AIGC model material attribution method and device

Abstract

The embodiment of the invention provides a method and device for generating a material attribution of an AIGC model through artificial intelligence. In the method, a prediction material corresponding to a first cue word and a prediction click rate of the prediction material are determined through an AIGC model, data processing based on a model interpretable algorithm is executed, and contribution values of a plurality of feature values in the first cue word to the prediction click rate are obtained. Then, a plurality of first feature values with high contribution are determined, first interpretive verbal skills corresponding to the first feature values respectively are inquired from a plurality of pre-constructed triples, any triple comprises the feature values, the field scene and the interpretive verbal skills, and the interpretive verbal skills are used for representing the effect and / or influence of the corresponding feature values in the field scene; the second cue word is input into an attribution report generation model, an attribution report is obtained, the attribution report comprises a contribution description text of a plurality of first feature values to the predicted click rate, and a task instruction contained in the second cue word is combined with the first interpretive verbal skill to generate the contribution description text.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and more particularly to a material attribution method and apparatus for an AIGC model. Background Technology

[0002] In recent years, machine learning technology has been widely applied across various industries. With the continuous development of advanced algorithmic models (various ensemble models, deep neural models, large language models), model performance has been constantly improved. However, while increasingly complex models provide users with more accurate decision-making, the rationale behind these decisions is becoming increasingly difficult for operations personnel to understand. For example, in the field of image and text content generation, AIGC (Artificial Intelligence Generated Content) models can generate high-quality images based on text descriptions. However, these models typically operate as black boxes, making it difficult for operations personnel to understand why the model generates specific material elements or text styles. This lack of interpretability limits the effective evaluation and iterative optimization of the generated results. Therefore, providing understandable decision-making rationale for the generation model becomes crucial. Simultaneously, the data processing process needs to protect potentially sensitive data.

[0003] Currently, there is a desire for improved solutions that can more effectively interpret the results generated by the AIGC model and produce attribution reports that operations personnel can understand. Summary of the Invention

[0004] This specification describes one or more embodiments of a material attribution method and apparatus for an AIGC model, which more effectively interprets the generated results of the AIGC model and obtains attribution reports that can be understood by operations personnel. The specific technical solution is as follows.

[0005] In a first aspect, the embodiments provide a material attribution method for an AIGC model, including:

[0006] The first prompt word to be generated for display materials is input into the AIGC model. The AIGC model generates corresponding predicted materials and their predicted click-through rates based on the feature values ​​of several features contained in the first prompt word.

[0007] For the first prompt word and the predicted click-through rate, perform data processing based on a model-interpretable algorithm to obtain the contribution values ​​of several feature values ​​to the predicted click-through rate;

[0008] For a number of first feature values ​​whose contribution values ​​are greater than a preset threshold, query the first explanatory statements corresponding to the first feature values ​​from a number of pre-constructed triples; wherein, any triple contains a feature value, a domain scenario and an explanatory statement, and the explanatory statement is used to represent the role and / or influence of the corresponding feature value in the domain scenario;

[0009] The second prompt word is input into the attribution report generation model to obtain an attribution report; wherein, the attribution report contains descriptive text describing the contribution of the plurality of first feature values ​​to the predicted click-through rate; the second prompt word contains a task instruction, the plurality of first feature values ​​and their corresponding first explanatory statements, and the task instruction instructs the generation of the contribution description text in combination with the first explanatory statements.

[0010] In one implementation, the step of performing data processing based on a model-interpretable algorithm includes:

[0011] The AIGC model is used to determine the average predicted click-through rate of the background dataset;

[0012] Using the background dataset, the average predicted click-through rate, the first prompt word, and the predicted click-through rate, data processing based on a model interpretable algorithm is performed to determine the degree to which the prediction deviates from the baseline for each of the several feature values, and the degree value is used as the contribution value of the corresponding feature value.

[0013] In one implementation, the feature values, domain scenarios, and explanatory statements in the plurality of triples are stored in a knowledge graph as nodes, and the corresponding nodes form triples based on the relationships between the nodes; the step of querying the first explanatory statements corresponding to the plurality of first feature values ​​includes:

[0014] For any given first feature value, retrieve the feature value node that matches the first feature value from the knowledge graph;

[0015] The first explanatory statement is obtained by identifying explanatory statement nodes from the knowledge graph that are associated with the feature value nodes through triples.

[0016] In one implementation, the knowledge graph is further improved in the following ways:

[0017] When the similarity of the feature values ​​between two feature value nodes is greater than the similarity threshold, a similarity relationship edge is established between the two feature value nodes.

[0018] When a specific relationship exists between two explanatory speech nodes, an edge representing the specific relationship is established between the two explanatory speech nodes; the specific relationship includes a progressive or complementary relationship.

[0019] When there is a hierarchical relationship between two domain scene nodes, an edge representing the hierarchical relationship is established between the two domain scene nodes.

[0020] In one implementation, after the step of determining the explanatory speech nodes that are associated with the feature value nodes in the knowledge graph through triples, the method further includes:

[0021] The first explanatory statement is obtained by searching the knowledge graph for other explanatory statement nodes that are associated with the triplet containing the feature value node.

[0022] In one implementation, the step of inputting the second prompt word into the attribution report generation model includes:

[0023] By engaging with users, we receive limiting information related to attribution reports from user input.

[0024] The second prompt containing the limiting information is input into the attribution report generation model; wherein the task instruction specifically instructs the generation of the contribution description text by combining the first explanatory phrase and the limiting information.

[0025] In one implementation, the method further includes: obtaining a domain knowledge base corresponding to the domain scenario in which the first explanatory phrase is located; inputting a second prompt word containing the domain knowledge base and the limiting information into an attribution report generation model; and the task instruction specifically instructs the generation of the contribution description text by combining the first explanatory phrase, the limiting information, and the domain knowledge base.

[0026] In one implementation, the method further includes:

[0027] A third prompt word is constructed based on the aforementioned first feature values ​​and the attribution report. The third prompt word is then input into the AIGC model to obtain new predicted material, which is then added to the utilization group.

[0028] A fourth cue word not obtained based on the aforementioned first feature values ​​and the attribution report is constructed. The fourth cue word is input into the AIGC model to obtain new predicted material, and the new predicted material is added to the exploration group.

[0029] The materials selected from the utilization group and the exploration group will be assigned to the application's display area for presentation.

[0030] In one implementation, the method further includes: collecting user interaction data on the materials displayed at the booth;

[0031] An online test report is generated based on the user interaction data; the AIGC model is then fine-tuned based on the attribution report and the online test report.

[0032] Secondly, the embodiment provides a material attribution device for an AIGC model, comprising:

[0033] The model prediction module is configured to input the first prompt word of the display material to be generated into the AIGC model, and generate the corresponding predicted material and its predicted click rate through the AIGC model based on the feature values ​​of several features contained in the first prompt word;

[0034] The model interpretation module is configured to perform data processing based on a model interpretable algorithm on the first prompt word and the predicted click-through rate to obtain the contribution values ​​of several feature values ​​to the predicted click-through rate.

[0035] The script query module is configured to query, from a pre-constructed set of triples, the first explanatory scripts corresponding to the first feature values ​​whose contribution values ​​are greater than a preset threshold among the aforementioned feature values; wherein, any triple contains a feature value, a domain scenario, and an explanatory script, and the explanatory script is used to represent the role and / or influence of the corresponding feature value in the domain scenario;

[0036] The material attribution module is configured to input the second prompt word into the attribution report generation model to obtain an attribution report; wherein, the attribution report contains descriptive text describing the contribution of the plurality of first feature values ​​to the predicted click-through rate; the second prompt word contains a task instruction, the plurality of first feature values ​​and their corresponding first explanatory statements, and the task instruction instructs the generation of the contribution description text in combination with the first explanatory statements.

[0037] Thirdly, the embodiments provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in any one of the first aspects.

[0038] Fourthly, an embodiment provides a computing device including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described in any one of the first aspects.

[0039] In the methods and apparatus provided in the embodiments of this specification, a triplet containing feature values, domain scenarios, and explanatory statements is pre-constructed. The explanatory statements can represent the role and / or impact of the feature values ​​in the domain scenarios. Through the attribution report generation model, combined with the first explanatory statements corresponding to high-contribution feature values ​​found from several triplets, a contribution description text of the high-contribution feature values ​​to the predicted click-through rate can be generated. This contribution description text enables operators to easily understand the shared impact of feature values ​​on the predicted click-through rate. Therefore, the embodiments can more effectively interpret the generation results of the AIGC model and obtain attribution reports that operators can understand. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0041] Figure 1 This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this application;

[0042] Figure 2 A flowchart illustrating a material attribution method for an AIGC model provided as an example;

[0043] Figure 3 A flowchart illustrating the interpretation of the AIGC model;

[0044] Figure 4 An example diagram of the structure of a knowledge graph corresponding to triples provided in this embodiment;

[0045] Figure 5 A flowchart illustrating the process of exploring and utilizing materials and fine-tuning the AIGC model;

[0046] Figure 6 This is a schematic block diagram of a material attribution device for an AIGC model provided in an embodiment. Detailed Implementation

[0047] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0048] Figure 1This is a schematic diagram illustrating an implementation scenario of one embodiment disclosed in this application. It includes a user device, a service platform, an AIGC model, and an attribution report generation model. An operator submits a user request (query) to the service platform via the user device to request the generation of materials (e.g., an image). For example, the query entered by the operator could be: "Please generate an image with the following characteristics: a luxury van parked next to a CBD building, a blue background with marketing text, and a sky and clouds in the background." This user request includes the feature values ​​of the image the operator needs to generate, such as the luxury van, the CBD building, the blue color, the marketing text, and the sky and clouds.

[0049] After receiving a query submitted by operations personnel, the service platform constructs prompts containing various feature values ​​based on the user query. These prompts are then input into the AIGC model, from which predicted creatives and predicted click-through rates (CTRs) are obtained. The predicted CTR is the expected click-through rate when the predicted creative is displayed in the application's showcase. The service platform packages the generated multiple predicted creatives into different formats and pushes them to the showcase for individual users to view or use. User devices can collect user interaction data from individual users, such as CTR, conversion rate, and dwell time. The service platform can aggregate user interaction data from multiple individual users to obtain data on the display effect of the predicted creatives in the showcase, thus creating a "creative-effect" dataset, which can be used for attribution analysis.

[0050] Based on the prompt word and predicted click-through rate, the service platform can interpret the model's prediction results, obtain the contribution values ​​of multiple feature values ​​in the prompt word, and identify the high-contribution feature values. Then, the service platform can use pre-constructed triples to determine the explanatory text corresponding to the high-contribution feature values.

[0051] The service platform can input high-contribution feature values ​​and definitive explanatory statements along with the aforementioned "material-effect" dataset into the attribution report generation model and obtain an attribution report. The attribution report can be a holistic analysis report for multiple predicted materials or an analysis report for a single predicted material.

[0052] Any triple contains a feature value, a domain context, and an explanatory statement, which represents the role and / or influence of the corresponding feature value in the domain context. The explanatory statement is expressed in natural language.

[0053] Figure 1The material attribution method provided in this application outlines an execution flow that can be implemented in one application scenario. In practical applications, this method can also be implemented in various other scenarios. For example, the AIGC model and attribution report generation model can be implemented on a service platform or on other computing devices outside the service platform. Triples can be stored as arrays in memory or on disk, or as knowledge graphs in a graph database (GDB). Figure 1 The execution flow shown can also be replaced by other actual execution flows. For example, material updates can be performed on a certain schedule, and attribution reports can also be executed on a certain schedule.

[0054] The following explains several concepts mentioned above. A service platform is a platform used to provide material generation services to users. In other implementation scenarios, materials can also be generated using computing devices other than the service platform, and assist the service platform in providing material generation services through data interaction. Operations personnel are business personnel who perform operational tasks related to the services provided by the service platform. The service platform and the computing devices mentioned above can be implemented through any device, equipment, platform, or cluster of devices with computing and processing capabilities. The concepts of model and materials involved in this application are explained below.

[0055] AIGC models are a class of advanced models that utilize artificial intelligence techniques to automatically generate multimodal content such as text, images, audio, and video. They are also supervised models that can be optimized for parameters. For example, the DALL-E model and the StableDiffusion model both belong to the AIGC model category. These models are typically trained on large-scale data and integrate multiple core algorithmic architectures, such as Generative Adversarial Networks (GANs), Diffusion Models, and Transformer architectures, or combinations thereof, to achieve high-quality and diverse content generation capabilities.

[0056] Large models are the core technological foundation of AIGC models. Current mainstream AIGC models (such as ChatGPT, MidJourney, and Sora) are almost all based on large models with a massive number of parameters (usually over a billion). AIGC is one of the important application scenarios for large models.

[0057] The full name of a large language model is Large Language Model (LLM). A large language model is a natural language processing model based on deep learning techniques, typically with billions to hundreds of billions or even more parameters, possessing powerful language understanding and generation capabilities. Large language models can employ the Transformer architecture or its variants (such as GPT, BERT, etc.), which utilizes an attention mechanism to globally model sequential data, efficiently handling long-distance dependencies and thus performing excellently in natural language tasks. Large language models learn the statistical features and semantic relationships of language through pre-training on large-scale corpora, giving them excellent generalization capabilities. The core capabilities of large language models include, but are not limited to: understanding contextual semantics, generating coherent and grammatically correct text, performing logical reasoning, and handling multi-task scenarios. Its usage typically includes two modes: direct inference and fine-tuning. In direct inference mode, the user guides the large language model to generate specific outputs by designing prompts. Prompts can be task descriptions or instructions in text form, used to stimulate the large language model's semantic understanding and generation capabilities. In fine-tuning mode, large language models are further trained on small-scale datasets within a specific domain to optimize their performance on specific tasks. The powerful generalization capabilities and flexibility of large language models make them an important tool in the field of artificial intelligence, providing efficient and accurate solutions for automated text generation and understanding.

[0058] The key technologies of large language models can be found in the detailed description in the paper "A Survey of Large Language Models" (paper number: arXiv:2303.18223v16, published on March 11, 2025), which will not be repeated here.

[0059] With technological advancements, some AIGC models have expanded their large language models to multimodal capabilities, forming Multimodal Large Language Models (MLLMs). MLLMs can simultaneously understand and generate data from multiple modalities, including text, images, and audio, enabling cross-modal information fusion and interaction. For example, MLLMs can generate descriptive text from images or generate corresponding images from text commands, significantly improving the naturalness and richness of human-computer interaction. This cross-modal capability is a key support for AIGC models to achieve collaborative generation of multiple content types.

[0060] The AIGC model, relying on various advanced AI architectures including large language models, can not only efficiently generate high-quality text, images, audio, and video content, but also apply this content to real-world scenarios such as news, advertising, and recommendation systems. For example, it can be used as exhibition material and further evaluated for click-through rates using predictive models. Therefore, the aforementioned material can be text, images, audio, or video.

[0061] AIGC models typically operate as a black box. As model performance improves, while providing accurate decisions, the rationale behind those decisions becomes increasingly difficult for operations personnel to understand, hindering their ability to effectively analyze the decisions. However, in many predictive scenarios involving content generation, the rationale behind the model's content generation is just as important as the generated content itself. Operations personnel often struggle to understand why the model generates specific content elements or copywriting styles.

[0062] It's understandable that although the predicted content is generated by the AIGC model based on the feature values ​​in the prompt words, the AIGC model does not strictly execute the instructions implied by each feature value like a programming language. It generates probabilistically based on statistical patterns learned from massive amounts of data. For example, some keywords may be ignored by the model, and some frequently co-occurring elements that are not explicitly mentioned may be automatically filled in.

[0063] This lack of interpretability of the model limits the effective evaluation and iterative optimization of the generated results by operators, especially in scenarios with specific business requirements for the materials (such as brand tone and regulatory compliance), which may lead to a deviation between the generated results and business objectives.

[0064] In practical applications, several model interpretation tools can be used to interpret the model's output data. For example, tools such as Shapley Additive exPlanations (SHAP), LIME, or Anchors can be used to interpret the model. These interpretation tools are model-independent post-processing tools that provide methods for measuring the contribution of a single feature vector to a single dimension of features.

[0065] Taking the SHAP tool as an example, its core objective is to quantify the contribution of each input model's feature value to the model's prediction results. For each specific prompt word, there is a difference between the AIGC model's predicted click-through rate and the baseline value. The contribution of each feature value in the prompt word to this difference is the interpretation result for that AIGC model. This contribution value is often expressed numerically, such as 0.3, 0.5, etc. While this interpretation result can help operations personnel understand the magnitude of the feature value's contribution to some extent, the specific content of its contribution remains incomprehensible to them.

[0066] To generate natural language attribution explanations that are easily understood by operations personnel, this application provides a material attribution method for the aforementioned AIGC model. This method utilizes pre-constructed triples containing feature values, domain context, and explanatory statements to identify explanatory statements with high-contribution feature values. Based on this, an attribution report generation model generates an attribution report described in natural language. This attribution method directly produces natural language attribution explanations that are easily understood by operations personnel, significantly reducing the comprehension barriers caused by the model's black box nature and promoting the deep application and trust of the model in business production systems.

[0067] The following is combined with Figure 2 The methods of the embodiments will be described in detail.

[0068] Figure 2 This is a flowchart illustrating a material attribution method for an AIGC model provided as an example. The method can be executed via a service platform or computing device. The method includes the following steps.

[0069] Step S210: Input the first prompt word prompt1 to be generated into the AIGC model. The AIGC model generates the corresponding predicted material and its predicted click-through rate based on the feature values ​​of several features contained in the first prompt word prompt1.

[0070] The first prompt word, prompt1, can be any prompt word used to generate the required content. This content can be text, an image, audio, or video, etc. This content is suitable for display in a booth on a specific page of an application. The predicted click-through rate (CTR) represents the predicted click-through rate when this content is displayed in the booth.

[0071] The first prompt word, prompt1, can be generated based on the user's input query, following a prompt template. The prompt template can contain task instructions, examples, and query placeholders. Filling the user's query into the query placeholders of this prompt template yields the first prompt word, prompt1. Several feature values ​​contained in the first prompt word, prompt1, are also feature values ​​contained in the user's query; these feature values ​​can exist in text form.

[0072] The AIGC model can be a model trained on a training dataset T. Where T = {(x1,y1),(x2,y2),…,(x…} N ,y N )}. Where, y i Standard values ​​representing the model's output, including creative assets and click-through rate. i =(x i 1,x i 2 ,…,x i D ) represents the eigenvector, x i k It is x i The k-th eigenvector, i∈{1,2,…,N}, k∈{1,2,…,D}.

[0073] The first prompt word, prompt1, can be one of the feature vectors in the set of feature vectors to be predicted, T′. T′ = {x1′, x2′, ..., x M M is the number of feature vectors in the set of feature vectors to be predicted, T′.

[0074] Step S220: For the first prompt word prompt1 and the predicted click-through rate, perform data processing based on the model interpretable algorithm to obtain the contribution values ​​of several feature values ​​to the predicted click-through rate.

[0075] The interpretable algorithm for the model can be SHAP, LIME, or Anchors, among others. Taking SHAP as an example, the contribution value represents the contribution of each feature value to the difference between the predicted click-through rate and the baseline value. Each feature value has a contribution value.

[0076] When performing data processing based on model-interpretable algorithms, the AIGC model can be used to determine the average predicted click-through rate of the background dataset. Using this average predicted click-through rate as a benchmark, and with the background dataset, the average predicted click-through rate, the first prompt word "prompt1", and the aforementioned predicted click-through rate, data processing based on model-interpretable algorithms can be performed to determine the degree to which several feature values ​​cause the prediction to deviate from the benchmark. This degree value is then used as the contribution value of the corresponding feature value.

[0077] The background dataset can contain multiple prompt words. Inputting each prompt word into the AIGC model yields its corresponding predicted probability and predicted click-through rate (CTR). The average of these predicted CTRs is then calculated as the average predicted CTR.

[0078] Using the background dataset, the average predicted click-through rate, the first prompt word "prompt1", and the aforementioned predicted click-through rate, the process of determining several feature values ​​that cause the prediction to deviate from the baseline can be performed according to the execution flow of the SHAP algorithm, and the specific details will not be elaborated here.

[0079] The set of feature vectors T′ to be predicted corresponds to the set of contribution values ​​S′:

[0080] S′={shap1′,shap1′,…,shap M ′} (1)

[0081] Among them, shap z ′ = {shap z ′ 1 , shap z ′ 2 , …, shap z ′ D} is the set of contribution values corresponding to the feature vector x z ′, and shap z ′ k is the contribution value of the k-th dimensional eigenvalue in the feature vector x z ′. z ∈ {1, 2, …, M}.

[0082] Figure 3 is a schematic flow diagram of a process for interpreting an AIGC model. Among them, the eigenvalues included in the first prompt prompt1 are: theme1, style1, color1, position1, condition1, text1, etc. After inputting the first prompt prompt1 into the AIGC model, a predicted image and a predicted click-through rate of 0.5 are obtained. It is known that the average predicted click-through rate obtained from the background dataset is 0.1, and the difference value between the predicted click-through rate of 0.5 and the average predicted click-through rate of 0.1 is 0.4. According to the above information, performing SHAP interpretation on the first prompt prompt1 can obtain the contribution values of each eigenvalue to this difference value as follows: theme1 is positive 0.2, style1 is positive 0.24, color1 is negative 0.1, position1 is negative 0.2, condition1 is positive 0.14, and text1 is positive 0.12, and their sum value is 0.4.

[0083] Using model-agnostic post-explainable methods such as SHAP can accurately quantify the contribution degree (i.e., Shapley value) of the feature single dimension of the material to be generated to the output result of the AIGC model.

[0084] Step S230, for a number of first eigenvalues whose contribution values are greater than a preset threshold among a number of eigenvalues, query a number of first interpretive statements respectively corresponding to the number of first eigenvalues from a number of pre-constructed triples.

[0085] For any feature vector x z ′, traverse the contribution value set shap z ′ of the feature vector x z ′, and add the eigenvalues corresponding to the contribution values greater than the preset threshold to the high-contribution eigenvalue set C z ′. That is, when the following formula is satisfied:

[0086] shap z ′ k > shap_threshold q (2)

[0087] The corresponding feature value feature_key_x k Add to set C z 'middle.

[0088] Any triple contains a feature value, a domain context, and explanatory language. The explanatory language is used to represent the role and / or impact of the corresponding feature value in the domain context and is expressed in natural language.

[0089] For example, in the marketing and advertising scenarios of car mortgage loans, the role or influence of luxury business vehicles (i.e., explanatory language) can be expressed as: they easily give people a sense of dignity and solemnity, as well as an impression of maturity and reliability, and are regarded as a manifestation of success, wealth and social status.

[0090] The set of triplet sequences can be represented as:

[0091] R = { <feature_key_x 1 ,domain q script 1q >, <feature_key_x 2 ,domain q script 2q >,…, <feature_key_x n ,domain q script nq >} (3)

[0092] Among them, feature_key_x n It is the nth eigenvalue, domain q It is the domain scenario q, script nq It is the explanatory language for the k-th dimension feature value in the domain scenario q.

[0093] Initially, triples can be constructed manually. The feature values, domain context, and explanatory statements in any triple are stored as arrays in memory or on disk.

[0094] As the number of triples increases, these manually constructed triples can be used to build a knowledge graph, within which new edge relationships can be established. When constructing the knowledge graph, feature values, domain scenarios, and explanatory statements can each be used as nodes, and triple relationships can be established between them. In other words, the constructed knowledge graph includes feature value nodes, domain scenario nodes, and explanatory statement nodes.

[0095] In addition, the nodes in the knowledge graph can be improved in at least one of the following ways: constructing new edges so that the edges between nodes in the knowledge graph are not limited to those that represent triple relationships, but can also include edges that represent other association relationships.

[0096] For any two feature value nodes, calculate the feature similarity between them. When the feature similarity between two feature value nodes is greater than a similarity threshold, establish a similarity relationship edge between the two feature value nodes. The similarity threshold can be preset and set to a relatively large threshold, thus making the establishment of similarity relationships between highly similar feature value nodes more flexible.

[0097] For any two feature value nodes, determine whether a specific relationship exists between the two explanatory phrase nodes. When a specific relationship exists between the two explanatory phrase nodes, establish an edge between them representing that specific relationship. This specific relationship can include progressive or complementary relationships.

[0098] For any two feature value nodes, determine whether a hierarchical relationship exists between the two domain scene nodes. When a hierarchical relationship exists between two domain scene nodes, establish an edge representing the hierarchical relationship between the two domain scene nodes.

[0099] Figure 4 This is an example diagram of a knowledge graph structure corresponding to triples provided in the embodiment. There are triple relationships between feature value 1, domain scenario 1, and explanatory statement 1; between feature value 1, domain scenario 2, and explanatory statement 2; and between feature value 2, domain scenario 1, and explanatory statement 3. Similarity edges are established between nodes corresponding to feature value 1 and feature value 2; hierarchical relationships are established between nodes corresponding to domain scenario 1 and domain scenario 2; and progressive or complementary relationships are established between explanatory statement 1 and explanatory statement 4. Figure 4 This is merely an example of the relationships between nodes and edges in a knowledge graph and does not constitute a limitation of this application.

[0100] When the triples are stored in memory or on disk as arrays, when querying the first explanatory statement, for any first feature value feature1, the first feature value feature1 is matched with the feature values ​​in several arrays, several matching arrays are determined based on the matching results, and the first explanatory statement corresponding to the first feature value feature1 is determined based on the several matching arrays.

[0101] For the set C containing the first eigenvalue z ', Traverse C z The eigenvalues ​​in ' are given when the following expression exists:

[0102] feature_key_x k ∈C z ′ (4)

[0103] Combine the triple sequence set R with feature_key_x k Corresponding explanatory script kq Add the set of eigenvectors x z The first set of explanatory rhetoric corresponding to ′ z ′.

[0104] Matching between feature values ​​can be done using the vectors corresponding to those feature values. These vectors can be the embedding vectors of the text. The matching array is the array containing the feature values ​​when the similarity between two feature values ​​is greater than a threshold.

[0105] When determining the first explanatory statement, the explanatory statements in several matching arrays can be directly used as the first explanatory statement of the first feature value feature1, or the explanatory statements in the specified domain scenarios in several matching arrays can be used as the first explanatory statement of the first feature value feature1.

[0106] The specified domain scenario can be the current business scenario, such as the marketing and advertising scenario for vehicle mortgage loans.

[0107] When triples are stored in the form of a knowledge graph, when querying the first explanatory statement, for any first feature value feature1, the feature value node that matches the first feature value feature1 is retrieved from the knowledge graph, and the explanatory statement node that has a triple association with the feature value node is determined from the knowledge graph. The explanatory statement corresponding to the explanatory statement node is taken as the first explanatory statement.

[0108] When determining the feature value node that matches the first feature value 1 from the knowledge graph, each feature value node in the knowledge graph can be matched with the first feature value 1 based on similarity, and the feature value node with a similarity greater than the threshold can be used as the matched feature value node.

[0109] After that, in order to obtain more and richer first explanatory statements, other explanatory statement nodes that are related to the nodes in the above triples can be found in the knowledge graph, and the explanatory statements corresponding to the other explanatory statement nodes can be used as the first explanatory statements.

[0110] For example, with Figure 4For example, suppose feature value 1 is a feature value node that matches the first feature value 1, and feature value 1-domain scenario 1-explanatory statement 1 and feature value 1-domain scenario 2-explanatory statement 2 are two triples that are associated with feature value 1. Next, other explanatory statement nodes that are associated with these two triples can be found. Here, other explanatory statement nodes can be explanatory statement 4 that has a progressive relationship with explanatory statement 1, or explanatory statement 3 in the triple containing feature value 2 (i.e., feature value 2-domain scenario 1-explanatory statement 3) that has a similar relationship with feature value 1.

[0111] Step S240: Input the second prompt word prompt2 into the attribution report generation model to obtain attribution report R1.

[0112] The attribution report R1 can be generated for a set of predicted materials or for a single predicted material. Taking a single predicted material as an example, the attribution report R1 may contain descriptive text describing the contribution of several first feature values ​​from the first prompt word to the predicted click-through rate. This contribution description text is natural language text. The second prompt word prompt2 contains a task instruction, several first feature values, and their corresponding first explanatory statements. This task instruction instructs the attribution report R1 generation model to combine the first explanatory statements to generate the corresponding contribution description text.

[0113] When constructing the second prompt word "prompt2", it can be based on the corresponding prompt template. This prompt template can contain task instruction text, feature values, and explanatory phrase placeholders. Filling these placeholders with several first feature values ​​and their corresponding first explanatory phrases yields the second prompt word "prompt2". The second prompt word "prompt2" can also include the display performance data of the predicted material. Display performance data can include the material's click-through rate, conversion rate, and dwell time, etc.

[0114] The attribution report generation model can be implemented using a large model or a retrieval-enhanced RAG model. The attribution report generation model is used to determine the contribution description text for each first feature value based on the second prompt word prompt2, and then generate an attribution report R1 from several contribution description texts.

[0115] In practical applications, the contribution values ​​of each first feature value can also be added to the second prompt word, prompt2. Here, the task instruction instructs the generation of corresponding contribution description text by combining the contribution values ​​of each first feature value with the first explanatory phrase.

[0116] The attribution report generation model can comprehensively analyze the textual contribution descriptions of each high-contribution feature based on the contribution value of each first feature and the first explanatory statement. Such contribution description text is material attribution content that operations personnel can understand.

[0117] Retrieval-Augmented Generation (RAG) models, which combine a retrieval unit and a generative unit, are among the most popular and cutting-edge technologies for large-scale modeling. They combine language modeling and information retrieval techniques. Specifically, when a model needs to generate text or answer a question, it first retrieves relevant information from a large document collection. This retrieved information then guides text generation, improving the quality and accuracy of predictions. By referencing information from external knowledge bases, it helps large models generate answers or content, offering strong interpretability and customization capabilities. It is suitable for various natural language processing tasks, including question-answering systems, document generation, and intelligent assistants.

[0118] When using the retrieval-enhanced generative model to implement step S240, relevant information can be retrieved from the knowledge base corresponding to the relevant domain scenario through the retrieval tool, and the relevant information can be added as context to the second prompt word prompt2. The obtained second prompt word prompt2 is then input into the generative model, and the attribution report R1 is obtained through the generative model.

[0119] During step S240, a dialogue can also be conducted with the operations personnel. Through this dialogue, the service platform can receive restrictive information related to the attribution report R1, input by the operations personnel. This restrictive information may include limitations on the domain scenario and / or limitations on the format of the attribution report R1. In other words, through user dialogue, the platform can determine in which domain scenario the operations personnel need to interpret the model, thereby filtering out the first explanatory statements from a large pool of first explanatory statements within the specified domain scenario.

[0120] Next, the service platform can add limiting information to the second prompt word "prompt2" and input the resulting second prompt word "prompt2" into the attribution report generation model. At this point, the specific task instructions in the second prompt word "prompt2" are combined with the first explanatory phrase and limiting information to generate contribution description text. In this way, the attribution report generation model can analyze the contribution of each first feature value by combining the first explanatory phrase and limiting information, and obtain the contribution description text corresponding to each first feature value.

[0121] Taking the marketing and advertising scenario of car mortgage loans as an example, the primary characteristic values ​​and their corresponding contribution values ​​include: luxury business vehicle (0.3), blue background (0.2), next to CBD building (0.2), and xx marketing and advertising text (0.15). Furthermore, in the marketing and advertising scenario of car mortgage loans, the explanatory language for luxury business vehicles is: it easily evokes a sense of prestige, solemnity, and maturity, and is seen as a manifestation of success, wealth, and social status; the explanatory language for blue background is: it can create a calm, credible, and professional sense of financial trust; the explanatory language for CBD building is: it clearly points to a business setting, enhancing the rationality of the loan purpose; and the explanatory language for xx marketing and advertising text is: it revolves around promoting financial services and avoids information confusion.

[0122] The above information is used to construct the second prompt word prompt2, which is then input into the attribution report generation model. The attribution report R1 can be obtained through the attribution report generation model, as shown in Table 1.

[0123] Table 1

[0124]

[0125] The attribution weights are the weights that the model summarizes for each first feature value, and they are not necessarily the same as their contribution values. The contribution description text has been summarized and analyzed holistically by the model, and is obtained with the contribution values ​​of each first feature value, the first explanatory statement, and other information as context.

[0126] After obtaining the attribution report R1, the retrieval enhancement model can also display the attribution report R1 to the user as a response to the dialogue.

[0127] Step S240 can also generate an attribution report R1 for multiple predicted materials. That is, steps S210–S230 can be executed multiple times as an execution unit to obtain high-contribution feature values ​​and first explanatory phrases corresponding to multiple predicted materials. When executing step S240, the relevant data from multiple predicted materials can be combined to construct a second prompt word, prompt2. This relevant data can include task instruction text, high-contribution feature values ​​and first explanatory phrases for each predicted material, limiting information, and display performance data for multiple predicted materials. The attribution report R1 can contain contribution description text for a single predicted material, or it can contain an overall analysis and description of multiple predicted materials; for example, high-conversion materials commonly include the keyword "limited-time offer."

[0128] In one specific implementation, the domain scenario in the triple can also be replaced by a domain knowledge base. This domain knowledge base contains a large number of relevant knowledge documents for that domain.

[0129] Alternatively, a relationship can be established between the domain context and the domain knowledge base in the triplet. Once each first explanatory phrase is determined, the domain knowledge base corresponding to the domain context in which that first explanatory phrase is located can be obtained. The second prompt word prompt2, which contains the domain knowledge base and limiting information, is then input into the attribution report generation model.

[0130] Here, the task instruction specifically instructs the generation of contribution description text by combining the first explanatory statement, limiting information, and the domain knowledge base. The domain knowledge base contains richer background knowledge related to the domain scenario. By using the domain knowledge base as the dialogue context input to the attribution report generation model, the model can refer to the rich knowledge in the domain knowledge base, thereby summarizing a more reasonable contribution description text.

[0131] In this embodiment, the predicted content is generated by the AIGC model based on various feature values. However, the AIGC model does not strictly execute the implicit instructions of each feature value. Model interpretability tools can quantitatively determine the contribution of each feature value to the predicted click-through rate exceeding the baseline value, but isolated numerical values ​​do not provide rich, easily understandable explanatory content. In the method of this embodiment, the attribution report generation model combines high-contribution feature values ​​with their explanatory terminology within the domain context. Through comprehensive analysis, it obtains contribution description text for each high-contribution feature value. This contribution description text contains rich content that operations personnel can understand. Therefore, this embodiment can provide better explanation of the model and improve the perceptibility, acceptability, and iterative capabilities of the model results in the field of content generation.

[0132] The model interpretation method (such as SHAP) used in the implementation examples is post-processed and model-independent. This design completely decouples model selection from interpretation capabilities, allowing for the free selection of more complex models (such as large AIGC models) with stronger representational capabilities and better suited to material generation tasks, without sacrificing model performance for interpretability. This ensures that the model maintains strong representational capabilities and a high ceiling for generation effects even in complex business data scenarios.

[0133] Once the attribution report R1 fully understands why the model generates such materials, the high-contribution feature values ​​can be fully utilized based on the attribution report R1 to generate higher-quality materials, thereby increasing the click-through rate of the booth materials.

[0134] In another embodiment of this application, the process of exploring and utilizing the materials and fine-tuning the AIGC model can continue. For example, after seeing the attribution report R1, the operator can initiate a request on their user device to optimize material generation. This request can carry instruction information, which may be an adjustment direction based on the attribution report R1, or it may directly accept the suggestions in the attribution report R1. The service platform can generate optimized prompt words based on the request and input the optimized prompt words into the AIGC model to obtain a new round of optimized predicted materials. The service platform can then repackage the optimized predicted materials and push them to the booth for individual users to view or use.

[0135] Figure 5 This is a flowchart illustrating the exploration and utilization of materials and the fine-tuning of the AIGC model. It includes the following steps 1-6.

[0136] Step 1: Construct a third prompt word (prompt3) based on several first feature values ​​and attribution report R1. Input the third prompt word (prompt3) into the AIGC model to obtain new prediction materials, and add the new prediction materials to the utilization group. The third prompt word (prompt3) is the optimized prompt word.

[0137] Among these, several first feature values ​​are high-contribution feature values. The attribution report R1 contains the attribution weights of each high-contribution feature value. Feature values ​​can be selected from each first feature value based on the attribution weights and the contribution values ​​of each first feature value to construct the third prompt word prompt3. The third prompt word prompt3 can be constructed using the same prompt template as the first prompt word prompt1.

[0138] The material used in this step was generated under the guidance of high contribution feature values ​​and attribution report R1, and has a good potential for exposure and click-through rate.

[0139] Step 2: Construct a fourth prompt word (prompt4) that is not based on several first feature values ​​and attribution report R1. Input the fourth prompt word (prompt4) into the AIGC model to obtain new prediction material, and add the new prediction material to the exploration group.

[0140] The fourth prompt word, prompt4, in this step was generated without the guidance of high-contribution feature values ​​and attribution report R1. The resulting exploration group serves as the control group for the utilization group. The construction process of the fourth prompt word, prompt4, can be the same as that of the first prompt word, prompt1, using the same prompt template. The fourth prompt word, prompt4, along with the third prompt word, prompt3, and the first prompt word, prompt1, are all prompt words used to generate presentation materials; their difference lies in the feature values ​​they contain.

[0141] In practical applications, the total number of materials in the utilization group and the exploration group is set, for example, it can be M. The number of materials in each group can be controlled using the ε-Greedy algorithm. Specifically, the ratio of materials in the utilization group and the exploration group can be determined by the following formula:

[0142]

[0143] Where π1 is the proportion of materials in the corresponding group, e represents the material, and p v Representative booth, |E exploit | is an estimated value based on the effect after the group goes live, |E explore | represents the estimated effect after the exploration group goes online, and ε is the pre-set exploration probability, usually set to 0.1 to 0.3.

[0144] Using the formula (5) above, the corresponding π values ​​for the utilization group and the exploration group are a and b, respectively. Therefore, the material quantity ratio for the utilization group is a / (a+b), and the material quantity is K = aM / (a+b). The material quantity ratio for the exploration group is b / (a+b), and the material quantity is L = bM / (a+b). Thus, the following utilization group and exploration group are obtained respectively:

[0145] Using group E exploit ={e1,e2,…,e K}

[0146] Exploration Group E explore ={e K+1 ,e K+2 ,…,e K+L}(6)

[0147] Where K+L≤M. For the specific execution process, please refer to the Explore-Exploit strategy.

[0148] Step 3 involves assigning the materials selected from the Utilization and Exploration groups to the application's showcase for display.

[0149] From using group E exploit and Exploration Group E explore When selecting materials, you can optimize materials for each group, choosing the best materials for display. During material optimization, a ranking model can be used to sort the materials within a group. Ranking models such as learn2Rank can be used. The processing of each material by the ranking model can be represented as follows:

[0150]

[0151] Among them, g(e i ) represents material ei Based on the predicted quality scores, the ranking results of each material can be obtained from the model's output as follows:

[0152] E exploit ranked =(e (1) ,e (2) ,…,e (K) (8)

[0153] Among them, e (i) This represents the i-th element in the sorted sequence, satisfying g(e (1) )≥g(e (2) )≥…≥g(e (K) ).

[0154] The materials in the exploration group can also be sorted in the same way.

[0155] Next, based on feature label groups, diversity recall can be performed on the exploitation group and the exploration group. A label group set B = {b1, b2, ..., b...} can be predefined. Q}. Among them, b q This represents the set of materials labeled 'q'. When the materials are images, the labels can be image style, saturation, or other category tags. Taking a group as an example, the subset of recalled materials obtained for each group is as follows:

[0156]

[0157] Here, tag(e) represents the set of feature tags for material e. The same method can be used to obtain the subset of recalled materials for each group in the exploration group. The total number of recalled materials can be determined based on the total number of booths.

[0158] The specific recall process can refer to existing diversity recall algorithms, such as the Determinantal Diversity Promotion (DDP) algorithm and the Maximum Marginal Relevance (MMR) algorithm.

[0159] When the recalled material subsets for the exploit and explore groups are obtained, A / B test booths can be assigned using algorithms such as the Upper Confidence Bound (UCB) algorithm. The UCB algorithm can also be replaced by algorithms such as the Linear Upper Confidence Bound (LinUCB) algorithm or the Diversity-Promoting Thompson Sampling (DP-TS) algorithm to optimize the exploration strategy. Here, the booth set is P = {p1, p2, ..., p...}v When allocating materials to each booth, the allocation probability of each material e can be determined using the following formula:

[0160]

[0161] Among them, UCB(e,p) v The formula for calculating ) is:

[0162]

[0163] Where β is a pre-defined hyperparameter, and the first term on the right side of equation (11) is CTR. e The first term is the utilization term, representing the historical average click-through rate of material e. If the material has never been displayed, this term is 0. The second term is the exploration term. c is a constant, usually taken as c = 2. 1 / 2 Or a value between 1 and 2, used to control the exploration intensity. e Material e is at booth p v The historical exposure count on the page, where 't' represents all materials displayed on booth p. v The total number of times the content is exposed. For a new source material, its UCB value is infinite.

[0164] The above formula can be used to determine the distribution probability of each material in each booth, and the recalled material e can be assigned to the corresponding booth according to the distribution probability.

[0165] Step 4: After a period of display, collect user interaction data on the materials displayed at the booths.

[0166] Step 5: Generate online test report R2 based on user interaction data.

[0167] Step 6: Fine-tune the AIGC model based on the attribution report R1 and the online test report R2.

[0168] When fine-tuning the AIGC model, you can adjust all parameters within a set small range, or you can adjust only some parameters.

[0169] User interaction data can include click counts, conversion counts, user dwell time, and impression counts. The collected user interaction data can be represented as: D = {e i ,p v ,c i,v ,v i,v ,t i,v User interaction data can also be called display effect data. Among them, e i Let i be the i-th material, and i belongs to the set b of materials with the tag q. q p vThis is the v-th booth, with the subscript v indicating the v-th booth. i,v v represents the number of clicks. i,v t represents the number of transformations. i,v This refers to the duration of stay.

[0170] The online test report R2 may include the following:

[0171] R2={CTR q CVR q ,Δg q} q＝1～Q (12)

[0172] Among them, CTR q =∑c i,v / ∑imp i,v CVR q =∑v i,v / ∑imp i,v , Δg q =g(e) post -g(e) pre imp i,v For the number of exposures, Δg q This represents the score difference of the ranking model before and after online testing. ∑imp i,v The material with the tag 'q' represents the collection b. q Exposure counts and values ​​across all booths. ∑c i,v This represents the total number of clicks and values ​​for the content tagged with "q" across all display booths. ∑v i,v This represents the number and value of conversions for the product tag "q" across all display booths. q CVR represents the average click-through rate (CVR) after a specific recall creative is launched. q This represents the average exposure conversion rate after a particular recall material is launched.

[0173] In other words, the content of the online test report R2 may include, but is not limited to, the average exposure click-through rate, average exposure conversion rate, and the difference in scores of the ranking model before and after the recall materials went live.

[0174] In this embodiment, the AIGC model is fine-tuned using attribution report R1 and online test report R2, which can improve the accuracy of the AIGC model and generate higher quality display materials.

[0175] This embodiment deeply integrates the exploration-exploitation (EE) balance strategy into the entire process of material generation and testing. Through a systematic EE framework, it achieves multi-dimensional, guided exploration (using historical knowledge to guide the exploitation team and unleashing the creativity of the AIGC model). By dynamically allocating exploration budgets and adjusting exploration strategies based on feedback (such as confidence interval terms in UCB), it achieves in-depth exploration, more efficiently finding potential high-value points in the material space. The embodiment also establishes a closed-loop feedback mechanism: the R2 metric in the online test report directly quantifies the effectiveness of the explored materials and is used to update the ranking model (Learn2Rank) and adjust the exploration strategy (UCB parameters), forming a continuous co-evolutionary closed loop of interpretation → exploration → verification → optimization. This systematically solves the problems of single-dimensional exploration, insufficient depth, and broken closed loops, enabling the service platform to continuously discover, verify, and utilize new high-value material patterns in a dynamic environment.

[0176] It is understood that before or during the use of the technical solutions in the various embodiments of this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner in accordance with relevant laws and regulations, and user authorization will be obtained. Users here include individual users and operators.

[0177] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to choose, based on the prompt message, whether to provide personal information to the software or hardware such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution. Personal information includes click actions and dwell time, etc.

[0178] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0179] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0180] In this manual, the terms "first," "first feature value," and "first explanatory phrase," as well as the corresponding "second" (if present) in the text, are used merely for ease of distinction and description and do not have any limiting meaning.

[0181] The foregoing description describes specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than those shown in the embodiments, and the desired result may still be achieved. Furthermore, the processes depicted in the drawings do not necessarily need to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0182] Figure 6 This is a schematic block diagram of a material attribution device for an AIGC model provided in an embodiment. This device embodiment is related to... Figure 2 The method embodiment shown corresponds to this. The device 600, deployed in a computing device, includes:

[0183] The model prediction module 610 is configured to input the first prompt word of the display material to be generated into the AIGC model, and generate the corresponding predicted material and its predicted click rate through the AIGC model based on the feature values ​​of several features contained in the first prompt word.

[0184] The model interpretation module 620 is configured to perform data processing based on a model interpretable algorithm on the first prompt word and the predicted click-through rate to obtain the contribution values ​​of several feature values ​​to the predicted click-through rate.

[0185] The script query module 630 is configured to query, from a pre-constructed set of triples, the first explanatory scripts corresponding to the first feature values ​​whose contribution values ​​are greater than a preset threshold among the plurality of feature values; wherein, any triple contains a feature value, a domain scenario and an explanatory script, and the explanatory script is used to represent the role and / or influence of the corresponding feature value in the domain scenario;

[0186] The material attribution module 640 is configured to input the second prompt word into the attribution report generation model to obtain an attribution report; wherein, the attribution report contains descriptive text describing the contribution of the plurality of first feature values ​​to the predicted click-through rate; the second prompt word contains a task instruction, the plurality of first feature values ​​and their corresponding first explanatory statements, and the task instruction instructs the generation of the contribution description text in combination with the first explanatory statements.

[0187] In one implementation, the model interpretation module 620 includes a determination submodule 21 and an interpretation submodule 22. The determination submodule 21 is configured to determine the average predicted click-through rate (CTR) of the background dataset using the AIGC model. The interpretation submodule 22 is configured to perform data processing based on a model interpretable algorithm using the background dataset, the average predicted CTR, the first prompt word, and the predicted CTR to determine the degree to which the prediction deviates from the baseline for each of the plurality of feature values, and to use this degree value as the contribution value of the corresponding feature value.

[0188] In one implementation, the feature values, domain scenarios, and explanatory statements in several triples are stored as nodes in a knowledge graph, and the corresponding nodes form triples based on the relationships between them. The statement query module 630 includes a retrieval submodule 31 and an interpretation submodule 32. The retrieval submodule 31 is configured to retrieve feature value nodes matching any first feature value from the knowledge graph. The interpretation submodule 32 is configured to determine explanatory statement nodes in the knowledge graph that have triple associations with the feature value node, thereby obtaining the first explanatory statement.

[0189] In one implementation, the device 600 further includes a knowledge graph improvement module 650, configured to improve the knowledge graph in the following ways: when the feature value similarity between two feature value nodes is greater than a similarity threshold, a similarity relationship edge is established between the two feature value nodes; when there is a specific association between two explanatory speech nodes, an edge representing the specific association is established between the two explanatory speech nodes; the specific association includes a progressive or complementary relationship; when there is a hierarchical relationship between two domain scene nodes, an edge representing the hierarchical relationship is established between the two domain scene nodes.

[0190] In one implementation, the interpretation submodule 32 is specifically configured as follows: after determining the interpretive script node that has a triplet association with the feature value node from the knowledge graph, it also searches for other interpretive script nodes that have a triplet association with the feature value node from the knowledge graph to obtain the first interpretive script.

[0191] In one implementation, the material attribution module 640 includes a receiving submodule 41 and an attribution factor module 42. The receiving submodule 41 is configured to receive limiting information related to the attribution report input by the user through a dialogue. The attribution factor module 42 is configured to input a second prompt containing the limiting information into the attribution report generation model. Specifically, the task instruction instructs the generation of the contribution description text by combining the first explanatory phrase and the limiting information.

[0192] In one implementation, the material attribution module 640 further includes: an acquisition submodule 43, configured to acquire a domain knowledge base corresponding to the domain scenario where the first explanatory phrase is located; and an attribution module 42, configured to input a second prompt word containing the domain knowledge base and the limiting information into the attribution report generation model; the task instruction specifically instructs the generation of the contribution description text by combining the first explanatory phrase, the limiting information, and the domain knowledge base.

[0193] In one implementation, the device 600 further includes: a utilization group generation module 660, an exploration group generation module 670, and a material display module 680. The utilization group generation module 660 is configured to construct a third prompt word based on the plurality of first feature values ​​and the attribution report, input the third prompt word into the AIGC model to obtain new predicted material, and add the new predicted material to the utilization group. The exploration group generation module 670 is configured to construct a fourth prompt word not obtained based on the plurality of first feature values ​​and the attribution report, input the fourth prompt word into the AIGC model to obtain new predicted material, and add the new predicted material to the exploration group. The material display module 680 is configured to allocate materials selected from the utilization group and the exploration group to the application's display area for display.

[0194] In one implementation, the device 600 further includes: a data collection module 710, a report generation module 720, and a model fine-tuning module 730. The data collection module 710 is configured to collect user interaction data related to the materials displayed at the booth. The report generation module 720 is configured to generate an online test report based on the user interaction data. The model fine-tuning module 730 is configured to fine-tune the AIGC model based on the attribution report and the online test report.

[0195] The above-described apparatus embodiments correspond to the method embodiments, and detailed descriptions can be found in the description of the method embodiments section, which will not be repeated here. The apparatus embodiments are derived based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments; detailed descriptions can be found in the corresponding method embodiments.

[0196] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform... Figures 1 to 5 Any one of the methods described.

[0197] This specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement... Figures 1 to 5 Any one of the methods described.

[0198] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for storage media and computing devices are basically similar to the method embodiments, so they are described more simply; relevant parts can be referred to the descriptions of the method embodiments.

[0199] Those skilled in the art will recognize that the functions described in the embodiments of the present invention in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0200] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made based on the technical solutions of the present invention should be included within the scope of protection of the present invention.

Claims

1. A material attribution method for an AIGC (Artificial Intelligence Generated Content) model, comprising: The first prompt word to be generated for display materials is input into the AIGC model, and the AIGC model generates corresponding predicted materials and their predicted click-through rates based on the feature values ​​of several features contained in the first prompt word. For the first prompt word and the predicted click-through rate, perform data processing based on a model-interpretable algorithm to obtain the contribution values ​​of several feature values ​​to the predicted click-through rate; For several first feature values ​​whose contribution values ​​are greater than a preset threshold, query the first explanatory statements corresponding to the several first feature values ​​from several pre-constructed triples; wherein, any triple contains a feature value, a domain scenario and an explanatory statement, and the explanatory statement is used to represent the role and / or influence of the corresponding feature value in the domain scenario; The second prompt word is input into the attribution report generation model to obtain an attribution report; wherein, the attribution report contains descriptive text describing the contribution of the plurality of first feature values ​​to the predicted click-through rate; the second prompt word contains a task instruction, the plurality of first feature values ​​and their corresponding first explanatory statements, and the task instruction instructs the generation of the contribution description text in combination with the first explanatory statements.

2. The method according to claim 1, wherein the step of performing data processing based on a model-interpretable algorithm comprises: The AIGC model is used to determine the average predicted click-through rate of the background dataset; Using the background dataset, the average predicted click-through rate, the first prompt word, and the predicted click-through rate, data processing based on a model interpretable algorithm is performed to determine the degree to which the prediction deviates from the baseline for each of the several feature values, and the degree value is used as the contribution value of the corresponding feature value.

3. The method according to claim 1, wherein the feature values, domain scenarios and explanatory statements in the plurality of triples are stored in the knowledge graph in the form of nodes, and the corresponding nodes form triples by the association relationship between the nodes; The step of querying the first explanatory statements corresponding to the plurality of first feature values ​​includes: For any given first feature value, retrieve the feature value node that matches the first feature value from the knowledge graph; The first explanatory statement is obtained by identifying explanatory statement nodes that are associated with the feature value nodes through triples from the knowledge graph.

4. According to claim 3, the knowledge graph is further improved in the following ways: When the similarity of the feature values ​​between two feature value nodes is greater than the similarity threshold, a similarity relationship edge is established between the two feature value nodes. When a specific relationship exists between two explanatory speech nodes, an edge representing the specific relationship is established between the two explanatory speech nodes; the specific relationship includes a progressive or complementary relationship. When there is a hierarchical relationship between two domain scene nodes, an edge representing the hierarchical relationship is established between the two domain scene nodes.

5. The method according to claim 4, further comprising, after the step of determining the explanatory speech node associated with the feature value node in the knowledge graph through a triplet, the method includes: The first explanatory statement is obtained by searching the knowledge graph for other explanatory statement nodes that are associated with the triplet containing the feature value node.

6. The method according to claim 1, wherein the step of inputting the second prompt word into the attribution report generation model comprises: By engaging in dialogue with users, we receive limiting information related to attribution reports from user input. The second prompt containing the limiting information is input into the attribution report generation model; wherein the task instruction specifically instructs the generation of the contribution description text by combining the first explanatory phrase and the limiting information.

7. The method of claim 6, further comprising: Obtain a domain knowledge base corresponding to the domain scenario in which the first explanatory statement is located, and input a second prompt word containing the domain knowledge base and the limiting information into the attribution report generation model; the task instruction specifically instructs the generation of the contribution description text by combining the first explanatory statement, the limiting information and the domain knowledge base.

8. The method according to claim 1, further comprising: A third prompt word is constructed based on the aforementioned first feature values ​​and the attribution report. The third prompt word is then input into the AIGC model to obtain new predicted material, which is then added to the utilization group. A fourth cue word not obtained based on the aforementioned first feature values ​​and the attribution report is constructed. The fourth cue word is input into the AIGC model to obtain new predicted material, and the new predicted material is added to the exploration group. The materials selected from the utilization group and the exploration group will be assigned to the application's display area for presentation.

9. The method of claim 8, further comprising: Collect user interaction data regarding the materials displayed at the booth; Generate an online test report based on the user interaction data; The AIGC model was fine-tuned based on the attribution report and the online test report.

10. A material attribution device for an AIGC model, comprising: The model prediction module is configured to input the first prompt word of the display material to be generated into the AIGC model, and generate the corresponding predicted material and its predicted click rate through the AIGC model based on the feature values ​​of several features contained in the first prompt word; The model interpretation module is configured to perform data processing based on a model interpretable algorithm on the first prompt word and the predicted click-through rate to obtain the contribution values ​​of several feature values ​​to the predicted click-through rate. The script query module is configured to query, from a pre-constructed set of triples, the first explanatory scripts corresponding to the first feature values ​​whose contribution values ​​are greater than a preset threshold among the aforementioned feature values; wherein, any triple contains a feature value, a domain scenario, and an explanatory script, and the explanatory script is used to represent the role and / or influence of the corresponding feature value in the domain scenario; The material attribution module is configured to input the second prompt word into the attribution report generation model to obtain an attribution report; wherein, the attribution report contains descriptive text describing the contribution of the plurality of first feature values ​​to the predicted click-through rate; the second prompt word contains a task instruction, the plurality of first feature values ​​and their corresponding first explanatory statements, and the task instruction instructs the generation of the contribution description text in combination with the first explanatory statements.

11. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-9.

12. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-9.

Images