Method, apparatus, electronic device, and program product for data attribution

Abstract

Embodiments of the present disclosure relate to methods, devices, electronic devices, and program products for data attribution. The method includes obtaining data in the form of a text description, and generating a parsing tree based on the data in the form of the text description and preset dimensions, wherein the parsing tree includes at least one of node attribution data and whole tree attribution data. The method also includes pruning the parsing tree based on a first parameter and a second parameter to obtain data in the form of a pruned multi-branch tree structure. The method also includes normalizing and serializing the pruned data in the form of a multi-branch tree structure to generate normalized serialized data, and the normalized serialized data is presented in the form of a text description. In addition, the method also includes generating an attribution summary for the data based on the normalized serialized data by a language model. Therefore, according to the solution of the embodiment of the present disclosure, data can be automatically attributed for users in massive data and presented in concise text, thereby improving the user experience.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more particularly to a method, apparatus, electronic device, and computer program product for data attribution. Background Art

[0002] With the advent of the big data era, data volumes are rapidly growing and becoming increasingly complex, making data analysis increasingly important across all industries. By efficiently processing and analyzing massive amounts of data, data analysis can help users deeply explore its value, optimize decision-making processes, and improve operational efficiency. Consequently, the importance of data analysis is constantly increasing.

[0003] Given the explosive growth in data volumes, extracting valuable information from such vast and complex data systems has become a key challenge in data analysis. Data attribution, a crucial component of the data analysis process, has become increasingly important in this context. Simply put, data attribution is the process of identifying the causes and influencing factors behind data results. Accurate data attribution enables a deeper understanding of the underlying logic, providing a solid and reliable basis for decision-making. Summary of the Invention

[0004] Embodiments of the present disclosure provide a method, apparatus, electronic device, and computer program product for data attribution.

[0005] According to a first aspect of the present disclosure, a method for data attribution is provided. The method includes obtaining data in the form of a textual description, wherein the data includes changes in the data within a date range. The method includes generating a parse tree based on the data in the textual description and preset dimensions, wherein the parse tree includes at least one of node attribution data and whole-tree attribution data, wherein the node attribution data indicates that the data is presented in a node-level nested form, each node in the node attribution data includes a key numerical description of the node, and the whole-tree attribution data is determined based on the node's decomposition path. The method also includes pruning the parse tree based on a first parameter and a second parameter to obtain data in the form of a pruned multi-tree structure, wherein the first parameter indicates whether to prune a node based on the contribution of the node in the multi-tree structure, and the second parameter indicates whether to prune a node based on the importance of the node in the multi-tree structure. The method also includes normalizing and serializing the pruned data in the form of a multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in the form of a textual description. In addition, the method also includes generating an attribution summary for the data based on the normalized serialized data by a language model.

[0006] According to a second aspect of the present disclosure, a device for data attribution is provided. The device includes a data acquisition module, which is configured to acquire data in the form of text description, wherein the data includes changes in the data within a date range. The device includes an analysis tree generation module, which is configured to generate an analysis tree based on the data in the form of text description and preset dimensions, wherein the analysis tree includes at least one of node attribution data and whole tree attribution data, the node attribution data indicating that the data is presented in the form of node hierarchical nesting, each node in the node attribution data includes a key numerical description of the node, and the whole tree attribution data is determined based on the node's disassembly path. The device also includes an analysis tree pruning module, which is configured to prune the analysis tree based on a first parameter and a second parameter to obtain data in the form of a pruned multi-branch tree structure, wherein the first parameter indicates whether to prune the node based on the contribution of the node in the multi-branch tree structure, and the second parameter indicates whether to prune the node based on the importance of the node in the multi-branch tree structure. The apparatus further includes a data generation module configured to normalize and serialize the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in the form of a text description. Furthermore, the apparatus further includes an attribution summary generation module configured to generate an attribution summary for the data based on the normalized serialized data using a language model.

[0007] According to a third aspect of the present disclosure, an electronic device is provided, comprising a processor and a memory coupled to the processor, wherein the memory has instructions stored therein, and when the instructions are executed by the processor, the electronic device executes the method according to the first aspect.

[0008] In a fourth aspect of the present disclosure, a computer program product is provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions that, when executed, cause a computer to perform the steps of the method of the first aspect of the present disclosure.

[0009] This summary is intended to introduce a selection of concepts in a simplified form that are further described below in the detailed description. It is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, wherein:

[0011] Figures 1A-1EA schematic diagram illustrating an example environment in which devices and / or methods according to embodiments of the present disclosure may be implemented;

[0012] Figure 2 A flowchart of a method for data attribution according to an embodiment of the present disclosure is shown;

[0013] Figure 3 A schematic diagram illustrating an example framework for data attribution according to an embodiment of the present disclosure is shown;

[0014] Figure 4 A schematic diagram illustrating another example system framework for data attribution according to an embodiment of the present disclosure is shown;

[0015] Figure 5A A schematic diagram illustrating an example structure of rule-based data attribution according to an embodiment of the present disclosure;

[0016] Figure 5B A schematic diagram showing how to convert a text description into a multi-branch tree structure according to an embodiment of the present disclosure

[0017] Figure 5C An example schematic diagram of determining the balance of direct contributions of nodes in a candidate path according to an embodiment of the present disclosure is shown;

[0018] Figure 5D A schematic diagram illustrating an example of performing importance clipping according to an embodiment of the present disclosure is shown;

[0019] Figure 6A A schematic diagram illustrating an example structure of model-based data attribution according to an embodiment of the present disclosure;

[0020] Figures 6B-6D A schematic diagram showing a form of training data generated for training according to an embodiment of the present disclosure;

[0021] Figure 6E A schematic diagram showing example prompts for training generated according to an embodiment of the present disclosure;

[0022] Figure 7 A block diagram illustrating an apparatus for data attribution according to some embodiments of the present disclosure; and

[0023] Figure 8 A block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0024] Throughout the drawings, the same or similar reference numbers denote the same or similar elements. DETAILED DESCRIPTION

[0025] It is understandable that before using the technical solutions disclosed in the various embodiments of this disclosure, the type, scope of use, usage scenarios, etc. of the personal information (such as voice) involved in this disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0026] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0027] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc. can refer to different or the same objects, unless explicitly stated otherwise. Other explicit and implicit definitions may also be included below.

[0028] As mentioned earlier, the importance of data analysis is growing, and the role of data attribution, a crucial component of data analysis, is also becoming increasingly prominent. For example, when a key indicator in the data changes, users need to analyze the cause of the change—in other words, they expect to identify the primary factors causing the change. For example, for key indicator A, you can break it down along dimension B into sub-indicators X and Y, and along dimension C into sub-indicators P and Q, and so on, to identify the primary factors influencing changes in key indicator A. This breakdown path is called a breakdown dimension. The depth and dimensions of the breakdown can be determined by the user based on their data attribution needs.

[0029] Currently, one approach to data attribution is for users to enumerate the factors contributing to changes in key metrics based on their own dimensions and analyze them one by one to identify the primary factor. However, this manual analysis is not only time-consuming and labor-intensive, but also prone to high error rates. To avoid manual errors, some technologies use rule-based approaches for automatic attribution, which can achieve accurate attribution. However, this approach is not suitable for data attribution in frequently evolving businesses due to its lack of flexibility and inability to adapt to changing business rules and structures. Other technologies use deep learning-based models to achieve automatic data attribution, but these model-based approaches are not suitable for large and continuously growing data volumes. This is because, due to current technological limitations, models often make errors when attributing data with extremely large data volumes.

[0030] To this end, at least in order to solve the above-mentioned problems and potential problems, an embodiment of the present disclosure proposes a solution for automatically attributing data for users. The solution first obtains data in the form of text description, wherein the data includes changes in the data within a date range, and based on the data in the form of text description and preset dimensions, generates an analysis tree including at least one of node attribution data and whole tree attribution data. Node attribution data is a data form presented in the form of nested nodes, and each node in the node attribution data includes a description of the key values of the node, while the whole tree attribution data can be determined based on the node's disassembly path. With the help of data in the analysis tree structure, the parent-child relationship and brother relationship between the data can be clearly represented, thereby enhancing the readability of data processing. Then, the analysis tree can be pruned according to a first parameter indicating the contribution of the node and a second parameter indicating the importance of the node to obtain data in the form of a pruned multi-branch tree structure, so that subsequent steps can focus on nodes that have a greater influence on the data, while also avoiding the situation where the subsequent model is not general enough for data processing. Then, by normalizing and serializing the pruned multi-tree structure data, the data is described in text form, which improves the efficiency of subsequent data processing by the language model. Finally, the language model generates an attribution summary based on the normalized and serialized text data.

[0031] Therefore, according to the embodiments of the present disclosure, by generating analysis tree-structured data and combining multiple processing steps (including pruning, normalization and serialization), the accuracy and pertinence of data attribution are improved. Furthermore, by utilizing language models to generate attribution summaries, key information is effectively extracted. This enables users to efficiently and accurately obtain data attribution results when faced with complex data, significantly improving their operational experience in data attribution.

[0032] The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. Figure 1A 1 is a schematic diagram of an example environment 100 in which devices and / or methods according to embodiments of the present disclosure may be implemented. Figure 1A As shown, text description 110 shows the details of changes in key indicator A over the time range, including the magnitude of the change and the rate of change. When key indicator A changes, users need to summarize the possible causes of the change to inform business decisions related to key indicator A. In other words, attribution requires outputting an interpretation of the primary factors that caused the change in key indicator A.

[0033] However, due to the varying roles played by users, some lack a deep understanding of the relevant business and are unable to accurately analyze the causes of changes in key indicator A. To help these users understand the causes of changes in business indicators and make attribution a common practice, we can automatically attribute data by combining rule-based and model-based processing. This combined approach ensures the stability of data attribution results.

[0034] refer to Figure 1A In order to deeply explore the reasons that affect the changes in key indicator A, a tree-structured data representation method can be used to deconstruct key indicator A. The reason is that the tree structure can display data in a hierarchical manner. This hierarchical structure can clearly show the parent-child relationship, brother relationship, etc. between the data, which is convenient for the subsequent processing steps to quickly locate the nodes with great influence and their related paths. For example, the text description 110 about key indicator A (data 120 in the form of a text description about the changes in key indicators) can be converted into a multi-branch tree structure 130, and then the next step of processing is performed on the data of the tree structure. It can be understood that key indicator A can be a root indicator or one of the sub-indicators under the root indicator.

[0035] In some embodiments, data in the form of text descriptions may be converted into an analysis tree according to preset dimensions. Figure 1B FIG2 shows a schematic diagram of an example disassembly path 100B according to an embodiment of the present disclosure. Figure 1BKey indicator A can be broken down into sub-indicators X and Y along dimension A. It can also be broken down into sub-indicators M and N along dimension B. Sub-indicators M and N can then be broken down further along dimension C into sub-indicators P and Q. Similarly, sub-indicators can be further broken down along other dimensions. The pre-defined dimensions here can be user-defined breakdown dimensions or breakdown paths. For example, taking key indicator A as order completion rate, there are multiple breakdown dimensions for order completion rate. For example, by breaking it down by industry category, the order completion rate can be divided into sub-indicators such as beauty and cosmetics order completion rate, watch and accessories order completion rate, and 3C digital appliance order completion rate. Furthermore, by breaking it down by business identity, the order completion rate can be broken down into sub-indicators such as cross-border order completion rate, next-day order completion rate, and instant retail order completion rate. Users can freely choose the breakdown dimensions (paths) based on their needs. For ease of description, this article will refer to (root) node and (key) indicator interchangeably.

[0036] In some embodiments, the generated analysis tree may be in at least one of a node attribution data format or a whole tree attribution data format. Figure 1C FIG2 shows a schematic diagram of an example node attribution data form 100C according to an embodiment of the present disclosure. Figure 1C Node attribution data is presented as a nested node hierarchy. Each node in the node attribution data includes a description of its key numerical values. Node attribution aims to identify the node (sub-metric) that contributes the most to the root metric within the current breakdown dimension. Contribution can be measured using contribution rate data.

[0037] Figure 1D FIG. 1 shows a schematic diagram of an example root node 100D according to an embodiment of the present disclosure. Figure 1D As shown, the root node includes the key numerical description of the sub-indicator, possible trends, possible abnormal points and other text descriptions, and Figure 1A The text description 110 shown is identical.

[0038] Figure 1E FIG. 1 is a schematic diagram illustrating an example whole tree attribution data form 100E according to an embodiment of the present disclosure. Figure 1E The whole tree attribution data includes a disassembly path consisting of a root node and multiple disassembly nodes, and can also include the disassembly of possible multiplication factors. Usually, there is only one level of disassembly of possible multiplication formulas. Figure 1BTo describe, for example, for the key indicator Q, its possible multiplication factor decomposition can be R*S*T, because Q=R*S*T. The whole tree attribution data is determined based on the node decomposition path. Whole tree attribution requires first determining which dimension decomposition method is most appropriate, and then determining the node with the largest contribution to the root indicator from the decomposition method according to the node attribution method. It can be understood that the contribution rate value is the proportion of the contribution made by a certain part to the overall change in the whole, that is, the proportion of the contribution made by the sub-indicator to the change of the root indicator in the whole.

[0039] In some embodiments, to ensure the accuracy of subsequent language model data processing, the parse tree can be adaptively pruned. Specifically, the textual description of the parse tree is converted into a standard multi-branch tree structure that can distinguish the importance of tree nodes at the same level. This allows for node pruning. This not only removes the impact of the data's text format on model calculations, but also allows for the selection of the most important sub-indicators, improving model calculation efficiency.

[0040] In some embodiments, for efficient data processing, whether to prune a node can be determined based on a first parameter indicating the contribution of the node in the multi-tree structure. This is because a node or sub-indicator with a too small contribution cannot be the main cause of the change in the root indicator and can therefore be pruned. In some embodiments, whether to prune a node can also be determined based on a second parameter indicating the importance of the node in the multi-tree structure. This is because when the multi-tree structure is deep, the model may make errors when processing related data. Therefore, in order to avoid unnecessary errors, some nodes can be pruned based on their importance.

[0041] Back to Figure 1A In some embodiments, to further avoid the impact of data format on model calculations, the data in the multi-tree structure 130 can be normalized and serialized so that it is presented as normalized and serialized text data 140. For example, the format description in the data can be replaced with a universal descriptor so that the subsequent model can focus on the specific logic rather than the format when attributing the data, which can improve the efficiency of the model output. In some embodiments, before obtaining the normalized and serialized text data 140, the data in the multi-tree structure 130 needs to be output as a text description.

[0042] In some embodiments, after the relevant data of key indicator A is processed according to the processing rules, the attribution summary 160 of key indicator A can be obtained with the help of language model 150. In some embodiments, the language model is a language model built based on deep learning technology, with a large number of parameters and powerful language processing capabilities.

[0043] Combine Figure 1B In some embodiments, when a user wishes to view a breakdown dimension for key indicator A, they can click on the corresponding node, such as clicking on the "Breakdown by Dimension B" control 131, to obtain the main factors causing changes in key indicator A according to that dimension, along with a series of detailed statistical indicator information 132. In some embodiments, when a user wishes to focus on the reasons for the overall change in indicator A, they can click on a button, which will display a description of the overall change in key indicator A and the most critical factors of change from a global perspective.

[0044] Therefore, combined Figures 1A-1E According to the solution provided by the embodiment of the present disclosure, the data is first structured by generating an analysis tree, and combined with processing steps such as clipping, normalization and serialization, the accuracy and pertinence of data attribution are significantly enhanced. In this process, the key information is efficiently refined by generating an attribution summary with the help of a language model. In this way, through the solution of the embodiment of the present disclosure, the results of data attribution can be obtained quickly and accurately, and the user experience in the data attribution operation process is comprehensively improved, allowing users to complete tasks more conveniently and efficiently in data attribution work, freeing themselves from the complicated data processing and focusing on making decisions based on the attribution results.

[0045] It should be understood that the architecture and functions in the example environments 100-100E are described for exemplary purposes only and do not imply any limitation on the scope of the present disclosure. The embodiments of the present disclosure may also be applied to other environments with different structures and / or functions.

[0046] The following will be combined Figures 2 to 8 The process according to the embodiment of the present disclosure is described in detail. For ease of understanding, the specific data mentioned in the following description are exemplary and are not intended to limit the scope of protection of the present disclosure. It is understood that the embodiments described below may also include additional actions not shown and / or may omit the actions shown, and the scope of the present disclosure is not limited in this respect.

[0047] Figure 2 A flowchart of a method 200 for data attribution according to an embodiment of the present disclosure is shown. The method 200 is schematically illustrated below, using a device for data attribution as an example. The method 200 includes blocks 202, 204, 206, 208, and 210.

[0048] At block 202, data in the form of a text description may be obtained, wherein the data includes how the data has changed within a date range. Figure 1AAs shown, text description 110 shows the change details, change amount, and change rate of key indicator A within the time range. When key indicator A changes, the user needs to summarize the possible reasons that affect the change of key indicator A to make decisions related to key indicator A.

[0049] At block 204, an analysis tree may be generated based on the data in the form of text description and preset dimensions, wherein the analysis tree includes at least one of node attribution data and whole tree attribution data, the node attribution data indicating that the data is presented in a node hierarchical nested form, each node in the node attribution data includes a key numerical description of the node, and the whole tree attribution data is determined based on the node decomposition path. Figure 1A In order to deeply explore the reasons that affect the changes in key indicator A, a tree-structured data representation method can be used to deconstruct key indicator A. In some embodiments, data in the form of text descriptions can be converted into an analysis tree based on preset dimensions. Figure 1B , the key indicator A can be decomposed into sub-indicators X and Y according to the A dimension, and can also be decomposed into sub-indicators M and N according to the B dimension. For the sub-indicators M and N, they can be decomposed downward into sub-indicators P and Q according to the C dimension. Similarly, the sub-indicators can continue to be decomposed layer by layer according to other dimensions. The preset dimensions here can be decomposition dimensions or decomposition paths pre-defined by the user. In some embodiments, the generated analysis tree can be in the form of node attribution data or in the form of whole tree attribution data. Combined Figure 1C Node attribution data is presented in the form of nested node levels. Each node in the node attribution data includes a key numerical description of the node. Node attribution needs to find the node (sub-indicator) with the largest contribution rate to the root indicator under the current decomposition dimension. Figure 1E The whole tree attribution data includes a disassembly path consisting of a root node and multiple disassembly nodes. The whole tree attribution data is determined based on the disassembly path of the node.

[0050] At box 206, the analysis tree can be pruned based on the first parameter and the second parameter to obtain data in the form of a pruned multi-branch tree structure, wherein the first parameter indicates whether to pruned the node based on the contribution of the node in the multi-branch tree structure, and the second parameter indicates whether to pruned the node based on the importance of the node in the multi-branch tree structure. In some embodiments, in order to ensure the accuracy of the subsequent language model processing of the data, the analysis tree can also be adaptively pruned. That is, the text description of the analysis tree is converted into a standard multi-branch tree structure that can distinguish the importance of tree nodes between the same levels to achieve node pruning. This can not only remove the influence of the text format of the data on the model calculation, but also filter out the most important sub-indicators to improve the efficiency of the model calculation. In some embodiments, in order to improve the efficiency of data processing, it can be determined whether to pruned a node based on the first parameter indicating the contribution of the node in the multi-branch tree structure. The reason is that a node or sub-indicator with a too small contribution cannot be the main cause of the change of the root indicator, so the node can be pruned. In some embodiments, whether to prune a node can also be determined based on a second parameter indicating the importance of the node in the multi-tree structure. This is because when the depth of the multi-tree structure is deep, the model may make mistakes when processing related data. Therefore, in order to avoid unnecessary errors, some nodes can be pruned based on the importance of the nodes.

[0051] At block 208, the pruned data in the multitree structure may be normalized and serialized to generate normalized serialized data, wherein the normalized serialized data is presented in a text description form. Figure 1A In some embodiments, in order to further avoid the influence of data format on model calculation, the data of the multi-tree structure 130 can be normalized and serialized so that it is presented as normalized and serialized text data 140. For example, the description with format in the data can be replaced with a universal descriptor so that the subsequent model can focus on the specific logic rather than the format in the process of attributing the data. This re-normalization and serialization method can improve the efficiency of model output. In some embodiments, before obtaining the normalized and serialized text data 140, the data of the multi-tree structure 130 needs to be output as a text description.

[0052] At block 210, a language model may generate an attribution summary for the data based on the normalized serialized data. Figure 1A In some embodiments, after the relevant data of key indicator A is processed according to the processing rules, the attribution summary 160 of key indicator A can be obtained with the help of language model 150. In some embodiments, the language model is a language model built based on deep learning technology, which has a large number of parameters and powerful language processing capabilities.

[0053] Therefore, according to the embodiments of the present disclosure, by generating analysis tree-structured data and combining multiple processing steps (including pruning, normalization and serialization), the accuracy and pertinence of data attribution are improved. Furthermore, by utilizing language models to generate attribution summaries, key information is effectively extracted. This enables users to efficiently and accurately obtain data attribution results when faced with complex data, significantly improving their operational experience in data attribution.

[0054] Figure 3 FIG. 3 is a schematic diagram of an example framework 300 for data attribution according to an embodiment of the present disclosure. Figure 3 As shown, the example framework 300 includes a data pre-processing step 310, a model prediction step 320, and a data post-processing step 330. Among them, the data pre-processing step 310 also includes an adaptive clipping step 312 and a normalization serialization step 313. The model prediction step 320 includes a model reasoning step 322. In addition, the data post-processing step 330 includes a normalization deserialization step 332. With the help of the data pre-processing step 310, a general data format suitable for model reasoning can be obtained, so that the model used for reasoning can focus on the logic of specific data attribution and improve the efficiency of model output. After obtaining the output result of the model reasoning, it is necessary to further perform a normalization deserialization step 332 on the result of the model output, so that the attribution summary of the model output can be presented in a user-readable form.

[0055] In some embodiments, after the original data undergoes adaptive cropping step 312 and normalized serialization step 313, it can be converted into a data format suitable for model inference step 322. After model inference step 322 and normalized deserialization 332, a readable and refined final data attribution result for the original data can be obtained. This method, combining data processing rules and model inference, not only ensures the stability of data attribution but also ensures that the model has good generalization output, thereby improving the user experience of data attribution.

[0056] Figure 4 A schematic diagram of another example system framework 400 for data attribution according to an embodiment of the present disclosure is shown. The example system framework 400 mainly includes two modules: data processing 410 and model construction 420. The data processing 410 module is responsible for optimizing the input of the model, including adaptive clipping 402, normalized serialization 403, and normalized deserialization 405. It can be understood that there is a sequential dependency between adaptive clipping 402, normalized serialization 403, and normalized deserialization 405. The model construction 420 is responsible for training the model for reasoning, including training data construction 422, model training 424, and model testing 426. Combined Figure 3, Figure 3 The steps of data pre-processing step 310 and data post-processing step 330 can be Figure 4 The steps performed by the data processing 410 module are shown. Figure 5A as well as Figure 6A The detailed steps that can be performed by the two modules of data processing 410 and model building 420 are further described. Figure 5A A schematic diagram illustrating an example structure 500A for rule-based data attribution according to an embodiment of the present disclosure is shown. Figure 6A A schematic diagram of an example structure 600A of model-based data attribution according to an embodiment of the present disclosure is shown. Figures 5A-6A To further describe the example system framework 400. For ease of description, the following model may refer to a large language model based on deep learning.

[0057] refer to Figure 4 In the data processing module 410, adaptive cropping 402 may be performed on the original data 401. In some embodiments, the original data 401 may be as follows: Figures 1B-1E The data in the form of node attribution shown or the data in the form of whole tree attribution. Figure 5A When performing adaptive pruning 402 , data structuring 501 can be performed first. Specifically, raw data 401 can be converted from a text description into a standard multi-tree structure. Data structuring 501 can refine the input content used for model inference and remove the impact of text formatting on model inference. In some embodiments, the levels of the multi-tree are distinguished by the indentation of the text description. Sub-indicators with less indentation represent higher levels in the multi-tree structure. Figure 5B FIG5 shows an example 500B schematic diagram of converting a text description into a multi-branch tree structure according to an embodiment of the present disclosure. Figure 5B , the multi-branch tree structure includes multiple nodes, each node corresponds to a sub-index. Figure 5B The method of converting the text description into a standard multi-branch tree structure can clearly distinguish the importance of tree nodes at the same level.

[0058] Continue to refer Figure 4In order to improve the speed of model reasoning, the input required for model reasoning can be simplified by contribution clipping 502. Contribution refers to the proportion of the contribution of each child node in the multi-branch tree structure to the change of the root node in the whole (which can be measured by the contribution rate), so contribution clipping 502 is for the child nodes in the multi-branch tree structure. Given that the child nodes with small contribution are not the main influencing factors of the root node change, the input of model reasoning can be simplified by clipping the child nodes with low contribution. In some embodiments, whether to clip a node can be determined by comparing the relationship between the contribution of each node and the first parameter used to indicate the contribution. Furthermore, given that the multi-branch tree structure can be divided into node attribution data form and whole tree attribution data form, the execution of contribution clipping 502 can also be divided into two sets of logic.

[0059] In some embodiments, when the multi-tree structure is data in the form of node attribution, given that the nodes in the multi-tree structure are decomposed according to a certain dimension to the parent indicator, the logic for cutting the contribution of node attribution is that in the same level, it is necessary to find the node with the largest contribution rate, and determine its contribution as the maximum contribution rate, and compare the contribution rate of all other nodes in the level with the maximum contribution rate. If the ratio of the current node's contribution rate to the maximum contribution rate is less than the first parameter, then the node will be cut; if the ratio of the current node's contribution rate to the maximum contribution rate is greater than or equal to the first parameter, then the node can be retained. In some embodiments, the first parameter for indicating the contribution of the node can be used as a hyperparameter and configured in the process of model training and reasoning. In some embodiments, for all nodes in the multi-tree, standardized algorithms such as top-k and same-layer information can be used to traverse and determine whether certain nodes in each level need to be cut.

[0060] In some embodiments, when the multi-branch tree structure is data in the form of whole-tree attribution, given that the focus of the whole-tree attribution format is to screen out the most important decomposition directions and the nodes with the highest contribution in the entire tree, and it is not limited to a specific decomposition dimension, it is necessary to select the decomposition path when performing contribution clipping 502. Taking key indicator A as an example, key indicator A can be decomposed according to both dimension A and dimension B. If the contribution rate of the four sub-indicators decomposed according to dimension A is all 25%, and one of the four sub-indicators decomposed according to dimension B has a contribution rate of 70%, then the decomposition path decomposed according to dimension B is a decomposition path that can highlight the key points. In this case, the decomposition according to dimension B is a better decomposition path than the decomposition according to dimension A, because the decomposition according to dimension B can find the factor with the greatest impact on key indicator A. Therefore, once the user finds the main factor affecting key indicator A, it can implement subsequent adjustments to the response strategy for business changes related to key indicator A.

[0061] Therefore, in some embodiments, when performing contribution pruning for the entire tree, it is necessary to traverse the contribution rates of all leaf nodes to the root node, and then select K (number of candidates) disassembly paths as candidate paths. It is understandable that the paths selected at this time are all paths that completely reach the leaf nodes. In the process of selecting candidate paths, whether the path is optional can be determined based on the balance of the direct contribution rates of the child nodes under each candidate path. Combined with Figure 5C , Figure 5C A schematic diagram shows an example 500C of determining a balance degree of direct contributions of nodes in a candidate path according to an embodiment of the present disclosure. Figure 5C The Gini coefficient is used to characterize the degree of balance of the direct contribution rate of the sub-nodes in the disassembly path. If the Gini coefficient is closer to 0, the degree of balance of the direct contribution rate of the sub-nodes in the disassembly path is more uneven; if the Gini coefficient is closer to 1, the degree of balance of the direct contribution rate of the sub-nodes in the disassembly path is more even. The calculation formula of the Gini coefficient is as follows:

[0062] Among them, Figure 5C As shown, A is the area of the upper half of the curve and B is the area of the lower half of the curve.

[0063] To describe more clearly Figure 5CThe Gini coefficient is used to characterize the degree of balance in the direct contribution rate of child nodes under a decomposition path. The following example illustrates this: Assume that under a certain decomposition path, the contribution rates of child nodes X, Y, and Z are 10%, 20%, and 70%, respectively; and the ratios of the change values of child nodes X, Y, and Z to the root node are 30%, 40%, and 30%, respectively. Then we can first draw a graph based on the contribution rate in ascending order. At this time, the area under the X curve is (10%+0) * 30% / 2 = 0.015. Similarly, the area under the Y curve is (10%+30%) * 40% / 2 = 0.08, and the area under the Z curve is (30%+100%) * 30% / 2 = 0.195. Therefore, the area of B is the sum of the three parts, 0.29, and the area of A is 0.21. At this time, according to formula (1), we can calculate G = 0.21 / 0.5 = 0.42, which is close to 0, indicating that the contribution rates of the sub-nodes under this decomposition path are still different.

[0064] In some embodiments, after determining the K candidate paths for the entire tree attribution, it is also necessary to determine whether the disassembly depth D of these paths is necessary. Figure 5B , the depth of the key indicator A of the root node is 0. The depth of the node P is 2. If the disassembly depth D of the candidate path is greater than the preset disassembly depth, then all candidate paths with a depth greater than D can be discarded. In some embodiments, the disassembly depth D and the candidate number K of the candidate path can be used as hyperparameters, i.e., first parameters, for trimming the contribution rate of the whole tree attribution. Users can choose different first parameters according to the different types of businesses related to the root node. In summary, in the process of trimming the contribution of the whole tree attribution, it is necessary to consider the depth D of the disassembly path and the candidate number K of the candidate path.

[0065] refer to Figure 5A After executing contribution clipping 502, importance clipping 503 can be executed to further ensure that the logical capability of model processing is in a normal state. The reason is that after contribution clipping 502, the structure of the multi-branch tree may not be within the scope of model processing. Therefore, importance clipping 503 can be used as a backup operation to ensure the availability of model input content.

[0066] In some embodiments, after contribution pruning 502 is performed on the multitree, the contributions of the multitree layers can be reordered so that nodes with greater contributions within the same layer are ordered at the top (left), and then importance pruning 503 can be performed again. In some embodiments, pruning can be performed starting from a subsequent traversal until the depth of the multitree meets the requirement (or until the number of lines in the text description meets the requirement of the second parameter indicating node importance).

[0067] In some embodiments, in view of the difference between whole tree attribution and node attribution, the strength of executing importance clipping 503 is also different. In node attribution, the minimum strength of importance clipping of node attribution is non-leaf nodes. The reason is that when a user clicks on a node and wants to view the details of the contribution rate of each child node under the disassembly dimension, if the child node is clipped, then after the child node is clipped, only the contribution rate of the parent node of the child node will be left. At this time, the user will want to see the details of the contribution rate of the child node, but because of the clipping, there will be a situation where there are no child nodes to see. Therefore, when executing the importance clipping of node attribution, only non-leaf nodes can be clipped, so that the parent node with a lower contribution rate among the parent nodes of the child node can be found and deleted together, while the parent node with a higher contribution rate can retain the disassembly of its child node. In whole tree attribution, the strength of executing importance clipping 503 is not limited. Figure 5D FIG. 5 is a schematic diagram showing an example 500D of performing importance clipping according to an embodiment of the present disclosure. Figure 5D When performing importance pruning of node-attributed data, M, P, and Q in the solid box can be deleted together; and when performing importance pruning of the whole-tree-attributed data, P in the dashed box can be deleted.

[0068] Continue to refer Figure 5A After performing contribution clipping 502 and importance clipping 503 on the original data 501, the structured data can be converted to text 504. In some embodiments, converting the structured data to text is to output the data in the form of a multi-tree structure as a text description. In some embodiments, to ensure data integrity, a depth-first traversal (DFS) algorithm can be used to output the data in the text description form.

[0069] Back to Figure 4 After adaptive pruning 402 is performed, normalization serialization 403 may be performed to enable the model to obtain input focused on logic. Figure 5A Normalization and serialization 403 can be divided into template matching 505 and key description replacement 506. In some embodiments, template matching 505 is performed to extract key information from specific patterns in the text description. For example, for the following description, information can be obtained, including the decomposition dimension "[Decomposition Dimension_1]", the indicators "Sub-Indicator_2" and "Sub-Indicator_3", and multiple fields such as key numerical descriptions.

[0070]

Disassembly dimension_1

[0071]

Sub-indicator_2

[0072]

Sub-indicator_3

[0073] In some embodiments, regular expressions can be used to implement template matching 505 of key information. In some embodiments, after the template matching 505 operation of key information is completed, key description replacement 506 can be performed, that is, the decomposed dimensions, indicators, and key values obtained through template matching 505 are replaced. In some embodiments, preset fixed symbols such as token_1, token_2, etc. can be used to replace key fields. This can reduce the number of numbers occupied by format characters, while also accelerating the model's understanding of input fields and improving the efficiency of model output. At the same time, the method of using fixed symbols to replace key fields can also prevent data leakage.

[0074] Back to Figure 4 After the normalization and serialization 403 is completed, the model (inference) 404 can be executed, as shown in Figure 5A In some embodiments, when the data processed text is input into the language model for reasoning, an output with fixed symbols can be obtained, and then it is necessary to perform Figure 4 The normalized deserialization shown is 405. Figure 5A , that is, key description restoration 508 can be performed. In some embodiments, key description restoration 508 refers to restoring the key description replaced with fixed symbols in the normalization serialization 403 operation (including template matching 505 and key description replacement 506) to the original description, so that the user can obtain a readable attribution summary of the original data 401. This method uses adaptive clipping in the rule part to streamline the information input to the model, and supplements the logical capabilities of the large model through partial rule judgment. At the same time, the process of normalization serialization, model inference and normalization deserialization is adopted to reduce the interference of data format on logic and reduce the time required for large model answers.

[0075] By generating analysis tree-structured data and combining multiple data processing steps (including pruning, normalization, and serialization), we improve the accuracy and pertinence of data attribution. Furthermore, we leverage language models to generate attribution summaries, effectively extracting key information. This allows users to efficiently and accurately obtain data attribution results even when faced with complex data, significantly enhancing their data attribution experience.

[0076] Continue to refer Figure 4 As mentioned above, data processing 410 is responsible for optimizing the input of the model, while model building 420 is responsible for training the model for inference, which includes training data construction 422, model training 424, and model testing 426. It can be understood that there is a sequential dependency between training data construction 422, model training 424, and model testing 426.

[0077] The following will be combined Figure 6A Let's further describe model building 420. Figure 6A FIG. 6 is a schematic diagram showing an example structure 600A of model-based data attribution according to an embodiment of the present disclosure. Figure 6A As shown, before executing model training 424, training data construction 422 must be performed first. This is because the model needs to adjust model parameters that determine model behavior and performance based on the training data. In some embodiments, training data construction 422 can be further divided into forged data generation 601, data processing 602, and legitimacy verification 603.

[0078] In some embodiments, given the powerful text generation capability of the language model, training data can be generated based on a predetermined format using the language model. Figures 6B-6D That is, the predetermined format is described. In some embodiments, the training data includes training node attribution data and training whole tree attribution data. As previously mentioned, the node attribution data and the whole tree attribution data are various types of multi-branch trees, so the purpose of generating training data is to generate various types of multi-branch tree structures.

[0079] In order to enhance the generalization of the model, in the process of generating training data, a fixed symbol similar to token_1 can be used to represent the key indicators described by each multi-branch tree node, and the key values in the key value description can be randomly generated. In some embodiments, for the effectiveness of model training, the depth of the generated training multi-branch tree can be set to a maximum of 7, and the number of child nodes per layer can be 1-5. Figure 6B , Figure 6B A schematic diagram illustrating an example node attribution 600B generated for training according to an embodiment of the present disclosure is shown.

[0080] In some embodiments, the constructed training root node may include a time range (baseline period and observation period), key data descriptions of key nodes (key indicators), outliers detected during the observation period, trends detected during the observation period, and so on. In some embodiments, the dates of the constructed training root node can be randomly selected from any X days between January 1, 1900, and December 31, 2030, as the base period and the observation period. It is also necessary to ensure that the base period and the observation period do not overlap, and that the observation period occurs after the base period. In some embodiments, the key data descriptions of key nodes (key indicators) in the constructed training root node can be randomly generated. In some embodiments, the outlier type in the constructed training root node can be set based on "fluctuation exceeding 2 standard deviations of the 30-day average fluctuation," "month-on-month fluctuation exceeding 10.0% compared to the previous day," or "fluctuation exceeding 2 standard deviations of the 7-day average fluctuation." The training data for the root node is constructed based on permutations and combinations of these three types of outliers. The probability of occurrence of an outlier can also be set to 50% to generate the training data for the root node. In some embodiments, the probability of a trend occurring during the observation period in the constructed training root node can also be set to 50%, where the trend is rising and falling. Figure 6C , Figure 6C FIG. 6 is a schematic diagram showing an example root node 600C generated for training according to an embodiment of the present disclosure.

[0081] In some embodiments, given that the multi-branch tree structure in the form of a whole tree includes a root node, a disassembly node, and a formula node, the above-mentioned process of generating training node attribution data and training root node data can be used as a reference when generating training whole tree data. Formula nodes (such as sub-indicators A=B*C*D*E) are special disassembly nodes in the multi-branch tree structure in the form of a whole tree, but their hierarchical depth is 1 layer. Therefore, when generating training whole tree data, the process of generating disassembly nodes can also be used as a reference, but the depth of its multi-branch tree structure needs to be set to a maximum of 1. Figure 6D , Figure 6D FIG. 6 is a schematic diagram showing an example whole tree 600D generated for training according to an embodiment of the present disclosure.

[0082] Back to Figure 6A In some embodiments, after the forged data generation 601 is performed with the help of a language model, data processing 602 is required to obtain labels in the training data. In some embodiments, labels in the training data can be generated with the help of data processing (model) such as a language model. Figure 6E , Figure 6E A schematic diagram illustrating an example prompt 600E generated for training according to an embodiment of the present disclosure is shown.

[0083] refer to Figure 6AIn some embodiments, after the steps of generating and processing the training data are completed, in order to ensure that the generated training data does not affect the training of the model, it is also necessary to perform a validity verification 603. For example, during the training process, if the input description is "token_1 contribution rate<r 50.70%> **", and the model answer may be "token_1 contributes the most, with a contribution rate of<r 50.70> ", this incomplete quotation is not allowed because<r 50.70%> The corresponding is ** <font color="red">50.70%< / font> **, where r indicates red font formatting. This means that if a symbol is missing from the answer (for example, a percent sign), normalized deserialization will fail. Therefore, validation primarily verifies whether key values in the model's answer appear in the vocabulary. If a key value not in the vocabulary appears in an answer, it indicates that the model does not fully reference the data format, and such data will be removed.

[0084] Back to Figure 4 After the training data construction 422 is completed, the model training 422 can be performed. Figure 6A , model training 424 may include supervised fine-tuning 604, that is, the model may be trained using the SFT method. In some embodiments, the specific parameters of the training process may be as shown in Table 1 below:

[0085]

[0086] refer to Figure 4 After completing model training 424, it is necessary to test the performance of the model, so it is also necessary to perform model testing 426. Figure 6A In some embodiments, the model test 426 also needs to perform an evaluation standard test 605 and a secondary evaluation test 606. In some embodiments, the evaluation standard test 605 can be performed according to the rules in Table 2 below.

[0087]

[0088] In some embodiments, in order to ensure the full release of model performance, a secondary evaluation test 606 can also be performed on the model to achieve a comprehensive evaluation of the model. In some embodiments, real business data can be used to test the model, and the above six indicators can be used to comprehensively obtain the model score from five dimensions. In some embodiments, a manual method can also be used to score the model results. Through this unsupervised automatic construction of training data method, various types of training data for improving model performance can be generated. At the same time, a variety of evaluation and inspection methods are combined to ensure the effectiveness of model performance verification.

[0089] Therefore, according to the embodiments of the present disclosure, by generating analysis tree-structured data and combining multiple processing steps (including pruning, normalization and serialization), the accuracy and pertinence of data attribution are improved. Furthermore, by utilizing language models to generate attribution summaries, key information is effectively extracted. This enables users to efficiently and accurately obtain data attribution results when faced with complex data, significantly improving their operational experience in data attribution.

[0090] Figure 7 FIG. 7 is a block diagram of an apparatus 700 for data attribution according to some embodiments of the present disclosure. Figure 7 As shown, the device 700 includes a data acquisition module 702, which is configured to acquire data in the form of text descriptions, wherein the data includes changes in the data within a date range. The device 700 also includes a parse tree generation module 704, which is configured to generate a parse tree based on the data in the form of text descriptions and preset dimensions, wherein the parse tree includes at least one of node attribution data and whole tree attribution data, the node attribution data indicating that the data is presented in the form of node hierarchical nesting, each node in the node attribution data includes a key numerical description of the node, and the whole tree attribution data is determined based on the node's disassembly path. The device 700 also includes a parse tree pruning module 706, which is configured to prune the parse tree based on a first parameter and a second parameter to obtain data in the form of a pruned multi-branch tree structure, wherein the first parameter indicates whether to prune a node based on the contribution of the node in the multi-branch tree structure, and the second parameter indicates whether to prune a node based on the importance of the node in the multi-branch tree structure. The apparatus 700 further includes a data generation module 708 configured to normalize and serialize the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in the form of a text description. Furthermore, the apparatus 700 further includes an attribution summary generation module 710 configured to generate an attribution summary for the data based on the normalized serialized data using a language model.

[0091] Figure 8 A block diagram of an electronic device 800 is shown, in accordance with certain embodiments of the present disclosure. Figure 8 FIG1 shows a block diagram of an electronic device 800 according to some embodiments of the present disclosure. The device 800 may be a device or apparatus described in an embodiment of the present disclosure. Figure 8As shown, the device 800 includes a central processing unit (CPU) and / or a graphics processing unit (GPU) 801, which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) 802 or loaded from a storage unit 808 into a random access memory (RAM) 803. Various programs and data required for the operation of the device 800 can also be stored in the RAM 803. The CPU / GPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804. Although not shown in FIG. Figure 8 As shown in FIG, device 800 may further include a co-processor.

[0092] Various components in device 800 are connected to I / O interface 805, including an input unit 806, such as a keyboard, mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a magnetic disk, optical disk, etc.; and a communication unit 809, such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows device 800 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0093] The various methods or processes described above may be performed by the CPU / GPU 801. For example, in some embodiments, the methods may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed onto the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the CPU / GPU 801, one or more steps or actions in the methods or processes described above may be performed.

[0094] In some embodiments, the methods and processes described above may be implemented as a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present disclosure.

[0095] A computer-readable storage medium can be a tangible device that can hold and store instructions used by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure within a groove on which instructions are stored, and any suitable combination thereof. As used herein, a computer-readable storage medium is not to be construed as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse passing through a fiber optic cable), or an electrical signal transmitted through an electrical wire.

[0096] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions to be stored in the computer-readable storage medium in each computing / processing device.

[0097] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages and conventional procedural programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, the state information of the computer-readable program instructions is used to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), so that the electronic circuit can execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

[0098] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device is generated that implements the functions / actions specified in one or more blocks in the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium, where these instructions cause the computer, programmable data processing device, and / or other device to operate in a specific manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing various aspects of the functions / actions specified in one or more blocks in the flowchart and / or block diagram.

[0099] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device, so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0100] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and the part of the module, program segment or instruction contains one or more executable instructions for realizing the prescribed logical function. In some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart, can be implemented by a special hardware-based system that performs the prescribed function or action, or can be implemented by a combination of special hardware and computer instructions.

[0101] The embodiments of the present disclosure have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, practical applications, or technical improvements to existing technologies, or to enable other persons skilled in the art to understand the embodiments disclosed herein.

[0102] Some example implementations of the present disclosure are listed below.

[0103] Example 1. A method for data attribution, comprising:

[0104] Obtaining data in a text description format, including changes in the data within a date range;

[0105] Generate an analysis tree based on the data in the text description form and preset dimensions, the analysis tree including at least one of node attribution data and whole-tree attribution data, the node attribution data indicating that the data is presented in a node-level nested form, each node in the node attribution data including a key numerical description of the node, and the whole-tree attribution data determined based on a node decomposition path;

[0106] Pruning the analysis tree based on a first parameter and a second parameter to obtain data in a pruned multi-tree structure, wherein the first parameter indicates whether to prune the node based on a contribution of the node in the multi-tree structure, and the second parameter indicates whether to prune the node based on an importance of the node in the multi-tree structure;

[0107] Normalizing and serializing the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in a text description form; and

[0108] The language model generates an attribution summary for the data based on the normalized serialized data.

[0109] Example 2. The method of Example 1, wherein pruning the parse tree based on the first parameter and the second parameter to obtain pruned data in a multi-branch tree structure comprises:

[0110] Converting the analysis tree into data in a multi-tree structure, wherein the multi-tree structure includes a plurality of nodes; and

[0111] Based on the first parameter and the second parameter, pruning nodes in the multi-tree structure to obtain pruned data in the multi-tree structure; and

[0112] The pruned data in the multi-tree structure form is converted into the pruned data in the multi-tree structure form presented in a text form.

[0113] Example 3. The method of any one of Examples 1-2, wherein the contribution is determined based on a contribution rate of a node, wherein pruning the nodes in the multi-tree structure based on the first parameter and the second parameter to obtain pruned data in the multi-tree structure comprises:

[0114] In response to the analysis tree being node attribution data, traversing downward in sequence a plurality of nodes at each level from a level where a highest node of data in a multi-branch tree structure corresponding to the node attribution data is located;

[0115] Determining a maximum contribution rate, the maximum contribution rate indicating the contribution rate of a node having the largest contribution rate to the highest node among each node at the same level in the data in the multi-branch tree structure corresponding to the node attribution data;

[0116] Determine a ratio of a contribution rate of each node in each node of the same level in the data in the form of a multi-branch tree structure corresponding to the node attribution data to the maximum contribution rate; and

[0117] In response to the ratio being less than the first parameter, in each node at the same level, each node having the ratio less than the first parameter is pruned.

[0118] Example 4. The method according to any one of Examples 1-3, further comprising:

[0119] In response to the analysis tree being adjusted based on the first parameter, sorting each node in the same level in the analysis tree;

[0120] In response to the sorted analysis tree being node attribution data, determining the number of levels of each leaf node in the data in the multi-branch tree structure corresponding to the node attribution data; and

[0121] In response to the number of levels being greater than the second parameter, each node whose number of levels is greater than the second parameter and the previous non-leaf node of the node are pruned cyclically.

[0122] Example 5. The method of any one of Examples 1-4, wherein the first parameter includes a disassembly depth and a number of candidate disassembly paths, the disassembly depth and the number of candidates being related to the disassembly paths, and wherein pruning nodes in the multi-tree structure based on the first parameter and the second parameter to obtain pruned data in the multi-tree structure includes:

[0123] In response to the analysis tree being the whole-tree attribution data, traversing each leaf node in the data in the multi-branch tree structure corresponding to the whole-tree attribution data;

[0124] Determining the contribution rate of each leaf node to the root node in the data in the multi-branch tree structure corresponding to the whole tree attribution data; and

[0125] Whether to prune each node is determined based on the disassembly depth, the number of candidates, and the contribution rate of each leaf node to the root node in the multi-branch tree structure data corresponding to the whole tree attribution data.

[0126] Example 6. The method according to any one of Examples 1-5, further comprising:

[0127] In response to the analysis tree being adjusted based on the first parameter, sorting each node in the same level in the analysis tree;

[0128] Determining the number of levels of each leaf node in the sorted data in the multi-branch tree structure corresponding to the whole tree attribution data; and

[0129] Based on the second parameter and the number of levels, it is determined in a loop whether to prune each leaf node in the multi-branch tree structure data corresponding to the whole tree attribution data.

[0130] Example 7. The method of any one of Examples 1-6, wherein converting the pruned data in the multi-tree structure form into the pruned data in the multi-tree structure form presented in text form comprises:

[0131] Depth-first traversal of the pruned data in the multi-tree structure; and

[0132] The pruned data in the multi-tree structure form is converted into the pruned data in the multi-tree structure form presented in a text form.

[0133] Example 8. The method of any one of Examples 1-7, wherein normalizing and serializing the pruned data in the multitree structure to generate normalized serialized data comprises:

[0134] Determining a plurality of fields in the pruned data in the multi-tree structure presented in text form, the plurality of fields including key fields and data fields; and

[0135] The key field is replaced with a preset symbol to generate the normalized serialized data.

[0136] Example 9. The method of any one of Examples 1-8, wherein the language model is a first language model, wherein generating, by the language model based on the normalized serialized data, an attribution summary for the data comprises:

[0137] The first language model generates an attribution summary with preset symbols based on the normalized serialized data; and

[0138] The attribution summary with the preset symbol is normalized and deserialized to generate an attribution summary for the data.

[0139] Example 10. The method of any one of Examples 1-9, further comprising:

[0140] The second language model generates training data for the first language model based on a predetermined format, wherein the training data includes training node attribution data and training whole tree attribution data; and

[0141] The first language model is trained based on the training data.

[0142] Example 11. The method of any one of Examples 1-10, wherein generating, by the second language model based on a predetermined format, training data for the first language model comprises:

[0143] The second language model generates a training label for the training data based on a preset prompt and the training data.

[0144] Example 12. The method of any one of Examples 1-11, further comprising:

[0145] Determining whether the generated training data satisfies a legality check; and

[0146] In response to the generated training data not satisfying the legality check, the generated training data not satisfying the legality check is removed.

[0147] Example 13. The method of any one of Examples 1-12, further comprising:

[0148] In response to detecting a user interaction with the first control, displaying an attribution summary of node attribution data regarding the data; and

[0149] In response to detecting a user interaction with the first control, an attribution summary of the entire tree of attribution data for the data is displayed.

[0150] Example 14. An apparatus for data attribution, comprising:

[0151] A data acquisition module is configured to acquire data in a text description format, wherein the data includes changes in the data within a date range;

[0152] a parsing tree generation module configured to generate a parsing tree based on the data in the text description form and preset dimensions, the parsing tree including at least one of node attribution data and whole-tree attribution data, the node attribution data indicating that the data is presented in a node-level nested form, each node in the node attribution data including a key numerical description of the node, and the whole-tree attribution data determined based on a node decomposition path;

[0153] a parse tree pruning module configured to prune the parse tree based on a first parameter and a second parameter to obtain pruned data in a multi-branch tree structure, wherein the first parameter indicates whether to prune a node based on a contribution of the node in the multi-branch tree structure, and the second parameter indicates whether to prune the node based on an importance of the node in the multi-branch tree structure;

[0154] a data generation module configured to normalize and serialize the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in a text description form; and

[0155] The attribution summary generation module is configured to generate an attribution summary for the data based on the normalized serialized data using a language model.

[0156] Example 15. The apparatus of any one of Example 14, wherein the parse tree pruning module comprises:

[0157] A first conversion module is configured to convert the analysis tree into data in a multi-tree structure, wherein the multi-tree structure includes a plurality of nodes; and

[0158] a first pruning module configured to prune nodes in the multi-tree structure based on the first parameter and the second parameter to obtain pruned data in the multi-tree structure; and

[0159] The second conversion module is configured to convert the pruned data in the multi-tree structure form into the pruned data in the multi-tree structure form presented in text form.

[0160] Example 16. The apparatus according to any one of Examples 14-15, wherein the contribution is determined based on a contribution rate of a node, wherein the first trimming module comprises:

[0161] A first traversal module is configured to, in response to the analysis tree being node attribution data, sequentially traverse downwards a plurality of nodes at each level starting from a level where a highest node of data in a multi-branch tree structure corresponding to the node attribution data is located;

[0162] A first determining module is configured to determine a maximum contribution rate, wherein the maximum contribution rate indicates a contribution rate of a node having the largest contribution rate to the highest node among each node at the same level in the data in the multi-tree structure corresponding to the node attribution data;

[0163] A second determining module is configured to determine a ratio of a contribution rate of each node in each node of the same level in the data in a multi-tree structure corresponding to the node attribution data to the maximum contribution rate; and

[0164] The second pruning module is configured to, in response to the ratio being less than the first parameter, prune each node whose ratio is less than the first parameter in each node at the same level.

[0165] Example 17. The apparatus of any of Examples 14-16, further comprising:

[0166] a first sorting module configured to sort each node in the same level of the analysis tree in response to the analysis tree being adjusted based on the first parameter;

[0167] a third determining module configured to, in response to the sorted analysis tree being node attribution data, determine the number of levels of each leaf node in the data in the multi-branch tree structure corresponding to the node attribution data; and

[0168] The third pruning module is configured to, in response to the number of levels being greater than the second parameter, cyclically prune each node whose number of levels is greater than the second parameter and the previous non-leaf node of the node.

[0169] Example 18. The apparatus of any one of Examples 14-17, wherein the first parameter comprises a disassembly depth and a number of candidate disassembly paths, the disassembly depth and the number of candidate disassembly paths being related to the disassembly paths, wherein the first cropping module comprises:

[0170] a second traversal module configured to, in response to the analysis tree being whole-tree attribution data, traverse each leaf node in the data in the multi-branch tree structure corresponding to the whole-tree attribution data;

[0171] a fourth determining module configured to determine a contribution rate of each leaf node to a root node in the data in a multi-branch tree structure corresponding to the entire tree attribution data; and

[0172] The fifth determination module is configured to determine whether to prune each node based on the disassembly depth, the number of candidates, and the contribution rate of each leaf node to the root node in the multi-branch tree structure data corresponding to the whole tree attribution data.

[0173] Example 19. The apparatus of any one of Examples 14-18, further comprising:

[0174] a second sorting module configured to sort each node in the same level of the analysis tree in response to the analysis tree being adjusted based on the first parameter;

[0175] a sixth determining module configured to determine the number of levels of each leaf node in the sorted data in the multi-branch tree structure corresponding to the whole tree attribution data; and

[0176] The seventh determination module is configured to cyclically determine whether to prune each leaf node in the multi-branch tree structure data corresponding to the whole tree attribution data based on the second parameter and the number of levels.

[0177] Example 20. The apparatus of any of Examples 14-19, wherein the second conversion module comprises:

[0178] A third traversal module is configured to perform a depth-first traversal of the pruned data in the multi-tree structure; and

[0179] The third conversion module is configured to convert the pruned data in the multi-tree structure form into the pruned data in the multi-tree structure form presented in text form.

[0180] Example 21. The apparatus of any one of Examples 14-20, wherein the data generation module comprises:

[0181] an eighth determining module configured to determine a plurality of fields in the data in the multi-tree structure after being pruned and presented in text form, the plurality of fields including a key field and a data field; and

[0182] The first generating module is configured to replace the key field with a preset symbol to generate the normalized serialized data.

[0183] Example 22. The apparatus of any one of Examples 14-21, wherein the language model is a first language model, wherein the attribution summary generation module comprises:

[0184] A second generating module is configured to generate an attribution summary with preset symbols based on the normalized serialized data by the first language model; and

[0185] The third generating module is configured to normalize and deserialize the attribution summary with the preset symbols to generate an attribution summary for the data.

[0186] Example 23. The apparatus of any one of Examples 14-22, further comprising:

[0187] a fourth generating module configured to generate training data for the first language model based on a predetermined format using the second language model, the training data including training node attribution data and training whole tree attribution data; and

[0188] The training module is configured to train the first language model based on the training data.

[0189] Example 24. The apparatus of any one of Examples 14-23, wherein the fourth generating module comprises:

[0190] The training label generation module is configured to generate a training label for the training data based on a preset prompt and the training data by the second language model.

[0191] Example 25. The apparatus of any one of Examples 14-24, further comprising:

[0192] a ninth determining module, configured to determine whether the generated training data satisfies a legality check; and

[0193] The removal module is configured to remove the generated training data that does not meet the legality check in response to the generated training data not meeting the legality check.

[0194] Example 26. The apparatus of any one of Examples 14-25, further comprising:

[0195] a first display module configured to display an attribution summary of node attribution data regarding the data in response to detecting a user interaction with the first control; and

[0196] The second display module is configured to display an attribution summary of the entire tree of attribution data regarding the data in response to detecting a user interaction with the first control.

[0197] Example 27. An electronic device comprising:

[0198] processor; and

[0199] A memory coupled to the processor, the memory having instructions stored therein, wherein when the instructions are executed by the processor, the electronic device performs actions, the actions comprising:

[0200] Obtaining data in a text description format, including changes in the data within a date range;

[0201] Generate an analysis tree based on the data in the text description form and preset dimensions, the analysis tree including at least one of node attribution data and whole-tree attribution data, the node attribution data indicating that the data is presented in a node-level nested form, each node in the node attribution data including a key numerical description of the node, and the whole-tree attribution data determined based on a node decomposition path;

[0202] Pruning the analysis tree based on a first parameter and a second parameter to obtain data in a pruned multi-tree structure, wherein the first parameter indicates whether to prune the node based on a contribution of the node in the multi-tree structure, and the second parameter indicates whether to prune the node based on an importance of the node in the multi-tree structure;

[0203] Normalizing and serializing the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in a text description form; and

[0204] The language model generates an attribution summary for the data based on the normalized serialized data.

[0205] Example 28. The electronic device of Example 27, wherein pruning the analysis tree based on the first parameter and the second parameter to obtain pruned data in a multi-branch tree structure comprises:

[0206] Converting the analysis tree into data in a multi-tree structure, wherein the multi-tree structure includes a plurality of nodes; and

[0207] Based on the first parameter and the second parameter, pruning nodes in the multi-tree structure to obtain pruned data in the multi-tree structure; and

[0208] The pruned data in the multi-tree structure form is converted into the pruned data in the multi-tree structure form presented in a text form.

[0209] Example 29. The apparatus of any of Examples 27-28, wherein the contribution is determined based on a contribution rate of a node, wherein pruning the nodes in the multi-tree structure based on the first parameter and the second parameter to obtain pruned data in the multi-tree structure comprises:

[0210] In response to the analysis tree being node attribution data, traversing downward in sequence a plurality of nodes at each level from a level where a highest node of data in a multi-branch tree structure corresponding to the node attribution data is located;

[0211] Determining a maximum contribution rate, the maximum contribution rate indicating the contribution rate of a node having the largest contribution rate to the highest node among each node at the same level in the data in the multi-branch tree structure corresponding to the node attribution data;

[0212] Determine a ratio of a contribution rate of each node in each node of the same level in the data in the form of a multi-branch tree structure corresponding to the node attribution data to the maximum contribution rate; and

[0213] In response to the ratio being less than the first parameter, in each node at the same level, each node having the ratio less than the first parameter is pruned.

[0214] Example 30. The apparatus of any of Examples 27-29, further comprising:

[0215] In response to the analysis tree being adjusted based on the first parameter, sorting each node in the same level in the analysis tree;

[0216] In response to the sorted analysis tree being node attribution data, determining the number of levels of each leaf node in the data in the multi-branch tree structure corresponding to the node attribution data; and

[0217] In response to the number of levels being greater than the second parameter, each node whose number of levels is greater than the second parameter and the previous non-leaf node of the node are pruned cyclically.

[0218] Example 31. The apparatus of any of Examples 27-30, wherein the first parameter comprises a disassembly depth and a number of candidate disassembly paths, the disassembly depth and the number of candidate disassembly paths being related to the disassembly paths, wherein pruning nodes in the multi-tree structure based on the first parameter and the second parameter to obtain pruned data in the multi-tree structure comprises:

[0219] In response to the analysis tree being the whole-tree attribution data, traversing each leaf node in the data in the multi-branch tree structure corresponding to the whole-tree attribution data;

[0220] Determining the contribution rate of each leaf node to the root node in the data in the multi-branch tree structure corresponding to the whole tree attribution data; and

[0221] Whether to prune each node is determined based on the disassembly depth, the number of candidates, and the contribution rate of each leaf node to the root node in the multi-branch tree structure data corresponding to the whole tree attribution data.

[0222] Example 32. The apparatus of any of Examples 27-31, further comprising:

[0223] In response to the analysis tree being adjusted based on the first parameter, sorting each node in the same level in the analysis tree;

[0224] Determining the number of levels of each leaf node in the sorted data in the multi-branch tree structure corresponding to the whole tree attribution data; and

[0225] Based on the second parameter and the number of levels, it is determined in a loop whether to prune each leaf node in the multi-branch tree structure data corresponding to the whole tree attribution data.

[0226] Example 33. The apparatus of any one of Examples 27-32, wherein converting the pruned data in the multi-tree structure form into the pruned data in the multi-tree structure form presented in text form comprises:

[0227] Depth-first traversal of the pruned data in the multi-tree structure; and

[0228] The pruned data in the multi-tree structure form is converted into the pruned data in the multi-tree structure form presented in a text form.

[0229] Example 34. The apparatus of any one of Examples 27-33, wherein normalizing and serializing the pruned data in the multitree structure to generate normalized serialized data comprises:

[0230] Determining a plurality of fields in the pruned data in the multi-tree structure presented in text form, the plurality of fields including key fields and data fields; and

[0231] The key field is replaced with a preset symbol to generate the normalized serialized data.

[0232] Example 35. The apparatus of any of Examples 27-34, wherein the language model is a first language model, wherein generating, by the language model, an attribution summary for the data based on the normalized serialized data comprises:

[0233] The first language model generates an attribution summary with preset symbols based on the normalized serialized data; and

[0234] The attribution summary with the preset symbol is normalized and deserialized to generate an attribution summary for the data.

[0235] Example 36. The apparatus of any of Examples 27-35, further comprising:

[0236] The second language model generates training data for the first language model based on a predetermined format, wherein the training data includes training node attribution data and training whole tree attribution data; and

[0237] The first language model is trained based on the training data.

[0238] Example 37. The apparatus of any of Examples 27-36, wherein generating, by a second language model based on a predetermined format, training data for the first language model comprises:

[0239] The second language model generates a training label for the training data based on a preset prompt and the training data.

[0240] Example 38. The apparatus of any of Examples 27-37, further comprising:

[0241] Determining whether the generated training data satisfies a legality check; and

[0242] In response to the generated training data not satisfying the legality check, the generated training data not satisfying the legality check is removed.

[0243] Example 39. The apparatus of any of Examples 27-38, further comprising:

[0244] In response to detecting a user interaction with the first control, displaying an attribution summary of node attribution data regarding the data; and

[0245] In response to detecting a user interaction with the first control, an attribution summary of the entire tree of attribution data for the data is displayed.

[0246] Example 40. A computer-readable storage medium having one or more computer instructions stored thereon, wherein the one or more computer instructions are executed by a processor to implement the method according to any one of Examples 1 to 13.

[0247] Example 41. A computer program product tangibly stored on a computer-readable medium and comprising computer-executable instructions that, when executed by a device, cause the device to perform the method according to any one of Examples 1 to 13.

[0248] Although the present disclosure has been described in language specific to structural features and / or methodological logical acts, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms of implementing the claims.

Claims

1. A method for data attribution, comprising: Obtaining data in a text description format, including changes in the data within a date range; Generate an analysis tree based on the data in the text description form and preset dimensions, the analysis tree including at least one of node attribution data and whole-tree attribution data, the node attribution data indicating that the data is presented in a node-level nested form, each node in the node attribution data including a key numerical description of the node, and the whole-tree attribution data determined based on a node decomposition path; The analysis tree is pruned based on a first parameter and a second parameter to obtain data in a pruned multi-tree structure, wherein the first parameter indicates whether to prune the node based on a contribution of the node in the multi-tree structure, and the second parameter indicates whether to prune the node based on an importance of the node in the multi-tree structure, wherein pruning the analysis tree based on the first parameter and the second parameter includes: Converting the analysis tree into data in a multi-tree structure, wherein the multi-tree structure includes a plurality of nodes; In response to the analysis tree being the whole-tree attribution data, determining whether to prune each node in the data in the multi-tree structure form based on the first parameter and a contribution rate of each leaf node in the data in the multi-tree structure form to a root node, the first parameter including a disassembly depth and a candidate number of disassembly paths, the disassembly depth and the candidate number being related to the disassembly path; and Based on the second parameter, cyclically determining whether to pruned each leaf node in the data in the multi-tree structure adjusted by the first parameter; Normalizing and serializing the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in a text description form; and The language model generates an attribution summary for the data based on the normalized serialized data.

2. The method according to claim 1, further comprising: The pruned data in the multi-tree structure form is converted into the pruned data in the multi-tree structure form presented in a text form.

3. The method according to claim 2, wherein the contribution is determined based on a contribution rate of a node, wherein pruning the analysis tree based on the first parameter and the second parameter comprises: In response to the analysis tree being the node attribution data, traversing downward in sequence a plurality of nodes at each level from a level where a highest node of the data in the multi-branch tree structure corresponding to the node attribution data is located; Determining a maximum contribution rate, where the maximum contribution rate indicates the contribution rate of a node having the largest contribution rate to the highest node among each node at the same level in the data in the multi-branch tree structure corresponding to the node attribution data; Determine a ratio of a contribution rate of each node in each node of the same level in the data in the form of a multi-branch tree structure corresponding to the node attribution data to the maximum contribution rate; as well as In response to the ratio being less than the first parameter, in each node at the same level, each node having the ratio less than the first parameter is pruned.

4. The method according to claim 3, further comprising: In response to the analysis tree being adjusted based on the first parameter, sorting each node in the same level in the analysis tree; In response to the sorted analysis tree being node attribution data, determining the number of levels of each leaf node in the data in the multi-branch tree structure corresponding to the node attribution data; as well as In response to the number of levels being greater than the second parameter, each node whose number of levels is greater than the second parameter and the previous non-leaf node of the node are pruned cyclically.

5. The method according to claim 2, wherein pruning the analysis tree based on the first parameter and the second parameter further comprises: In response to the analysis tree being the whole-tree attributed data, traversing each leaf node in the data in the multi-branch tree structure corresponding to the whole-tree attributed data; Determine the contribution rate of each leaf node to the root node in the data in the multi-branch tree structure corresponding to the whole tree attribution data; as well as Whether to prune each node is determined based on the disassembly depth, the number of candidates, and the contribution rate of each leaf node to the root node in the multi-branch tree structure data corresponding to the whole tree attribution data.

6. The method according to claim 5, wherein, based on the second parameter, cyclically determining whether to prune each leaf node in the multi-tree structured data adjusted by the first parameter comprises: In response to the analysis tree being adjusted based on the first parameter, sorting each node in the same level in the analysis tree; Determining the number of levels of each leaf node in the sorted data in the multi-branch tree structure corresponding to the whole tree attribution data; as well as Based on the second parameter and the number of levels, it is determined in a loop whether to prune each leaf node in the multi-branch tree structure data corresponding to the whole tree attribution data.

7. The method according to claim 2, wherein converting the pruned data in the multi-tree structure form into the pruned data in the multi-tree structure form presented in text form comprises: Depth-first traversal of the pruned data in the multi-tree structure; as well as The pruned data in the multi-tree structure form is converted into the pruned data in the multi-tree structure form presented in a text form.

8. The method according to claim 7, wherein normalizing and serializing the pruned data in the multitree structure to generate normalized serialized data comprises: Determine a plurality of fields in the data in the multi-tree structure form after being pruned and presented in text form, wherein the plurality of fields include a key field and a data field; as well as The key field is replaced with a preset symbol to generate the normalized serialized data.

9. The method according to claim 1, wherein the language model is a first language model, wherein generating, by the language model based on the normalized serialized data, an attribution summary for the data comprises: The first language model generates an attribution summary with preset symbols based on the normalized serialized data; as well as The attribution summary with the preset symbol is normalized and deserialized to generate an attribution summary for the data.

10. The method according to claim 9, further comprising: The second language model generates training data for the first language model based on a predetermined format, wherein the training data includes training node attribution data and training whole tree attribution data; as well as The first language model is trained based on the training data.

11. The method according to claim 10, wherein generating, by the second language model based on a predetermined format, training data for the first language model comprises: The second language model generates a training label for the training data based on a preset prompt and the training data.

12. The method according to claim 11, further comprising: Determining whether the generated training data satisfies a legality check; as well as In response to the generated training data not satisfying the legality check, the generated training data not satisfying the legality check is removed.

13. The method according to claim 1, further comprising: In response to detecting a user interaction with the first control, displaying an attribution summary of node attribution data regarding the data; as well as In response to detecting a user interaction with the first control, an attribution summary of the entire tree of attribution data for the data is displayed.

14. A device for data attribution, comprising: A data acquisition module is configured to acquire data in a text description format, wherein the data includes changes in the data within a date range; a parsing tree generation module configured to generate a parsing tree based on the data in the text description form and preset dimensions, the parsing tree including at least one of node attribution data and whole-tree attribution data, the node attribution data indicating that the data is presented in a node-level nested form, each node in the node attribution data including a key numerical description of the node, and the whole-tree attribution data determined based on a node decomposition path; A parse tree pruning module is configured to prune the parse tree based on a first parameter and a second parameter to obtain pruned data in a polytree structure, wherein the first parameter indicates whether to prune a node based on a contribution of the node in the polytree structure, and the second parameter indicates whether to prune the node based on an importance of the node in the polytree structure, wherein pruning the parse tree based on the first parameter and the second parameter includes: In response to the analysis tree being the whole-tree attribution data, determining whether to prune each node in the data in the multi-tree structure form based on the first parameter and a contribution rate of each leaf node in the data in the multi-tree structure form to a root node, the first parameter including a disassembly depth and a candidate number of disassembly paths, the disassembly depth and the candidate number being related to the disassembly path; and Based on the second parameter, cyclically determining whether to pruned each leaf node in the data in the multi-tree structure adjusted by the first parameter; a data generation module configured to normalize and serialize the pruned data in the multi-tree structure to generate normalized serialized data, wherein the normalized serialized data is presented in a text description form; and The attribution summary generation module is configured to generate an attribution summary for the data based on the normalized serialized data using a language model.

15. An electronic device comprising: processor; as well as A memory coupled to the processor, the memory having instructions stored therein, wherein when the instructions are executed by the processor, the electronic device performs the method according to any one of claims 1 to 13.

16. A computer program product tangibly stored on a non-transitory computer-readable medium and comprising computer-executable instructions for performing the method according to any one of claims 1 to 13.

Images