasets is a crucial aspect of data management and analysis. This helps improve work efficiency, optimize decision-making processes, and support subsequent statistical analysis. For example, linking merchant data with brand data ensures the accuracy of each merchant's brand information; linking merchant data with business district data determines the location of each merchant within their business district; and linking merchant data with enterprise graph data identifies the relationships between merchants and other businesses.

[0003] Traditional dataset association methods primarily rely on simple rule engines or model analysis. While these methods can establish relationships between different datasets to some extent, they have limitations. For example, rule engines mainly determine the association between datasets by comparing whether keywords overlap, thus requiring the manual collection and organization of large keyword databases, which consumes significant human and material resources. Similarly, model analysis depends on whether the model training data includes relevant domain corpora. Insufficient relevant domain corpora can lead to poor model training performance, affecting the final association analysis results. Summary of the Invention

[0004] To address or at least alleviate one or more of the above problems, the following technical solutions are provided.

[0005] According to a first aspect of this application, a method for optimizing a data association model is provided, the data association model including a first sub-model and a second sub-model, the method comprising the following steps: determining candidate sample output results using the first sub-model based at least on candidate datasets associated with data in a second multi-source heterogeneous dataset determined from a first multi-source heterogeneous dataset; determining optimized data for the first sub-model using the second sub-model based at least on the candidate sample output results; optimizing the parameters of the first sub-model using the optimized data to obtain an optimized first sub-model; constructing a fine-tuning dataset based at least on user feedback, the user feedback being used to correct the association results generated by the optimized first sub-model based on the data to be associated; and co-optimizing the parameters of the optimized first sub-model and the parameters of the second sub-model based at least on the fine-tuning dataset to optimize the data association model.

[0006] According to an embodiment of this application, the method for optimizing a data association model has a smaller number of parameters in the first sub-model than in the second sub-model.

[0007] According to one embodiment or any of the above embodiments of the present application, the method for optimizing a data association model, wherein the first multi-source heterogeneous dataset and the second multi-source heterogeneous dataset are constructed respectively by: preprocessing data in the database into tuple-form data; generating a feature vector database based on the feature vectors of the tuple-form data; using a retrieval enhancement generation tool to retrieve and determine feature vectors that match the feature vectors of the user input question; and constructing the first multi-source heterogeneous dataset and the second multi-source heterogeneous dataset respectively based on the user input question and the tuple-form data corresponding to the feature vectors that match the feature vectors of the user input question.

[0008] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein determining candidate sample output results using the first sub-model based at least on candidate datasets associated with data in a second multi-source heterogeneous dataset determined from a first multi-source heterogeneous dataset includes: determining the similarity between the data in the second multi-source heterogeneous dataset and each candidate data in the candidate dataset using the first sub-model; and dividing the candidate dataset into candidate positive sample data and candidate negative sample data based on a comparison of the similarity with a similarity threshold to obtain the candidate sample output results.

[0009] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein determining the optimized data of the first sub-model using the second sub-model based at least on the candidate sample output results includes: generating formatted sample data based at least on the candidate sample output results; inputting the formatted sample data, the data in the second multi-source heterogeneous dataset, and the candidate dataset into the second sub-model to generate the optimized data of the first sub-model, wherein the optimized data of the first sub-model includes corrected positive sample data, corrected negative sample data, and candidate data in the candidate dataset that is most associated with the data in the second multi-source heterogeneous dataset.

[0010] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein optimizing the parameters of the first sub-model using the optimization data of the first sub-model to obtain an optimized first sub-model includes: determining an optimization loss function with the optimization objectives of maximizing the difference between the corrected positive sample data and the corrected negative sample data and maximizing the difference between the most associated candidate data and the candidate data in the candidate dataset other than the most associated candidate data; adjusting the parameters of the first sub-model using the optimization loss function; and obtaining the optimized first sub-model based on the adjusted parameters of the first sub-model.

[0011] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein constructing a fine-tuning dataset based at least on user feedback includes: preprocessing the user feedback into user feedback in tuple form, the user feedback in tuple form including at least two associated entities and entity relationships between the at least two associated entities; determining the entity attributes of the at least two associated entities and the similarity between the feature vectors of the at least two associated entities; generating question-answer data pairs based on the at least two associated entities and the entity relationships between the at least two associated entities; and constructing the fine-tuning dataset based on the question-answer data pairs, the entity attributes of the at least two associated entities, and the similarity between the feature vectors of the at least two associated entities.

[0012] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein co-optimizing the parameters of the optimized first sub-model and the parameters of the second sub-model based at least on the fine-tuning dataset includes: constructing fine-tuning data for the first sub-model and fine-tuning data for the second sub-model based at least on the fine-tuning dataset; determining a first fine-tuning loss function based on the fine-tuning data of the first sub-model and determining a second fine-tuning loss function based on the fine-tuning data of the second sub-model; and iteratively optimizing the parameters of the optimized first sub-model and the parameters of the second sub-model using the first fine-tuning loss function and the second fine-tuning loss function, respectively, until the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model is less than a difference threshold.

[0013] According to one embodiment or any of the above embodiments of the present application, the method for optimizing a data association model, wherein the first fine-tuning loss function is used at least to minimize the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model, and the second fine-tuning loss function is used at least to minimize the difference between the output distribution of the second sub-model and the output distribution of the optimized first sub-model.

[0014] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein the method further comprises: before and after co-optimizing the parameters of the optimized first sub-model and the parameters of the second sub-model based at least on the fine-tuning dataset, respectively obtaining a first association result output by the data association model before co-optimization and a second association result output by the data association model after co-optimization; determining whether the first association result and the second association result are consistent; and in response to determining that the first association result and the second association result are inconsistent, updating the fine-tuning dataset based at least on the second association result.

[0015] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein the method further includes: constructing a user feedback vector database based at least on the user feedback; determining, during online interaction with the user, candidate association datasets associated with user-inputted data to be associated from the user feedback vector database; determining, using a second sub-model, the most associated data with the user-inputted data to be associated from the candidate association datasets; and selectively returning the associated data to the user or determining, using the optimized first sub-model, the most associated data with the user-inputted data to be associated based on the similarity between the user-inputted data to be associated and the associated data.

[0016] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein selectively returning the associated data to the user or determining the most associated data with the user-inputted data based on the similarity between the user-inputted data to be associated and the associated data includes: returning the associated data to the user in response to the similarity being greater than or equal to a similarity threshold; or determining the most associated data with the user-inputted data with the user-inputted data using the optimized first sub-model in response to the similarity being less than a similarity threshold.

[0017] The method for optimizing a data association model according to one embodiment or any of the above embodiments of this application, wherein the data association model is used to determine the association between at least two types of data: data in a merchant database, data in a brand database, data in a business district database, and data in a business graph database.

[0018] According to a second aspect of this application, a system for optimizing a data association model is provided, the data association model including a first sub-model and a second sub-model, the system comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, the execution of the computer program causing the following operations: determining candidate sample output results using the first sub-model based at least on candidate datasets associated with data in a second multi-source heterogeneous dataset determined from a first multi-source heterogeneous dataset; determining optimized data for the first sub-model using the second sub-model based at least on the candidate sample output results; optimizing the parameters of the first sub-model using the optimized data to obtain an optimized first sub-model; constructing a fine-tuning dataset based at least on user feedback, the user feedback being used to correct association results generated by the optimized first sub-model based on data to be associated; and co-optimizing the parameters of the optimized first sub-model and the parameters of the second sub-model based at least on the fine-tuning dataset to optimize the data association model.

[0019] According to a third aspect of this application, a computer-readable storage medium is provided, comprising instructions that, when executed, perform the steps of the method for optimizing a data association model according to a first aspect of this application.

[0020] According to a fourth aspect of this application, a computer program product is provided, the computer program product including instructions that, when executed by a processor, implement the steps of the method for optimizing a data association model according to a first aspect of this application.

[0021] The scheme for optimizing a data association model according to one or more embodiments of this application determines the optimization data of the first sub-model by utilizing the output results of the second sub-model at least based on the candidate samples, and optimizes the parameters of the first sub-model using the optimization data of the first sub-model to obtain an optimized first sub-model. This allows the optimization of the parameters of the first sub-model with the help of the knowledge domain of the second sub-model, thereby improving the association analysis results of the data association model in various association analysis scenarios. By using a fine-tuning dataset constructed based on user feedback to collaboratively optimize the parameters of the optimized first sub-model and the parameters of the second sub-model, this collaborative optimization method can combine the actual needs and feedback of users, enabling the model to better adapt to different application scenarios and user needs, while achieving good model optimization results with less training data and a shorter training time. Attached Figure Description

[0022] The above and / or other aspects and advantages of this application will become clearer and more readily understood from the following description taken in conjunction with the accompanying drawings, in which the same or similar elements are denoted by the same reference numerals. The drawings include: Figure 1 A flowchart of a method for optimizing a data association model according to one or more embodiments of this application is shown.

[0023] Figure 2 A flowchart of a method for optimizing a data association model according to one or more embodiments of this application is shown.

[0024] Figure 3 A block diagram of a system for optimizing a data association model according to one or more embodiments of this application is shown. Detailed Implementation

[0025] The present application will now be described more fully with reference to the accompanying drawings, which illustrate exemplary embodiments thereof. However, the present application may be implemented in various forms and should not be construed as being limited to the embodiments given herein. The foregoing embodiments are intended to make the disclosure herein complete and thorough, so as to more fully convey the scope of protection of the present application to those skilled in the art.

[0026] In this specification, terms such as “comprising” and “including” indicate that, in addition to having the units and steps that are directly and explicitly stated in the specification and claims, the technical solution of this application does not exclude the presence of other units and steps that are not directly or explicitly stated.

[0027] Unless otherwise specified, terms such as “first” and “second” do not indicate the order of units in terms of time, space, size, etc., but are merely used to distinguish between units.

[0028] In the context of this application, the optimized data association model can be used to determine the association between multi-source heterogeneous data, which refers to data from different sources and with different data structures (e.g., data fields, text attributes, etc.). For example, multi-source heterogeneous data may include merchant data from registered merchants, brand data from brand websites, business district data from map applications, and enterprise maps from the State Administration for Industry and Commerce. In one embodiment, the optimized data association model can be used to determine the association between at least two types of data: data in a merchant database, data in a brand database, data in a business district database, and data in an enterprise map database, for example, by associating and matching data in the merchant database with corresponding data in the brand database.

[0029] In the following, various exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings.

[0030] Figure 1 A flowchart of a method for optimizing a data association model according to one or more embodiments of this application is shown, wherein the data association model includes a first sub-model and a second sub-model.

[0031] As Figure 1 shown in the figure, in step S101, a candidate sample output result is determined by using a first sub-model at least based on a candidate data set determined from a first multi-source heterogeneous data set and associated with data in a second multi-source heterogeneous data set, and optimization data of the first sub-model is determined by using a second sub-model at least based on the candidate sample output result.

[0032] In one embodiment, data in a database can be preprocessed into data in the form of multi-tuples, a feature vector database is generated based on feature vectors of the data in the form of multi-tuples, a retrieval enhancement generation tool is used to retrieve in the feature vector database to determine a feature vector matching the feature vector of a user input question, and a first multi-source heterogeneous data set and a second multi-source heterogeneous data set are respectively constructed based on the user input question and the data in the form of multi-tuples corresponding to the feature vector matching the feature vector of the user input question. For example, a combination of the user input question and the data in the form of multi-tuples corresponding to the feature vector matching the feature vector of the user input question can be constructed into a multi-source heterogeneous data set.

[0033] Exemplarily, the data in the form of multi-tuples can be data in the form of triples. Taking brand data as an example, the data in the brand data database can include brand ID, brand Chinese name, brand English name, brand abbreviation, brand-owned company, brand origin, brand score, brand classification, brand icon, brand description, etc. The data in the brand data database can be preprocessed into data in the form of triples, such as triples in the form (Ent, Rel, Attr), where Ent represents the entity name, such as brand ID; Rel represents the entity relationship, such as "brand Chinese name", "brand English name", etc.; Attr represents the entity attribute value, such as "Kentucky Fried Chicken", "KFC", etc. It should be noted that the same Ent can have multiple triples, and the same Ent and Rel can have multiple Attrs. For example, the brand English name of "Kentucky Fried Chicken" can be "KFC" and "Kentucky Fried Chicken". Exemplarily, the data in the form of triples can be transformed into feature vectors through a vectorization tool or model to generate a feature vector database.

[0034] In one embodiment, data T i in a second multi-source heterogeneous data set S i can be randomly selected (for example, an entity triple), and based on, such as the text edit distance, text similarity, etc. between data T i and data in a first multi-source heterogeneous data set S j , a candidate data set T j associated with data T i is retrieved from the first multi-source heterogeneous data set Sj ={T j1 ,T j2 ,…, T jn}, determine the data T in the second multi-source heterogeneous dataset S using the first sub-model i in i and the candidate dataset T j ={T j1 , T j2 ,…, T jn}, and based on the comparison between the similarity and the similarity threshold, divide the candidate dataset T j ={T j1 , T j2 ,…, T jn} into candidate positive sample data and candidate negative sample data to obtain the candidate sample output result. For example, the similarity threshold can be set to 0.8. The candidate dataset T j ={T j1 , T j2 ,…, T jn} can be sorted according to the similarity of each candidate data. The candidate dataset T j ={T j1 , T j2 ,…, T jn} corresponding to the candidate data with similarity greater than or equal to the similarity threshold is divided into candidate positive sample data, and the candidate dataset T j ={T j1 , T j2 ,…, T jn} corresponding to the candidate data with similarity less than the similarity threshold is divided into candidate negative sample data. The candidate positive sample data and the candidate negative sample data are used as the candidate sample output result.

[0035] In one embodiment, format sample data can be generated at least based on the candidate sample output result. For example, the generated format sample data can be "The data in the second multi-source heterogeneous dataset S i is T i , and T j ={T j1 ,T j2 ,…, T jn} in the candidate dataset T j1 , T j2 are candidate positive sample data, and T jn is candidate negative sample data". In one embodiment, the format sample data, the data T in the second multi-source heterogeneous dataset S i and the candidate dataset T i ={T j ={T j1 , Tj2 ,…, T jn The input is fed into the second sub-model to generate optimized data for the first sub-model. This optimized data may include corrected positive sample data and corrected negative sample data {T}. + , T -} and candidate dataset T j ={T j1 , T j2 ,…, T jn In the second multi-source heterogeneous dataset S i Data T in i Most relevant candidate data T m .

[0036] In one embodiment, the number of parameters in the first sub-model can be less than the number of parameters in the second sub-model. For example, the number of parameters in the first sub-model can be less than 600 million, while the number of parameters in the second sub-model can be greater than 32 billion. By utilizing the second sub-model to determine the optimized data of the first sub-model based at least on the candidate sample output results, and then using the optimized data of the first sub-model to optimize the parameters of the first sub-model to obtain an optimized first sub-model, the method of teaching the first sub-model with the second sub-model can fully combine the generalization ability advantage of the second sub-model with the scenario adaptation and inference latency advantage of the first sub-model, thereby improving the association analysis results of the data association model in various association analysis scenarios.

[0037] In step S103, the parameters of the first sub-model are optimized using the optimization data of the first sub-model to obtain an optimized first sub-model.

[0038] In one embodiment, as described above, the optimization data for the first sub-model may include corrected positive sample data and corrected negative sample data {T}. + , T -} and candidate dataset T j ={T j1 , T j2 ,…, T jn In the second multi-source heterogeneous dataset S i Data T in i Most relevant candidate data T m In step S103, the optimization objectives can be determined by maximizing the difference between the corrected positive sample data and the corrected negative sample data, and maximizing the difference between the most relevant candidate data and the candidate data in the candidate dataset other than the most relevant candidate data. The optimization loss function is then used to adjust the parameters of the first sub-model, and the optimized first sub-model is obtained based on the adjusted parameters of the first sub-model.

[0039] As an example, the optimization loss function can be constructed using the following formula (1). And the parameters of the first sub-model are adjusted by the following formula (2). : Formula (1) Formula (2) in, express and Similarity between them S represents the second multi-source heterogeneous dataset. i The data in This indicates corrected positive sample data. This represents the corrected mean of the positive sample data. This represents the corrected negative sample data. This represents the corrected mean of the negative sample data. This indicates that, in the candidate dataset, besides the most relevant candidate data T, there is... m Other candidate data, This indicates that, in the candidate dataset, besides the most relevant candidate data T, there is... m The mean of the candidate data other than and Indicates the weights assigned. During the optimization of the parameters of the first sub-model using the optimization data from the first sub-model, dynamic adjustments can be made. and The size of the variable is used to dynamically adjust the weights of the differences between the corrected positive sample data and the corrected negative sample data, as well as the differences between the most relevant candidate data and the candidate data in the candidate dataset other than the most relevant candidate data.

[0040] In step S105, a fine-tuning dataset is constructed based at least on user feedback, which is used to correct the association results generated by the optimized first sub-model based on the data to be associated.

[0041] In one embodiment, a user can input the data to be associated into an optimized first sub-model via human-computer interaction. The optimized first sub-model can then return the association result generated based on the data to be associated to the user. If the user determines that the association result is incorrect, the user can provide feedback via human-computer interaction to correct the association result generated by the optimized first sub-model based on the data to be associated. Taking merchant data as an example, the user can input the merchant data into the optimized first sub-model via human-computer interaction or by uploading a file. The optimized first sub-model can convert the merchant data into a feature vector, calculate the similarity between this feature vector and the feature vector in the feature vector database generated from the feature vector of the brand data in tuple form as described above, and return the brand data in tuple form corresponding to the feature vector with the highest similarity as the association result to the user. For example, a user can input the data to be associated, "What is the brand of Wenzhou Sam's Club Co., Ltd.", into the optimized first sub-model through human-computer interaction. The optimized first sub-model can return the association result "The brand of Wenzhou Sam's Club Co., Ltd. is KFC" generated based on the data to be associated to the user. The user can further provide feedback "The brand of Wenzhou Sam's Club Co., Ltd. is Sam" through human-computer interaction to correct the association result "The brand of Wenzhou Sam's Club Co., Ltd. is KFC" generated by the optimized first sub-model.

[0042] Optionally, in step S105, user feedback can be preprocessed into a tuple-based user feedback, which includes at least two related entities and the entity relationships between them. The similarity (e.g., similar or dissimilar) between the entity attributes of the at least two related entities and their feature vectors can be determined. Question-answer pairs are generated based on the at least two related entities and their entity relationships. A fine-tuning dataset is constructed based on the question-answer pairs, the entity attributes of the at least two related entities, and the similarity between their feature vectors. For example, taking the user feedback "The brand of Wenzhou Sam's Club is Sam" as an example, the second sub-model can preprocess this user feedback into a tuple-based user feedback according to the triple template (Ent1, Rel, Ent2), i.e., (Wenzhou Sam's Club, Chinese brand name, Sam), where Rel represents the Chinese brand name, which is the entity relationship predicted by the second sub-model based on the keyword "brand" in the user feedback, and Ent1 and Ent2 represent entity names, such as merchant name and brand name. For example, user feedback in the form of triplets (Wenzhou Sam's Club Co., Ltd., brand name in Chinese, Sam's) can be processed into a fine-tuned dataset that the model can recognize, such as: Query: What is the Chinese brand name of Wenzhou Sam's Club Co., Ltd. Context: Attributes of Wenzhou Sam's Club Co., Ltd., such as address, operating hours, and business scope.

[0043] Result: Sam.

[0044] CoT: Sam's Club's attributes, such as the brand's Chinese name, English name, and brand introduction; the similarity calculation results of the feature vectors between Wenzhou Sam's Club Co., Ltd. and Sam's Club (e.g., similar or dissimilar); and chat logs (if any) provided by users through the human-computer interaction interface.

[0045] In step S107, the parameters of the first sub-model and the second sub-model are optimized in a coordinated manner based on at least the fine-tuned dataset to optimize the data association model.

[0046] Optionally, in step S107, fine-tuning data for the first sub-model and fine-tuning data for the second sub-model can be constructed at least based on the fine-tuning dataset. A first fine-tuning loss function is determined based on the fine-tuning data of the first sub-model, and a second fine-tuning loss function is determined based on the fine-tuning data of the second sub-model. The parameters of the optimized first sub-model and the parameters of the second sub-model are iteratively optimized using the first and second fine-tuning loss functions respectively until the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model is less than a difference threshold. For example, the difference threshold can be set to 0.1. In one embodiment, the first fine-tuning loss function is used at least to minimize the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model, and the second fine-tuning loss function is used at least to minimize the difference between the output distribution of the second sub-model and the output distribution of the optimized first sub-model. By constructing fine-tuning data for the first sub-model and fine-tuning data for the second sub-model based on at least the fine-tuning dataset, and by using the first and second fine-tuning loss functions to collaboratively optimize the parameters of the first and second sub-models respectively, it can be ensured that both the first and second sub-models are optimized to move closer to the actual distribution of business scenarios reported by users. Compared with traditional models, this can achieve better user customization results.

[0047] In one embodiment, the fine-tuning data for the first sub-model may include positive sample data and negative sample data, wherein the positive sample data may include user feedback and candidate positive sample data from the candidate dataset as described above, and the negative sample data may include data pairs randomly obtained from two multi-source heterogeneous datasets. In one embodiment, the fine-tuning data for the second sub-model may include positive sample data and negative sample data, wherein the positive sample data may include positive sample data constructed based on user feedback (e.g., using Ent1 as positive sample data for Ent2 and Ent2 as positive sample data for Ent1) and candidate positive sample data from the candidate dataset as described above, and the negative sample data may include candidate negative sample data from the candidate dataset as described above. In one embodiment, the fine-tuning data for the second sub-model may also include general data (e.g., multiple-choice question data, true / false question data, calculation question data, etc.) to prevent the degradation of the general capabilities of the second sub-model.

[0048] As an example, the first fine-tuning loss function can be constructed using the following formulas (3)-(4). And the parameters of the first sub-model are adjusted by the following formula (5). : Formula (3) Formula (4) Formula (5) in, You can refer to formula (1) above to calculate. express and Similarity between them S represents the second multi-source heterogeneous dataset. i The data in This represents the positive sample data in the fine-tuning data of the first sub-model. This represents the mean of the positive sample data in the fine-tuning data of the first sub-model. This represents the negative sample data in the fine-tuning data of the first sub-model. This represents the mean of the negative sample data in the fine-tuning data of the first sub-model. This represents the positive sample data in the fine-tuning data of the second sub-model. This represents the mean of the positive sample data in the fine-tuning data of the second sub-model. This represents the negative sample data in the fine-tuning data of the second sub-model. This represents the mean of the negative sample data in the fine-tuning data of the second sub-model.

[0049] As an example, the second fine-tuning loss function can be constructed using the following formulas (6)-(7). And the parameters of the second sub-model are adjusted by the following formula (8). : Formula (6) Formula (7) Formula (8) in, You can refer to the above formula (1) to calculate. In formula (1), each variable is changed from the original first sub-model to the second sub-model, while the rest remain unchanged. express and Similarity between them S represents the second multi-source heterogeneous dataset. i The data in This represents the positive sample data in the fine-tuning data of the second sub-model. This represents the mean of the positive sample data in the fine-tuning data of the second sub-model. This represents the negative sample data in the fine-tuning data of the second sub-model. This represents the mean of the negative sample data in the fine-tuning data of the second sub-model. This represents the positive sample data in the fine-tuning data of the first sub-model. This represents the mean of the positive sample data in the fine-tuning data of the first sub-model. This represents the negative sample data in the fine-tuning data of the first sub-model. This represents the mean of the negative sample data in the fine-tuning data of the first sub-model.

[0050] In one embodiment, the method for optimizing a data association model may further include obtaining a first association result output by the data association model before co-optimization and a second association result output by the data association model after co-optimization, before and after co-optimizing the parameters of the first sub-model and the second sub-model; determining whether the first association result and the second association result are consistent; and updating the fine-tuning dataset at least based on the second association result if the first association result and the second association result are inconsistent, for example, by adding the second association result to the fine-tuning dataset. In one embodiment, when the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model is less than a difference threshold, optimization can be stopped and user feedback can be cleared, thereby solidifying the model knowledge of the user feedback content and alleviating the problem of excessively long model context caused by too much user feedback content.

[0051] The method for optimizing a data association model according to one or more embodiments of this application determines the optimization data of a first sub-model by utilizing a second sub-model at least based on the output results of candidate samples, and optimizes the parameters of the first sub-model using the optimization data of the first sub-model to obtain an optimized first sub-model. This allows the optimization of the parameters of the first sub-model with the help of the knowledge domain of the second sub-model, thereby improving the association analysis results of the data association model in various association analysis scenarios. By using a fine-tuning dataset constructed based on user feedback to collaboratively optimize the parameters of the optimized first sub-model and the parameters of the second sub-model, this collaborative optimization method can combine the actual needs and feedback of users, enabling the model to better adapt to different application scenarios and user needs, while achieving better model optimization results with less training data and a shorter training time.

[0052] Figure 2 A flowchart illustrating a method for optimizing a data association model according to one or more embodiments of this application is shown. Figure 2 The method shown can be used to update the model output online based on user feedback, enabling real-time online optimization of the data association model. This helps avoid misunderstandings or erroneous associations that may result from relying solely on the model itself, thereby improving the accuracy and relevance of the association results.

[0053] like Figure 2 As shown, in step S201, a user feedback vector database is constructed based at least on user feedback. Optionally, in step S201, user feedback can be preprocessed into tuple-based data, and the user feedback vector database is constructed based on the feature vectors of the tuple-based data and the feature vectors of user interaction history (e.g., user chat history). Through the user feedback mechanism, error cases encountered during model inference are recorded in the context dialogue, enabling online real-time optimization of the model output, reducing the model's dependence on knowledge updates and training iterations, and effectively improving the model's generalization ability.

[0054] In step S203, candidate association datasets associated with the user-inputted data to be associated are determined from the user feedback vector database during online interaction with the user. Optionally, in step S203, the association dataset can be based on data such as the user-inputted data S to be associated. i Text edit distance and text similarity between the data and the user feedback vector database are retrieved from the user feedback vector database to find the data S. i The associated candidate association dataset T={T1, T2,…, T n}

[0055] In step S205, the second sub-model is used to determine the most relevant associated data from the candidate association dataset that is most closely related to the user-input data to be associated. Optionally, in step S205, the second sub-model can be used to determine the user-input data S to be associated. i The candidate associated dataset T={T1, T2,…, T n The similarity between each data in the table is calculated, and the data with the highest similarity is identified as the most relevant associated data to the user-input data to be associated.

[0056] In step S207, it is determined whether the similarity between the user-inputted data to be associated and the associated data is greater than or equal to a similarity threshold. If the similarity between the user-inputted data to be associated and the associated data is greater than or equal to the similarity threshold, proceed to step S209; otherwise, proceed to step S211. For example, the similarity threshold can be set to 0.8.

[0057] In step S209, associated data is returned to the user. Optionally, the data in tuple form from the user feedback vector database corresponding to the associated data can be combined with the user's chat history to form a prompt word template, which is then called to return to the user via the second sub-model.

[0058] In step S211, the optimized first sub-model is used to determine the most relevant associated data to the user-input data to be associated. Taking merchant data as an example, the optimized first sub-model can convert the merchant data into feature vectors, and calculate the similarity between these feature vectors and the feature vectors in the feature vector database generated from the feature vectors of brand data in tuple form as described above. The tuple-form brand data corresponding to the feature vector with the highest similarity is returned to the user as the most relevant associated data to the user.

[0059] Figure 3 A block diagram of a system for optimizing a data association model according to one or more embodiments of this application is shown.

[0060] like Figure 3 As shown, the system 300 for optimizing a data association model includes a memory 310, a processor 320, and a computer program 330 stored in the memory 310 and executable on the processor 320. The processor 320 executes the computer program 330 to implement the method for optimizing a data association model according to one aspect of this application.

[0061] This application can also be implemented as a computer-readable storage medium including instructions that, when executed, perform a method for optimizing a data association model according to one aspect of this application.

[0062] Where applicable, the various embodiments provided in this application may be implemented using hardware, software, or a combination of hardware and software. Furthermore, where applicable, without departing from the scope of this application, the various hardware and / or software components described herein may be combined into composite components comprising software, hardware, and / or both. Where applicable, without departing from the scope of this application, the various hardware and / or software components described herein may be divided into sub-components comprising software, hardware, or both. Additionally, where applicable, it is contemplated that software components may be implemented as hardware components, and vice versa.

[0063] The software (such as program code and / or data) according to this application can be stored on one or more computer storage media. It is also contemplated that the software identified herein can be implemented using one or more networked and / or otherwise general-purpose or special-purpose computers and / or systems. Where applicable, the order of the various steps described herein can be changed, combined into compound steps, and / or divided into sub-steps to provide the features described herein.

[0064] The embodiments and examples presented herein are provided to best illustrate embodiments of this application and its particular applications, thereby enabling those skilled in the art to implement and use this application. However, those skilled in the art will understand that the above description and examples are provided for ease of illustration and example only. The descriptions presented are not intended to cover all aspects of this application or to limit this application to the precise forms disclosed.

Claims

1. A method for optimizing a data association model, characterized in that, The data association model includes a first sub-model and a second sub-model, and the method includes the following steps: The first sub-model is used to determine candidate sample output results based at least on candidate datasets that are associated with data in the second multi-source heterogeneous dataset and determined from the first multi-source heterogeneous dataset. The second sub-model is used to determine optimized data for the first sub-model based at least on the candidate sample output results. The parameters of the first sub-model are optimized using the optimization data of the first sub-model to obtain an optimized first sub-model; A fine-tuned dataset is constructed based at least on user feedback, which is used to correct the association results generated by the optimized first sub-model based on the data to be associated. as well as The parameters of the optimized first sub-model and the parameters of the second sub-model are optimized in a coordinated manner based on at least the fine-tuned dataset to optimize the data association model.

2. The method according to claim 1, wherein the number of parameters of the first sub-model is less than the number of parameters of the second sub-model.

3. The method according to claim 1, wherein the first multi-source heterogeneous dataset and the second multi-source heterogeneous dataset are constructed respectively in the following ways: Preprocess the data in the database into tuple format; A feature vector database is generated based on the feature vectors of the data in the form of tuples. The retrieval enhancement generation tool is used to retrieve feature vectors from the feature vector database to determine feature vectors that match the feature vectors of the user input question; and The first multi-source heterogeneous dataset and the second multi-source heterogeneous dataset are constructed based on the user input question and the tuple-form data corresponding to the feature vectors that match the feature vectors of the user input question.

4. The method of claim 1, wherein determining the candidate sample output using the first sub-model based at least on a candidate dataset associated with data in the second multi-source heterogeneous dataset, determined from the first multi-source heterogeneous dataset, comprises: The similarity between the data in the second multi-source heterogeneous dataset and each candidate data in the candidate dataset is determined using the first sub-model; as well as Based on the comparison between the similarity and the similarity threshold, the candidate dataset is divided into candidate positive sample data and candidate negative sample data to obtain the candidate sample output results.

5. The method according to claim 1, wherein determining the optimization data of the first sub-model using the second sub-model based at least on the candidate sample output results comprises: At least based on the output results of the candidate samples, formatted sample data is generated; The formatted sample data, the data in the second multi-source heterogeneous dataset, and the candidate dataset are input into the second sub-model to generate optimized data for the first sub-model. The optimized data for the first sub-model includes corrected positive sample data, corrected negative sample data, and candidate data in the candidate dataset that is most associated with the data in the second multi-source heterogeneous dataset.

6. The method according to claim 5, wherein optimizing the parameters of the first sub-model using the optimization data of the first sub-model to obtain an optimized first sub-model comprises: The optimization loss function is determined with the optimization objectives of maximizing the difference between the corrected positive sample data and the corrected negative sample data, and maximizing the difference between the most relevant candidate data and the candidate data in the candidate dataset other than the most relevant candidate data. The parameters of the first sub-model are adjusted using the optimized loss function; as well as The optimized first sub-model is obtained based on the adjusted parameters of the first sub-model.

7. The method of claim 1, wherein constructing the fine-tuning dataset based at least on user feedback comprises: The user feedback is preprocessed into a tuple-based user feedback, which includes at least two related entities and the entity relationships between the at least two related entities. Determine the similarity between the entity attributes of the at least two associated entities and the feature vectors of the at least two associated entities; Generate question-answer data pairs based on the at least two associated entities and the entity relationships between the at least two associated entities; as well as The fine-tuning dataset is constructed based on the question-and-answer data pairs, the entity attributes of the at least two related entities, and the similarity between the feature vectors of the at least two related entities.

8. The method of claim 1, wherein co-optimizing the parameters of the optimized first sub-model and the parameters of the second sub-model based at least on the fine-tuning dataset comprises: At least based on the fine-tuning dataset, fine-tuning data for the first sub-model and fine-tuning data for the second sub-model are constructed respectively; A first fine-tuning loss function is determined based on the fine-tuning data of the first sub-model, and a second fine-tuning loss function is determined based on the fine-tuning data of the second sub-model. as well as The parameters of the first sub-model and the second sub-model are iteratively optimized using the first fine-tuning loss function and the second fine-tuning loss function, respectively, until the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model is less than the difference threshold.

9. The method of claim 8, wherein the first fine-tuning loss function is used at least to minimize the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model, and the second fine-tuning loss function is used at least to minimize the difference between the output distribution of the second sub-model and the output distribution of the optimized first sub-model.

10. The method of claim 1, wherein the method further comprises: Before and after co-optimizing the parameters of the first sub-model and the second sub-model based at least on the fine-tuned dataset, obtain the first association result output by the data association model before co-optimization and the second association result output by the data association model after co-optimization, respectively. Determine whether the first association result and the second association result are consistent; as well as In response to determining that the first association result is inconsistent with the second association result, the fine-tuning dataset is updated based at least on the second association result.

11. The method according to claim 1, wherein the method further comprises: At least a user feedback vector database should be constructed based on the aforementioned user feedback; In online interactions with users, candidate association datasets associated with the user-input data to be associated are determined from the user feedback vector database; The second sub-model is used to determine the most relevant associated data from the candidate association dataset that is most closely related to the user-input data to be associated. as well as Based on the similarity between the user-inputted data to be associated and the associated data, the associated data is selectively returned to the user, or the optimized first sub-model is used to determine the associated data most closely related to the user-inputted data to be associated.

12. The method of claim 11, wherein selectively returning the associated data to the user or determining the most relevant associated data based on the similarity between the user-input data to be associated and the associated data comprises: In response to the similarity being greater than or equal to a similarity threshold, the associated data is returned to the user; or In response to the similarity being less than a similarity threshold, the optimized first sub-model is used to determine the associated data most closely related to the user-input data to be associated.

13. The method according to claim 1, wherein the data association model is used to determine the association between at least two types of data: data in the merchant database, data in the brand database, data in the business district database, and data in the enterprise graph database.

14. A system for optimizing a data association model, characterized in that, The data association model includes a first sub-model and a second sub-model, and the system includes: Memory; A processor coupled to the memory; and A computer program stored in the memory and running on the processor, the execution of which causes the following operations: The first sub-model is used to determine candidate sample output results based at least on candidate datasets that are associated with data in the second multi-source heterogeneous dataset and determined from the first multi-source heterogeneous dataset. The second sub-model is used to determine optimized data for the first sub-model based at least on the candidate sample output results. The parameters of the first sub-model are optimized using the optimization data of the first sub-model to obtain an optimized first sub-model; A fine-tuned dataset is constructed based at least on user feedback, which is used to correct association results generated by the optimized first sub-model based on the data to be associated; and The parameters of the optimized first sub-model and the parameters of the second sub-model are optimized in a coordinated manner based on at least the fine-tuned dataset to optimize the data association model.

15. The system of claim 14, wherein the number of parameters of the first sub-model is less than the number of parameters of the second sub-model.

16. The system of claim 14, wherein the first multi-source heterogeneous dataset and the second multi-source heterogeneous dataset are constructed respectively in the following ways: Preprocess the data in the database into tuple format; A feature vector database is generated based on the feature vectors of the data in the form of tuples. The retrieval enhancement generation tool is used to retrieve feature vectors from the feature vector database to determine feature vectors that match the feature vectors of the user input question; and The first multi-source heterogeneous dataset and the second multi-source heterogeneous dataset are constructed based on the user input question and the tuple-form data corresponding to the feature vectors that match the feature vectors of the user input question.

17. The system of claim 14, wherein the execution of the computer program causes the determination of candidate sample outputs using the first sub-model based at least on candidate datasets associated with data in the second multi-source heterogeneous dataset, determined from the first multi-source heterogeneous dataset: The similarity between the data in the second multi-source heterogeneous dataset and each candidate data in the candidate dataset is determined using the first sub-model; as well as Based on the comparison between the similarity and the similarity threshold, the candidate dataset is divided into candidate positive sample data and candidate negative sample data to obtain the candidate sample output results.

18. The system of claim 14, wherein the execution of the computer program causes the determination of optimized data for the first sub-model using the second sub-model, at least based on the candidate sample output, comprising: At least based on the output results of the candidate samples, formatted sample data is generated; The formatted sample data, the data in the second multi-source heterogeneous dataset, and the candidate dataset are input into the second sub-model to generate optimized data for the first sub-model. The optimized data for the first sub-model includes corrected positive sample data, corrected negative sample data, and candidate data in the candidate dataset that is most associated with the data in the second multi-source heterogeneous dataset.

19. The system of claim 18, wherein the execution of the computer program causes the parameters of the first sub-model to be optimized using the optimization data of the first sub-model to obtain an optimized first sub-model, comprising: The optimization loss function is determined with the optimization objectives of maximizing the difference between the corrected positive sample data and the corrected negative sample data, and maximizing the difference between the most relevant candidate data and the candidate data in the candidate dataset other than the most relevant candidate data. The parameters of the first sub-model are adjusted using the optimized loss function; as well as The optimized first sub-model is obtained based on the adjusted parameters of the first sub-model.

20. The system of claim 14, wherein the execution of the computer program results in the construction of a fine-tuning dataset based at least on user feedback, comprising: The user feedback is preprocessed into a tuple-based user feedback, which includes at least two related entities and the entity relationships between the at least two related entities. Determine the similarity between the entity attributes of the at least two associated entities and the feature vectors of the at least two associated entities; Generate question-answer data pairs based on the at least two associated entities and the entity relationships between the at least two associated entities; as well as The fine-tuning dataset is constructed based on the question-and-answer data pairs, the entity attributes of the at least two related entities, and the similarity between the feature vectors of the at least two related entities.

21. The system of claim 14, wherein the execution of the computer program results in the co-optimization of the parameters of the optimized first sub-model and the parameters of the second sub-model based at least on the fine-tuning dataset, comprising: At least based on the fine-tuning dataset, fine-tuning data for the first sub-model and fine-tuning data for the second sub-model are constructed respectively; A first fine-tuning loss function is determined based on the fine-tuning data of the first sub-model, and a second fine-tuning loss function is determined based on the fine-tuning data of the second sub-model. as well as The parameters of the first sub-model and the second sub-model are iteratively optimized using the first fine-tuning loss function and the second fine-tuning loss function, respectively, until the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model is less than the difference threshold.

22. The system of claim 21, wherein the first fine-tuning loss function is used at least to minimize the difference between the output distribution of the optimized first sub-model and the output distribution of the second sub-model, and the second fine-tuning loss function is used at least to minimize the difference between the output distribution of the second sub-model and the output distribution of the optimized first sub-model.

23. The system of claim 14, wherein the execution of the computer program causing the method to further include: Before and after co-optimizing the parameters of the first sub-model and the second sub-model based at least on the fine-tuned dataset, obtain the first association result output by the data association model before co-optimization and the second association result output by the data association model after co-optimization, respectively. Determine whether the first association result and the second association result are consistent; as well as In response to determining that the first association result is inconsistent with the second association result, the fine-tuning dataset is updated based at least on the second association result.

24. The system of claim 14, wherein the execution of the computer program causing the method to further include: At least a user feedback vector database should be constructed based on the aforementioned user feedback; In online interactions with users, candidate association datasets associated with the user-input data to be associated are determined from the user feedback vector database; The second sub-model is used to determine the most relevant associated data from the candidate association dataset that is most closely related to the user-input data to be associated. as well as Based on the similarity between the user-inputted data to be associated and the associated data, the associated data is selectively returned to the user, or the optimized first sub-model is used to determine the associated data most closely related to the user-inputted data to be associated.

25. The system of claim 24, wherein the execution of the computer program results in selectively returning the associated data to the user or determining, based on the similarity between the user-input data to be associated and the associated data, the most relevant associated data to the user-input data to be associated using the optimized first sub-model, comprising: In response to the similarity being greater than or equal to a similarity threshold, the associated data is returned to the user; or In response to the similarity being less than a similarity threshold, the optimized first sub-model is used to determine the associated data most closely related to the user-input data to be associated.

26. The system of claim 14, wherein the data association model is used to determine the association between at least two types of data: data in the merchant database, data in the brand database, data in the business district database, and data in the enterprise graph database.

27. A computer-readable storage medium, characterized in that, The computer storage medium includes instructions that, when executed, perform the method for optimizing a data association model according to any one of claims 1-13.

28. A computer program product, characterized in that, The computer program product includes instructions that, when executed by a processor, implement the method for optimizing a data association model according to any one of claims 1-13.

Images