Data processing method and electronic equipment

Abstract

The invention discloses a data processing method and electronic equipment, and the method comprises the steps: in response to obtaining target input data inputted to a target application, identifying a field category to which the target input data belongs and a represented user intention, the field category being determined at least based on the attribute of target content in the target input data; based on the field categories and the user intention, calling at least one processing model to generate the target input data, a calling strategy of the processing model being related to the number of the field categories; and processing the reasoning result of the at least one processing model into a target response result for the target input data.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data processing method and electronic equipment. Background Art

[0002] At present, when dealing with complex and diverse practical problems, intelligent systems often find it difficult to provide satisfactory answers to users, resulting in a poor user experience. Summary of the Invention

[0003] The technical solutions provided in this application are as follows:

[0004] The first aspect of the present application provides a data processing method, comprising:

[0005] In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intent represented, wherein the domain category is determined based on at least an attribute of target content in the target input data;

[0006] Based on the domain category and the user intention, calling at least one processing model to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories;

[0007] The inference result of the at least one processing model is processed into a target response result for the target input data.

[0008] The generating and processing the target input data by calling at least one processing model based on the domain category and the user intention includes at least one of the following:

[0009] In the case of identifying that the domain category is unique, calling a first processing model that matches the domain category from a target model library based on the user intention to generate and process the target input data;

[0010] In the case of recognizing that the domain category is not unique, calling a plurality of processing models corresponding to the non-unique domain categories from a target model library based on the user intention to generate and process the target input data;

[0011] The target model library is deployed on a local end of the electronic device or on a target processing device connected to the electronic device.

[0012] The step of calling a plurality of processing models corresponding to the non-unique domain categories to generate and process the target input data includes:

[0013] Determining contribution weights of the multiple processing models based on the user intent and / or target reference data, wherein the target reference data can represent the contribution degree of the content of each attribute in the target input data to the user intent;

[0014] Based on the contribution weights, the multiple processing models are called in parallel or serially to generate and process the attribute contents of the target input data respectively; or

[0015] The plurality of processing models are called based on the contribution weights to process the target input data into fused input data which is input into a second processing model for generation processing.

[0016] Determining the contribution weights of the multiple processing models based on the user intention and / or target reference data includes at least one of the following:

[0017] Determining contribution weights of the multiple processing models based on a contribution ratio of each attribute content in the target input data to the user intention;

[0018] Determining contribution weights of the plurality of processing models based on the degree of association between each attribute content and the context in the target input data;

[0019] Determining contribution weights of the multiple processing models based on the attribute categories and proportions of the attribute contents in the target input data;

[0020] The contribution weights of the multiple processing models are determined based on the degree of matching between the attribute content in the target input data and the user portrait information of the target user.

[0021] The identifying the domain category to which the target input data belongs and the user intention represented includes:

[0022] In a case where the target input data includes multiple data of different modalities, performing unified format processing on the target input data;

[0023] Perform semantic analysis on target input data in a unified format to determine the user intent represented;

[0024] The domain category to which the target input data belongs is determined according to the category attribute and / or semantic attribute of the target content and the context of the target input data.

[0025] Processing the inference result of the at least one processing model into a target response result for the target input data includes at least one of the following:

[0026] Performing weighted fusion processing on the respective inference results based on the contribution weights corresponding to the multiple processing models to obtain a target response result for the target input data;

[0027] Obtain user portrait data of the target user, and based on the user portrait data, process the inference results of multiple processing models into target feedback content and output it to the target feedback interface, wherein the user portrait data at least includes interaction data between the target user and the at least one processing model.

[0028] Also includes:

[0029] When feedback data of the target user regarding the target response result is obtained, generating corresponding first positive evaluation content based on the feedback data;

[0030] The first positive evaluation content is different from the second positive evaluation content generated for the target input data, and the first positive evaluation content includes target guidance information.

[0031] Also includes:

[0032] Obtaining interaction data between a target user and the target application, and updating the knowledge graph data of the target user based on the interaction data;

[0033] The target application is an application that can call the at least one processing model to execute a target processing function, and the interaction data includes data as input to the processing model and feedback data for the inference result output by the processing model.

[0034] Also includes:

[0035] When feedback data of the target user regarding the target response result is obtained, a corresponding target feedback interface is generated based on the feedback data, and the target feedback interface provides operation controls or recommended content for updating the knowledge graph data of the target user.

[0036] Another aspect of the present application provides an electronic device, comprising at least one processor and at least one processing model capable of running on the processor, wherein the processing model can be called by a target application to perform at least one of the following:

[0037] In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intent represented, wherein the domain category is determined based on at least an attribute of target content in the target input data;

[0038] Based on the domain category and the user intention, calling at least one processing model to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories;

[0039] The inference result of the at least one processing model is processed into a target response result for the target input data. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the originals and elements are not necessarily drawn to scale.

[0041] Figure 1 A flowchart of a data processing method provided in Example 1 of the present application;

[0042] Figure 2 A flowchart of a data processing method provided in Example 6 of the present application;

[0043] Figure 3 A flowchart of a data processing method provided in Example 8 of the present application;

[0044] Figure 4 A flowchart of a data processing method provided in Example 9 of the present application;

[0045] Figure 5 A flowchart of a data processing method provided in Example 10 of the present application;

[0046] Figure 6 A schematic diagram of a target feedback interface provided by this application;

[0047] Figure 7 A schematic diagram of the structure of a data processing device provided in this application. DETAILED DESCRIPTION

[0048] The following describes the embodiments of the present application in conjunction with the accompanying drawings. The terms used in the implementation methods of the present application are only used to explain the specific embodiments of the present application and are not intended to limit the present application.

[0049] The embodiments of the present application are described below in conjunction with the accompanying drawings. Those skilled in the art will appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0050] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, and this is merely a way of distinguishing the objects of the same attributes when describing them in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0051] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0052] Reference Figure 1 , is a flow chart of a data processing method provided in Example 1 of the present application, such as Figure 1 As shown, the method may include but is not limited to the following steps:

[0053] Step S101 : In response to obtaining target input data input into a target application, identifying the domain category to which the target input data belongs and the user intention represented.

[0054] In this embodiment, the target input data may include but is not limited to: question data and / or other interaction data.

[0055] Question data can be understood as a collection of information containing questions to be answered or knowledge points to be explored. Question data can be of different types, such as learning, daily life, or professional consultation.

[0056] For example, the target input data may include: an entire Chinese test paper (i.e., one implementation method of question data). In this Chinese test paper, in the modern Chinese reading comprehension section, the selected article may be a narrative text about the impact of scientific development on literary creation. It not only involves the examination of Chinese reading skills such as the main theme and writing techniques of the article, but also implies an understanding of the basic knowledge of science as a subject. Students are required to use basic knowledge in the field of science to assist in understanding the content in the text about how science changes the way of literary creation.

[0057] In the ancient poetry appreciation section, there may be poems written with historical events as the background. When appreciating, students must not only grasp the artistic conception, emotions, expression techniques and other language knowledge of the poems, but also have a certain understanding of the relevant historical events in order to appreciate the connotation of the poems more deeply.

[0058] The composition topic may also be interdisciplinary in nature. For example, the topic of "Looking at the Development of Human Civilization from the Perspective of Biological Evolution" requires students to not only have solid written expression and in-depth thinking skills, but also have a certain understanding of biological evolution as a biological knowledge, and to discuss it in connection with the development of human civilization. It comprehensively examines students' mastery of Chinese language knowledge and their ability to apply it comprehensively, as well as their ability to transfer and apply interdisciplinary knowledge.

[0059] Alternatively, the target input data may include: a modern Chinese reading comprehension section in a Chinese language test paper or one of the sub-questions in the modern Chinese reading comprehension section (ie, one implementation of question data).

[0060] For example, daily life type question data may include: when decorating a home, a user may ask "how to choose the appropriate wall color and lighting layout according to the room area and lighting conditions."

[0061] For example, professional consultation type question data may include: scientific researchers may ask questions about the preparation method of a new material, the selection of data analysis models, etc.

[0062] In this embodiment, other interaction data can be used to perform information interaction with the system.

[0063] For example, the target input data may include: a voice command issued by a user, "Help me plan a weekend trip that combines historical and cultural exploration with outdoor sports, requiring visits to places with ancient buildings and suitable for hiking, and providing introductions to attractions, transportation routes, and hiking equipment recommendations" (i.e., an implementation method of other interactive data).

[0064] Alternatively, the target input data may include: the user uploads a "medical image of an artificial joint designed in accordance with biomechanical principles", hoping that the system will identify whether the joint structure meets the biomechanical requirements of the human body and provide diagnostic suggestions (i.e., an implementation method of other interactive data).

[0065] Alternatively, the target input data may include: the entire Chinese test paper and the user's voice instruction to "summarize the knowledge points involved in the test paper and explain the key and difficult knowledge points in detail" (i.e., an implementation method of question data and other interactive data).

[0066] In this embodiment, the domain category may be determined based at least on the attribute of the target content in the target input data.

[0067] Field categories can include subject categories or application areas. Subject categories include, but are not limited to, language, history, geography, physics, chemistry, and other subjects. Application areas can cover a variety of vertical fields, such as architecture, civil engineering, economics, politics, biomedicine, urban development and construction, the automotive industry, the PC industry, renewable energy, ESG, and law.

[0068] User intent can at least reflect the purpose of the user's interaction, such as calculation, concept explanation, problem consultation, homework correction, etc., whether the purpose of the interaction is to obtain problem-solving ideas and process, or to obtain personalized recommendations, industry reports, personalized images, personalized text descriptions, application codes, etc.

[0069] Step S102: Based on the domain category and the user intention, at least one processing model is called to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories.

[0070] In this embodiment, a domain category may correspond to only one processing model, but the model has multiple functions and can meet the data processing needs of different users in the domain.

[0071] In most cases, a domain category can correspond to multiple processing models, each of which focuses on specific functions or tasks within the domain to achieve more professional and efficient data processing.

[0072] If the target input data belongs to only one domain category, the calling strategy can be used to directly call the processing model corresponding to the domain category.

[0073] If the target input data belongs to multiple domain categories, the calling strategy can be used to comprehensively consider each domain category and call the corresponding processing model for each domain category to meet the needs of multi-domain data processing.

[0074] Step S103: Process the inference result of the at least one processing model into a target response result for the target input data.

[0075] In this embodiment, the method of generating the target response result may vary according to the number of called processing models.

[0076] If only one processing model is called, the target response result can directly adopt the inference result of the model.

[0077] When multiple processing models are invoked, the target response is generated differently than when a single model is used. Because multiple processing models process the target input data from different domain categories, the inference results of a single processing model are not directly used. Instead, the inference results of each processing model are combined to form a comprehensive and accurate target response.

[0078] In this embodiment, in response to obtaining target input data inputted into a target application, the domain category to which the target input data belongs and the user intent represented can be identified. The domain category is determined based at least on the attributes of the target content in the target input data, enabling the target input data to be accurately categorized into corresponding domains, such as language, mathematics, medicine, architecture, and other subject areas, or application areas such as economics, politics, and environmental protection.

[0079] After accurately identifying the domain category and user intent, at least one processing model can be flexibly called to generate and process the target input data based on the domain category and user intent. This calling method can avoid deviation from user expectations during the reasoning process, while fully utilizing the professional advantages of the processing model corresponding to the domain category to achieve precise processing for specific fields, and then process the reasoning results of the at least one processing model into a target response result for the target input data, ensuring that the target response result is more comprehensive and professional, and improving the user experience.

[0080] As another optional embodiment of the present application, a data processing method is provided in Example 2 of the present application. This embodiment is mainly an implementation of the above step S102, and may specifically include but is not limited to at least one of the following:

[0081] Step S1021: When it is identified that the domain category is unique, a first processing model matching the domain category is called from a target model library based on the user intention to generate and process the target input data.

[0082] In this embodiment, the target model library may include processing models corresponding to various domain categories. One domain type may correspond to one processing model or multiple processing models.

[0083] The target model library can be deployed locally on the electronic device. When a user asks a question, the local processing model can quickly process the target input data without relying on an external network connection. This is suitable for scenarios with high requirements for data privacy and response speed, such as sensitive data analysis and processing within some enterprises.

[0084] The target model library can also be deployed on a target processing device connected to the electronic device. The target processing device can be a cloud server or other edge computing device. Deploying the target model library on the target processing device can leverage its powerful computing power and rich data resources to process more complex and larger amounts of data.

[0085] In this embodiment, if the target input data is identified as belonging to a unique domain category, it can be considered that the user's question focuses on a specific discipline or industry field. For example, the user's question only involves the discipline of history, such as inquiring about the specific process of a historical event, the life story of a historical figure, or exploring the political, economic, and cultural characteristics of a historical period.

[0086] Alternatively, the questions raised by users may only involve physics, for example, asking about the principles of a certain physical law, the operating steps and result analysis of a physical experiment, or exploring the application of physical phenomena in real life.

[0087] Or, the questions raised by users only involve the PC industry, such as troubleshooting computer hardware, optimizing and upgrading software systems; analyzing market trends and competitive situations in the PC industry, etc.

[0088] Alternatively, the questions raised by users may only involve the biomedical field, for example, consulting on treatment drugs for a certain disease.

[0089] If the domain category corresponds to only one processing model, the user's intent can determine the direction and method of data processing by the first processing model. For example, if the target input data belongs to the biomedical field and the user's intent is to perform disease diagnosis, the model will use its built-in disease diagnosis algorithms and knowledge base to analyze and process the data. If the user's intent is to process drug development data, the model can use its built-in drug development-related data processing algorithms to process and analyze drug development-related experimental data, clinical trial data, and other data.

[0090] If the domain category corresponds to multiple processing models, the user intent can be used to select the first processing model from the multiple processing models. For example, taking the biomedical field as an example, this field may include specialized processing models for disease diagnosis, processing models focused on drug development data, models for gene sequence analysis, and processing models for drug interaction prediction.

[0091] If the user's intention is to assist doctors in diagnosing a disease, the disease diagnosis model can be called. This model analyzes the input patient symptoms, test results and other data based on a large amount of medical case data and diagnostic rules, and gives possible disease diagnosis results and suggestions; if the user's intention is to evaluate the research and development prospects of a new drug, the processing model for drug research and development data can be called. This processing model can conduct a comprehensive analysis of the drug's chemical structure, pharmacodynamic data, pharmacokinetic data, etc., to provide decision support for drug research and development.

[0092] In this embodiment, when it is recognized that the domain category is unique, the first processing model matching the domain category is called from the target model library based on the user intention to generate and process the target input data. This ensures that the first processing model matching the domain category does not deviate from the user's expectations during the generation process, thereby providing the user with a more professional and accurate answer.

[0093] Step S1022: When it is identified that the domain category is not unique, based on the user intention, multiple processing models corresponding to the non-unique domain categories are called from the target model library to generate and process the target input data.

[0094] If the domain categories are not unique, the user's question can be considered complex, involving multiple disciplines or application areas. For example, a user might ask, "Analyze the forces acting on athletes' joints during running from a biomechanical perspective, and combine this with sports medicine knowledge to provide recommendations for preventing sports injuries." This question involves both biomechanics, requiring analysis of the mechanical structure and force principles of joints, and sports medicine, requiring the development of methods for preventing sports injuries based on the biomechanical analysis results.

[0095] Alternatively, the question a user asks might be, “Why did the Tang Dynasty become such a powerful empire?” This question involves both the disciplines of history and economics.

[0096] In this embodiment, the multiple processing models invoked may correspond one-to-one to the non-unique domain categories. For example, if the non-unique domain categories include history and economics, two processing models may be invoked: one corresponding to history and one corresponding to economics.

[0097] In this embodiment, the method of calling the processing model of each field category can refer to the relevant introduction in step S1021, which will not be repeated here.

[0098] In this embodiment, by identifying that the domain category is not unique, calling multiple processing models corresponding to the non-unique domain categories from the target model library based on the user intention to generate and process the target input data, it is possible to avoid deviation from user expectations during the reasoning process, and at the same time achieve cross-domain category collaborative processing, fully utilize the professional advantages of the processing models corresponding to different domain categories, and ensure that the target response results are more comprehensive and professional.

[0099] As another optional embodiment of the present application, a data processing method is provided in Example 3 of the present application. This embodiment is mainly an implementation method of calling multiple processing models corresponding to the non-unique domain categories in the above step S1022 to generate and process the target input data. Specifically, it may include but is not limited to the following steps:

[0100] Step S11: Determine contribution weights of the multiple processing models based on the user intention and / or target reference data.

[0101] In some complex application scenarios, user intent can often focus on multiple disciplines or application areas. However, the role and importance of different disciplines or application areas in solving the problem vary. For example, for the user question "Why did the Tang Dynasty become a powerful empire?", multiple processing models could include: a processing model for history and a processing model for economics. The history processing model, with its rich historical materials and research methods, can deeply analyze the development of the Tang Dynasty from multiple perspectives, such as politics, culture, military affairs, and diplomacy, thereby revealing the fundamental reasons for the Tang Dynasty's prosperity. Therefore, it can be assigned a higher contribution weight for this question.

[0102] Models in the economic application field can analyze the economic development of the Tang Dynasty, such as the advancement of agricultural production, the development of handicrafts, the prosperity of commerce, and the buoyancy of foreign trade. These economic factors did provide the material foundation for the prosperity of the Tang Dynasty, but they were more of an important aspect of historical development, operating within the broader framework of the history discipline. Therefore, their contribution weight is lower than that assigned to corresponding models in the history discipline.

[0103] The target reference data can represent the contribution degree of the content of each attribute in the target input data to the user intention.

[0104] The processing model corresponding to the content of attributes with high contribution should play a greater role in the overall data processing process, that is, obtain a higher contribution weight; conversely, the contribution weight of the processing model corresponding to the content of attributes with low contribution should be relatively low.

[0105] Step S12: Based on the contribution weights, the multiple processing models are called to generate and process the attribute contents of the target input data in parallel or in series.

[0106] In this embodiment, the first prompt information generated based on the target input data and the contribution weight of each processing model can be input to each processing model, and each processing model can generate processing on the target input data according to its contribution weight.

[0107] Of course, corresponding attributes and related content data can also be extracted from the target input data based on the contribution weight of each processing model, and input into each processing model, which will then generate and process the corresponding attributes and related content data in the target input data.

[0108] In this embodiment, when the contextual relationship between the various attribute contents in the target input data is weak, or the processing processes of the various processing models do not affect each other, a parallel processing method can be used. This method can fully utilize computing resources, improve processing efficiency, and shorten the overall processing time. For example, in the question "Why did the Tang Dynasty become a powerful empire?", the history discipline processing model and the economic field processing model can be processed in parallel. The history discipline processing model focuses on analyzing the development process of the Tang Dynasty from the dimensions of politics, culture, etc., while the economic field processing model focuses on analyzing the economic development status of the Tang Dynasty. Both perform data processing simultaneously.

[0109] In this embodiment, when the contextual relationships between the attributes in the target input data are close and subsequent processing depends on the results of previous processing, a serial processing approach can be used. For example, when analyzing "free fall motion and parabolic trajectory," physical principles are fundamental, and mathematical processing relies on accurate understanding of physical principles. Therefore, the processing model corresponding to the physics subject should be invoked first, followed by the processing model corresponding to the mathematics subject.

[0110] In this embodiment, by determining the contribution weights and reasonably selecting parallel or serial processing methods, the professional advantages of different processing models can be fully utilized, and efficient use of computing resources and improvement of processing efficiency can be achieved while ensuring the quality of the target response results.

[0111] As another optional embodiment of the present application, a data processing method is provided in Example 4 of the present application. This embodiment is mainly an implementation method of calling multiple processing models corresponding to the non-unique domain categories in the above step S1022 to generate and process the target input data. Specifically, it may include but is not limited to the following steps:

[0112] Step S21: Determine contribution weights of the multiple processing models based on the user intention and / or target reference data.

[0113] The detailed process of step S21 can be found in the relevant introduction of the above step S11, which will not be repeated here.

[0114] Step S22: Based on the contribution weights, the multiple processing models are called to process the target input data into fused input data, which is then input into a second processing model for generation processing.

[0115] In this embodiment, each processing model can re-interpret and process the target input data according to its own contribution weight in the overall processing process, and obtain input data to be used that can more comprehensively reflect the characteristics of the target input data.

[0116] The input data generated by each processing model is fused to produce fused input data. For example, if the user asks, "Why did the Tang Dynasty become a powerful empire?", the processing model for history has a higher contribution weight, so the information generated by it on the political system and cultural development of the Tang Dynasty (i.e., one implementation of the input data) can be presented in more detail in the fused input data. Furthermore, the analysis of the Tang Dynasty economy generated by the processing model for economics (i.e., one implementation of the input data) can serve as supplementary information in the fused input data.

[0117] The second processing model may include but is not limited to: a general generative model; or a large professional domain model formed by fine-tuning the general generative model through domain data, such as a medical diagnosis model, a legal document generation model, etc.

[0118] If it is a large model in a professional field, when fusing the input data to be used generated by each processing model, the conversion relationship between the large model in the professional field and the processing model that generates the input data to be used can be considered to ensure that the fused input data can meet the input specifications of the large model in the professional field without losing data accuracy.

[0119] In this embodiment, each processing model reinterprets the target input data based on its own contribution weight, generating ready-to-use input data that more comprehensively reflects the characteristics of the target input data. This ready-to-use input data is then fused, integrating information from different fields. This fused input data can describe and analyze the target input data from multiple perspectives, providing a richer source of information for the second processing model, enabling it to more fully understand the problem and generate more accurate and professional target responses.

[0120] As another optional embodiment of the present application, a data processing method is provided in Example 5 of the present application. This embodiment is mainly an implementation method of determining the contribution weights of the multiple processing models based on the user intent and / or target reference data, and may specifically include but is not limited to at least one of the following:

[0121] Step S31: Determine the contribution weights of the multiple processing models based on the contribution ratio of each attribute content in the target input data to the user intention.

[0122] The contribution ratio can reflect the relative importance of the attribute content in the target input data. The higher the contribution ratio, the more critical the attribute content is to achieving user intention.

[0123] The processing model corresponding to the attribute content with a high contribution ratio should play a greater role in the overall data processing process, that is, obtain a higher contribution weight; conversely, the processing model corresponding to the attribute content with a low contribution ratio has a relatively low contribution weight.

[0124] For example, the questions raised by the user may include: analysis of the free fall motion and parabolic trajectory of an object. The multiple processing models may include: a processing model corresponding to mathematics and a processing model corresponding to physics.

[0125] The processing model for mathematics is primarily responsible for handling attributes related to mathematical formula derivation, function calculations, and coordinate system establishment. For example, when analyzing parabolic trajectories, it is necessary to use quadratic function formulas to describe the trajectory of an object and calculate the object's position coordinates at different times.

[0126] The processing model for physics focuses on processing attributes related to physical principles, calculation of physical quantities, and interpretation of experimental data. For example, the calculation and analysis of physical quantities such as gravitational acceleration, initial velocity, and time in free fall.

[0127] By analyzing user questions, we can determine that physics-related content (e.g., "free fall") in the target input data can contribute up to 70% to the user's intent (accurately analyzing the free fall and parabolic trajectory). Mathematical content (e.g., "parabolic trajectory"), which primarily quantifies and accurately describes physical principles, is less important and may contribute only 30% to the user's intent.

[0128] Accordingly, the contribution weight of the processing model corresponding to the physics discipline is higher than the contribution weight of the processing model corresponding to the mathematics discipline.

[0129] In this embodiment, by determining the contribution ratio of each attribute to user intent, we can accurately identify the most critical components of the target input data for achieving user intent. Key attributes often directly relate to the core of user needs. Giving them a higher contribution weight in the corresponding processing model ensures that this critical information is more fully processed and analyzed, thereby improving the alignment of the final processing results with user intent and providing users with more accurate and valuable information.

[0130] Step S32: Determine the contribution weights of the multiple processing models based on the degree of association between each attribute content in the target input data and the context.

[0131] In this embodiment, the context can be understood as background content such as historical interaction information with each attribute content in the target input data, previously discussed issues, etc.

[0132] If the context is all about problems or interactions in a specific discipline or application field, then the attribute content related to the discipline or application field may be more important in the current processing, and the corresponding processing model should have a higher contribution weight.

[0133] For example, suppose a user has previously been discussing physics problems, such as mechanical principles and electromagnetic phenomena, on a self-learning and feedback platform (i.e., one implementation of the target application), thus establishing a physics context. Now, if the user enters new target input data, "Analysis of free-fall motion and parabolic trajectory," the content with the physics category attribute (e.g., free-fall motion) is more closely associated with the context. Therefore, the corresponding physics processing model should play a more critical role in processing the current target input data, and its contribution weight should be higher.

[0134] On the contrary, the content with the category attribute of mathematics (such as parabolic trajectory) has a relatively low degree of relevance to the context, and the corresponding mathematics subject processing model should have a lower contribution weight when processing the current target input data.

[0135] In this embodiment, context can reflect the user's previous interactions and discussions. Attributes highly relevant to the context often have greater significance in the current processing. Determining the contribution weight of the processing model based on this relevance ensures that the data processing process remains consistent with the user's interaction history, better understanding the user's current needs, avoiding processing results that deviate from user expectations due to ignoring context, and ultimately improving user satisfaction and trust in the target application.

[0136] Step S33: Determine the contribution weights of the multiple processing models based on the attribute categories and proportions of the attribute contents in the target input data.

[0137] In this embodiment, a contribution weight can be assigned to each attribute category. For example, in comprehensive problem solving, language skills may be considered important for understanding and expression, so the language attribute category is assigned a contribution weight of 25%. Mathematical skills are critical for logical analysis and calculation, so the mathematics attribute category is assigned a contribution weight of 30%. Physics knowledge plays a prominent role in some problems involving natural phenomena and engineering technology, so the physics attribute category is assigned a contribution weight of 40%. Other attribute categories may be relatively less important and have a corresponding contribution weight of 5%.

[0138] In this embodiment, the proportion of each attribute category in the target input data can also be counted, and then the contribution weight of the processing model can be assigned according to this proportion. If a certain attribute category appears frequently in the target input data and accounts for a large proportion of the content, then the corresponding processing model should have a higher contribution weight.

[0139] Of course, the contribution weight corresponding to the attribute category of the attribute content and the contribution weight corresponding to the proportion can be weighted to obtain the contribution weight of the processing model.

[0140] In this embodiment, the contribution weights assigned to each attribute category provide a basic reference framework for data processing. This pre-set weight distribution method can reasonably divide different attribute categories based on their general importance to problem solving, ensuring the efficiency and accuracy of the processing model.

[0141] Furthermore, by calculating the proportion of each attribute category in the target input data, the contribution weight of the processing model is assigned, allowing for dynamic adjustment based on the actual data. If a certain attribute category accounts for a large proportion of the data, it indicates that its content is more important to the current problem. Giving its corresponding processing model a higher weight can make the data processing more tailored to the characteristics of the actual data, improving the accuracy and effectiveness of the processing.

[0142] Step S34: Determine the contribution weights of the multiple processing models based on the degree of matching between the attribute content in the target input data and the user portrait information of the target user.

[0143] User profile information encompasses various characteristics of the target user, such as strengths and weaknesses in academic subjects, learning styles, interests, and hobbies. By comparing the degree of match between the attributes in the target input data and the user profile information, we can understand the user's ability level in the areas corresponding to these attributes. If the user is highly competent in this area (i.e., the degree of match is high), the weight of the processing model corresponding to this attribute can be appropriately weakened, as the user may already possess the relevant knowledge and does not need to rely too much on model processing. If the user is less competent in this area (i.e., the degree of match is low), the weight of the processing model corresponding to this attribute should be increased to provide more targeted help and support to the user.

[0144] For example, suppose the target user is a student. His or her user portrait information shows that the student has excellent grades in mathematics and often wins awards in mathematics competitions, indicating that mathematics is a strong subject; while his or her grades in Chinese are average, and his or her reading comprehension and writing skills need to be improved, indicating that Chinese is a weak subject.

[0145] At this time, the user inputs the target input data "solve a difficult math problem and analyze the central idea of ​​an ancient text".

[0146] The attribute contents are difficult mathematical problems and analysis of the central ideas of ancient Chinese texts, corresponding to the subjects of mathematics and Chinese language respectively.

[0147] The difficult mathematical problems match the advantageous subjects in the user profile with a high degree of matching.

[0148] The analysis of the central idea of ​​the ancient text matches the weak subjects in the user portrait, and the matching degree is low.

[0149] Since the user has strong mathematical ability, the contribution weight of the mathematical processing model can be appropriately reduced, for example, from the preset 30% to 20%.

[0150] Due to the user's weak language ability, the contribution weight of the language processing model can be appropriately increased, for example, from the preset 25% to 35%. The weights of other subject processing models are adjusted accordingly based on actual conditions to ensure that the total is 100%.

[0151] In this embodiment, user profile information includes multiple characteristics, including the user's strengths and weaknesses, learning style, and interests. By comparing the degree of match between each attribute and the user profile information, personalized data processing solutions can be provided for each user. For areas where the user excels, the corresponding processing model's weight is appropriately weakened to reduce unnecessary repetitive processing; for areas where the user is weak, the corresponding processing model's weight is increased to provide more assistance and support, thereby improving the pertinence and practicality of data processing.

[0152] This personalized data processing method enables users to feel the system's attention and understanding of them, meets their personalized needs, thereby increasing their participation and loyalty to the system and promoting long-term interaction and cooperation between users and the system.

[0153] As another optional embodiment of the present application, refer to Figure 2 , is a flow chart of a data processing method provided in Example 6 of the present application. This embodiment is mainly an implementation method of the above step S101. Figure 2 As shown, step S101 may include but is not limited to the following steps:

[0154] Step S1011 : In response to obtaining target input data input to a target application, if the target input data includes multiple data in different modalities, performing unified format processing on the target input data.

[0155] When users interact with the target application, they can choose from a variety of input methods to meet user needs in different scenarios (such as answering questions or submitting learning content). Specifically, they can include:

[0156] Text input: Users input content in natural language to answer questions. Whether it is a short reply or a long description, the target application can parse and understand different forms of text expression.

[0157] Image input: Users can upload image data such as handwritten problems and diagrams (such as geometric figures and flow charts). For example, a student can upload a photo of a math problem they wrote in their workbook, or upload an image containing complex geometric figures. The target application can then receive and process this image information.

[0158] Formula input: Supports users to input complex mathematical formulas using standard mathematical input methods (such as LaTeX).

[0159] Audio input: Users can submit information through voice input. For example, in language learning scenarios, users can read aloud an English article. The target application can receive the audio data and then perform speech recognition on the audio, converting it into text for grammatical analysis and semantic understanding. In meeting recording scenarios, users can also directly record the meeting audio, and the application will process the audio and extract key information.

[0160] Video input: Users can upload video files. For example, in a physics experiment course, students can record and upload a video of themselves conducting a physics experiment. The target application can then analyze the video frame by frame, identifying key steps, operating techniques, and experimental phenomena. In the education field, the application can also extract and analyze the teaching content of the teaching videos uploaded by teachers, assisting in the organization and optimization of teaching resources.

[0161] For data of different modalities, unified format processing methods may include:

[0162] Text processing: Perform grammatical analysis, keyword extraction, and semantic understanding on text input based on a multimodal large language model (LLM).

[0163] Multimodal LLM has powerful natural language processing capabilities and can accurately analyze the structure and semantic information of text, thereby providing accurate and comprehensive information for subsequent analysis. Compared with traditional methods, it can better handle complex contexts and diverse text expressions.

[0164] Image processing: Multimodal LLM is used to recognize text in images (such as handwritten homework content), while image analysis technology is used to interpret mathematical symbols, tables, and charts in images.

[0165] Formula recognition: Using the mathematical formula recognition technology of multimodal LLM, complex mathematical formulas are identified and converted into standardized expressions to facilitate subsequent calculations and analysis.

[0166] Audio processing: First, use speech recognition technology to convert audio data into text, and then perform subsequent operations according to text processing methods, such as grammatical analysis and semantic understanding, to extract valuable information from the audio.

[0167] Video processing: Analyze the video frame by frame, extract key frame images, extract the audio part of the video for audio processing, and use image recognition technology to identify and process the key frame images. Combine the image and audio processing results to fully extract the effective information in the video.

[0168] Step S1012: Perform semantic analysis on the target input data in a unified format to determine the represented user intention.

[0169] In this embodiment, it is possible but not limited to: the second prompt information generated based on the target input data in a unified format is input into a third processing model (such as a multimodal LLM), and the third processing model can process it to generate the represented user intention.

[0170] User intent reflects the purpose of a user's interaction, such as calculation, concept explanation, or homework grading. Different user intents can determine the type of results a user desires, such as problem-solving ideas and procedures, personalized recommendations, industry reports, personalized images, personalized text descriptions, or application code.

[0171] At the same time, user intent can also represent the complexity of processing the target input data. For example, if a user asks the system to fully interpret a complex scientific research paper and provide innovative research ideas, this is much more complex than simply asking for the explanation of a simple concept.

[0172] Step S1013: Determine the domain category to which the target input data belongs according to the category attribute and / or semantic attribute of the target content in the target input data and the context of the target input data.

[0173] In this embodiment, the category attribute may be determined by, but is not limited to, at least one of the following methods:

[0174] Step S41: Based on professional knowledge and experience, the target content in the target input data is analyzed and judged to determine its category attributes. For example, based on a subject knowledge base covering various disciplines, the input content is classified into the subject category, knowledge domain, etc., to determine whether it belongs to a subject category such as mathematics, physics, or chemistry, or a more specific sub-field.

[0175] Step S42: Determine the category attributes of the target content by extracting the category features of the target content according to the machine learning model.

[0176] In this embodiment, the semantic attributes of the target content may be determined by, but are not limited to, at least one of the following methods:

[0177] Step S51: Extract semantic information from the text using techniques such as part-of-speech tagging, named entity recognition, and semantic role tagging to determine the semantic attributes of words or concepts. For example, part-of-speech tagging can determine whether a word is a noun, verb, or adjective, thereby understanding its semantic role in the sentence.

[0178] Step S52: The information in the knowledge graph can be used to determine the semantic attributes of entities and concepts in the target input data. For example, searching for the entity "apple" in the knowledge graph can obtain its related semantic attributes, such as "fruit", "red", "round", etc.

[0179] Step S53: For some complex semantic attributes, which are difficult to accurately obtain through automated technology, manual annotation may be required. Manual annotation can combine professional knowledge and contextual understanding to determine semantic attributes more accurately.

[0180] Of course, the attributes of target content can also include emotional attributes. For example, target input data could include a voice command such as, "Every time I see solar panels converting sunlight into electricity, I find it truly magical. I'd love to learn more about the applications of solar photovoltaic systems in home settings!" This voice command clearly conveys the user's positive sentiment towards renewable energy, which can be applied to the renewable energy sector.

[0181] In this embodiment, when the category attribute of the target content is very clear, the domain category can be determined directly based on the category attribute. For example, if the target content is classified as algebra in the mathematics discipline after analysis, then the domain category of the target input data can be determined to be algebra in the mathematics discipline.

[0182] In the case of multiple overlapping categories, the weight and relevance of the attributes of each category should be comprehensively considered. For example, if the target content involves both mechanics in physics and chemical reaction principles in chemistry, but the mechanics component accounts for a larger and more core area, then the domain category can be determined as mechanics within the physics field. If the proportion of mechanics and chemical reaction principles is not much different, the domain category can be determined as mechanics within physics and chemical reaction principles within chemistry.

[0183] In this embodiment, semantic attributes can provide more detailed information about the target content, helping to further determine its domain category. For example, after determining that the target content is in the science field, semantic attribute analysis reveals that it contains a large number of concepts such as biological evolution and genetic variation. In this case, its domain category can be more accurately determined as evolutionary biology within the biology field.

[0184] Semantic attributes can play a key role in identifying target content whose categorization is unclear. For example, if the target content lacks clear disciplinary characteristics, but semantic attribute analysis reveals frequent occurrences of psychology-related concepts such as "emotion," "cognition," and "behavior," its domain category can be determined to be psychology.

[0185] In this embodiment, the contextual information of the target input data includes the data's source, usage scenario, preceding and following context, etc. For example, if the target input data comes from a medical research forum, and the preceding text discusses the diagnosis of a certain disease, then even if the target content itself does not have clear category attributes or semantic attributes pointing to the medical field, the contextual information can be combined to determine that the field category it belongs to is the medical field.

[0186] In complex situations, it's helpful to consider a combination of categorical attributes, semantic attributes, and contextual information. For example, if the target content involves the research and development of a new material, the categorical attributes suggest it may fall within the field of materials science, while the semantic attributes include some concepts related to electronic engineering. Furthermore, the contextual information indicates that the material will be used in the manufacture of electronic devices. Combining these factors, we can determine that the field category is at the intersection of materials science and electronic engineering, specifically, electronic materials.

[0187] In this embodiment, unified formatting transforms data from different modalities into a form that the system can uniformly recognize and process. This eliminates the barriers caused by data format differences and provides broader input compatibility for target applications. Regardless of the user's chosen input method, the application can smoothly receive and process data, thus meeting the diverse needs of users in different scenarios.

[0188] Furthermore, by performing semantic analysis on target input data in a unified format, the target application can gain a deeper understanding of the user's true needs, rather than just superficially analyzing the input content. This precise understanding enables the application to provide responses and results that better meet the user's expectations, improving user experience and satisfaction.

[0189] In addition, category attributes can provide a basic basis for classification, semantic attributes further refine the understanding of the target content, and contextual information takes into account the usage scenarios and background of the data. Comprehensive consideration of the category attributes and / or semantic attributes of the target content and the contextual information avoids classification bias that may be caused by a single factor, and can more comprehensively and accurately determine the domain category to which the target input data belongs.

[0190] As another optional embodiment of the present application, a data processing method is provided in Example 7 of the present application. This embodiment is mainly an implementation of the above-mentioned step S103. Step S103 may include but is not limited to at least one of the following:

[0191] Step S1031: Perform weighted fusion processing on the respective inference results based on the contribution weights corresponding to the multiple processing models to obtain a target response result for the target input data.

[0192] The method for determining the contribution weights corresponding to the allocation of multiple processing models can refer to the relevant introduction to the contribution weights introduced above, and will not be repeated here.

[0193] In this embodiment, the third prompt information generated based on the contribution weights and inference results of the multiple processing models can be input into the fourth processing model, which then processes the information to generate the target response. The fourth processing model possesses powerful text generation and logical reasoning capabilities, and can integrate information from different sources to generate a desired output.

[0194] Alternatively, for the inference result of each processing model, multiply it by the corresponding contribution weight, and then add the weighted results of all models to obtain the final target response result.

[0195] In this embodiment, different processing models may analyze the target input data from different perspectives, generating diverse inference results. Contribution weights reflect the relative importance of each model in the overall task. By weightedly fusing the inference results of multiple processing models based on their corresponding contribution weights, a more comprehensive and accurate target response can be generated.

[0196] Step S1032: Obtain user portrait data of the target user, and based on the user portrait data, process the inference results of multiple processing models into target feedback content and output it to the target feedback interface.

[0197] The user portrait data may include at least interaction data between the target user and the at least one processing model, which may reflect information such as the user's interests, needs, knowledge level, and error feedback.

[0198] In this embodiment, based on user profile data, the most valuable and most effective inference results from multiple processing models can be selected. At the same time, the filtered information needs to be integrated. This integration process is not a simple accumulation of information, but rather a reorganization and arrangement of the inference results based on the user's knowledge level and understanding ability as reflected by the user profile data, resulting in targeted feedback content that better aligns with the user's cognitive habits and acceptance level. Furthermore, the integration process can also incorporate historical interaction information from the user profile data to add personalized annotations, explanations, or supplementary information to the inference results, further improving the quality and practicality of the feedback content.

[0199] For example, in a self-learning and feedback platform, user profile data is like an accurate user portrait, clearly showing that the user has a strong interest in the algebra part of mathematics and that their knowledge level is at an intermediate level. After multiple processing models conduct comprehensive and in-depth reasoning analysis on the mathematical problems submitted by users, the platform can carefully screen the reasoning results based on the user profile data. Reasoning results that are not related to the algebra part, such as content involving other branches of mathematics such as geometry, probability and statistics, are excluded. At the same time, the results of the algebra part will be further screened to match the user's intermediate knowledge level, avoiding parts that are too simple or too complex and beyond the user's understanding.

[0200] Next, the autonomous learning and feedback platform can combine the user's answering situation, such as accuracy rate, answering time and other data, with the analysis of problem-solving ideas and methods in the reasoning results to generate a detailed and organized learning report, wrong question collection or interactive summary (that is, a way of implementing the target feedback content). The learning report can not only include the user's answering situation, but also conduct in-depth error analysis on the questions that the user got wrong, pointing out the reasons for the errors, such as poor grasp of knowledge points, deviations in problem-solving ideas, etc. At the same time, based on the user's interests reflected in the user portrait data, the important knowledge points of the algebra part are summarized and sorted out, and targeted learning suggestions are provided to users, such as recommending related exercises, learning materials or online courses. These learning suggestions will also be combined with the user's past interaction habits. For example, if the user used to like to learn through videos, some high-quality algebra teaching videos will be recommended first.

[0201] In this embodiment, by obtaining user profile data for the target user and processing the inference results of multiple processing models based on this user profile data into target feedback content, the target feedback content can be made more consistent with the user's cognitive habits and acceptance level. Different users have different levels of knowledge and comprehension abilities, and this personalized processing approach makes it easier for users to understand the target feedback content.

[0202] As another optional embodiment of the present application, refer to Figure 3 , is a flow chart of a data processing method provided in Example 8 of the present application, such as Figure 3 As shown, the method may include but is not limited to the following steps:

[0203] Step S201 : In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intention represented, wherein the domain category is determined based at least on an attribute of target content in the target input data.

[0204] Step S202: Based on the domain category and the user intention, at least one processing model is called to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories.

[0205] Step S203: Process the inference result of the at least one processing model into a target response result for the target input data.

[0206] The detailed process of steps S201-S203 can be found in the relevant introduction of steps S101-S103 in Example 1, and will not be repeated here.

[0207] Step S204: When feedback data of the target user regarding the target response result is obtained, corresponding first positive evaluation content is generated based on the feedback data, where the first positive evaluation content includes target guidance information.

[0208] After the target response generated for the target input data is presented to the target user, the user can provide feedback based on their understanding and needs. This feedback data may be presented in various forms, such as text comments entered directly by the user on the target application's interactive interface (such as "This answer is not very clear" or "I understand part of the content, but I still have some questions"), user thumbs-up or thumbs-down of the response, and further questions raised by the user regarding the response.

[0209] The first positive evaluation content is different from second positive evaluation content generated for the target input data.

[0210] The second positive evaluation content generally does not involve a specific response to the target response result feedback. For example, in response to a user's question about a textbook content they don't understand (i.e., one implementation of the target input data), the user might generate a comment like "You're really great! Taking the initiative to learn about knowledge points you don't understand is the beginning of progress" (i.e., one implementation of the second positive evaluation content).

[0211] First, positive evaluation content can not only contain encouraging elements, but also goal-guiding information.

[0212] Target guidance information can be mainly used to guide users to learn other content or guide users to understand through other means when users do not accept or understand the target response results.

[0213] Of course, goal-directed information can also be used to demonstrate learning achievements.

[0214] Guiding learning of other content can be understood as:

[0215] Based on user feedback, we can analyze potential knowledge gaps or interests and guide users to learn more about related content. For example, if a user doesn't understand a particular application scenario of a mathematical function, we can guide them to learn about other, simpler or more complex applications of that function, helping them understand functions from different perspectives.

[0216] For example, in a self-learning and feedback platform, a user did not understand the explanation of the application of trigonometric functions to complex geometric figures in a target response. Based on the user's feedback data, the self-learning and feedback platform generated the first positive evaluation content, which included the target guidance information: "Although the application of trigonometric functions to complex geometric figures is a bit difficult to understand, we can start by learning its application to simple triangles. Once we master the basics, it will be much easier to work on complex figures. Below are some recommended exercises and learning materials for applying trigonometric functions to simple triangles."

[0217] Guidance is understood in other ways, which can be understood as:

[0218] Different users have different learning styles and preferences. Some users may be more suited to understanding knowledge through intuitive methods such as videos and animations, while others prefer to learn through reading text and doing exercises. If a user does not understand the current text-based response, you can use user profile data (if previously collected) and feedback data to guide the user to other more suitable methods of understanding.

[0219] For example, in a self-learning and feedback platform, a user might not understand the text explanation of a grammatical point in the target response. The target guidance message in the first positive review generated by the platform could be: "It seems the text explanation might be difficult to understand. We've prepared some interesting and engaging grammar explanation videos for you to watch to better grasp this grammar point. Click here to watch."

[0220] Demonstrating learning achievements can be understood as:

[0221] Real-time tracking and recording of students' learning achievements, such as how many questions they completed, which knowledge points they learned, and how much study time they completed, can be displayed to users as goal guidance, allowing them to intuitively see their learning progress and results. By displaying learning achievements, users gain a sense of accomplishment and satisfaction, thereby enhancing their learning motivation.

[0222] In this embodiment, the generation method and focus of positive evaluation content can be flexibly and accurately adjusted according to the learning performance and progress shown by students at different learning stages, so as to always adapt to the students' learning status and effectively maintain and improve students' learning enthusiasm.

[0223] For example, when students repeatedly make mistakes on a certain knowledge point during the learning process, the system can keenly capture this situation and then adjust the strategy for generating positive evaluation content. At this time, the positive evaluation content will focus on giving students more encouragement and guidance. For example, positive evaluation content may include: "This knowledge point is indeed quite difficult, but you have tried hard again and again. This spirit of perseverance is worthy of praise. Come, let's analyze the relevant examples and explanations in depth. I believe that with this persistence, you will definitely master it thoroughly." Through such targeted feedback, students can feel the system's recognition of their efforts, while providing clear guidance and strengthening their confidence in overcoming difficulties.

[0224] If a student shows faster learning progress at a certain stage of learning, the system will also promptly adjust the way positive evaluation content is generated. At this time, the positive evaluation content will give students a higher degree of affirmation and encouragement, and appropriately make more challenging suggestions. For example, the positive evaluation content may be: "Wow, your learning speed is amazing. You have mastered so many knowledge points so quickly! Next, why not challenge some more difficult content to further improve your abilities. I believe you will be able to make greater breakthroughs in your studies." Such feedback not only affirms the students' excellent performance, but also inspires the students' desire to move towards higher goals, prompting students to continue to maintain a positive learning attitude.

[0225] By dynamically adjusting the way positive evaluation content is generated according to different learning stages, the system can always closely follow the students' learning status, provide students with just the right incentives and guidance, and allow students to feel the system's support and attention at all stages of learning, thereby always maintaining a high level of learning enthusiasm.

[0226] In this embodiment, when users see the first positive evaluation content generated by the system based on their feedback on the target response results, they will feel the system's attention and attention to them, and thus be more willing to interact with the system. The system can also continue to provide targeted feedback and guidance based on these user interactions, forming a virtuous cycle of interaction. For example, when interacting with students, the first positive evaluation content can enhance students' learning motivation and promote their enthusiasm through an incentive mechanism.

[0227] As another optional embodiment of the present application, refer to Figure 4 , is a flow chart of a data processing method provided in Example 9 of the present application, such as Figure 4 As shown, the method may include but is not limited to the following steps:

[0228] Step S301 : In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intention represented, wherein the domain category is determined based at least on an attribute of target content in the target input data.

[0229] Step S302: Based on the domain category and the user intention, at least one processing model is called to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories.

[0230] Step S303: Process the inference result of the at least one processing model into a target response result for the target input data.

[0231] The detailed process of steps S301-S303 can be found in the relevant introduction of steps S101-S103 in Example 1, and will not be repeated here.

[0232] Step S304: Obtain interaction data between the target user and the target application, and update the knowledge graph data of the target user based on the interaction data.

[0233] The target application is an application that can call the at least one processing model to execute a target processing function, and the interaction data includes data as input to the processing model and feedback data for the inference result output by the processing model.

[0234] Knowledge graph data can be used to reflect at least one of the target user's knowledge structure, skill structure, and experience structure.

[0235] The following uses students as an example to explain knowledge graph data updates in detail. For example, the raw information students actively input into the target application (such as submitted homework answers, problem-solving steps, and stated learning questions) can serve as learning data input to the processing model. This data can reflect students' learning behavior and current knowledge mastery.

[0236] After the processing model generates inference results (such as homework grading results and problem answers) based on the input learning data, students will provide feedback on these results, generating feedback data. Feedback can take various forms, including judgments on the correctness of answers, understanding of problem-solving ideas, and further questions about knowledge points.

[0237] Based on the learning data and feedback data, the students' learning progress, error types, test scores, etc. in the knowledge graph data can be adjusted to ensure that the knowledge graph data is consistent with the students' actual learning situation.

[0238] In this embodiment, the frequency and difficulty of students' errors can be analyzed based on learning data and feedback data, and the students' weak links in corresponding knowledge points can be marked in the knowledge graph data.

[0239] Based on the updated knowledge graph, we can analyze the correlation between the knowledge points mastered by students and learning resources such as course content, exercises, and video courses. For example, when the knowledge graph shows that a student has a good understanding of a mathematical theorem, the system will recommend related extension exercises, practical application cases, and relevant mathematical competition questions to help students further deepen their understanding and application of the theorem.

[0240] By using real-time updated knowledge graph data, we can accurately track students' learning paths and understand their key learning points, difficulties, and progress at different stages. For example, the system can plot a student's learning time curve for each knowledge point over the course of a semester, analyzing their learning status and efficiency at different time periods.

[0241] Furthermore, knowledge graph data can help the system identify bottlenecks in students' learning—critical knowledge points that hinder further learning or performance improvement. Based on the weak links identified in the knowledge graph, the system can analyze the connections between these weak links to identify the root causes of students' learning difficulties, thereby providing more targeted solutions.

[0242] Furthermore, based on the updated knowledge graph data, the system can analyze students' knowledge needs and interests and recommend personalized learning resources, functions, or services. For example, if the knowledge graph shows that a student has a strong interest in a certain subject area and has mastered a certain foundation, the system will recommend more in-depth courses, research reports, or related academic exchange activities in that field to meet the student's needs for further in-depth learning and exploration.

[0243] In this embodiment, step S304 may also be performed after the above-mentioned step S203.

[0244] As another optional embodiment of the present application, refer to Figure 5 , is a flow chart of a data processing method provided in Example 10 of the present application, such as Figure 5 As shown, the method may include but is not limited to the following steps:

[0245] Step S401 : In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intention represented, wherein the domain category is determined based at least on an attribute of target content in the target input data.

[0246] Step S402: Based on the domain category and the user intention, at least one processing model is called to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories.

[0247] Step S403: Process the inference result of the at least one processing model into a target response result for the target input data.

[0248] The detailed process of steps S401-S403 can be found in the relevant introduction of steps S101-S103 in Example 1, and will not be repeated here.

[0249] Step S404: When feedback data of the target user regarding the target response result is obtained, a corresponding target feedback interface is generated based on the feedback data, and the target feedback interface provides operation controls or recommended content for updating the knowledge graph data of the target user.

[0250] If the user triggers an operation control for updating the knowledge graph data of the target user, the target application can update the knowledge graph data of the target user based on the feedback data.

[0251] Based on the updated knowledge graph data, recommended content can be generated. Still using the target user as a student as an example, a comprehensive and in-depth learning report can be automatically generated based on the updated knowledge graph data. This report can cover multiple important dimensions, including learning progress, error analysis, and knowledge mastery.

[0252] In terms of learning progress, the report will list in detail the tasks that the students have completed, such as the chapters studied, the homework completed, etc., and clearly point out the weak links, that is, the knowledge points with low mastery or loosely related knowledge systems in the knowledge map.

[0253] In the error analysis section, the report will provide a detailed statistical analysis of error types, such as conceptual misunderstandings, calculation errors, logical reasoning errors, etc., and analyze the error frequency. By combining the knowledge graph, it will identify frequently occurring knowledge points and their related knowledge points, and conduct an in-depth analysis of the root causes of the errors.

[0254] Knowledge point mastery is displayed through intuitive charts and text descriptions, demonstrating the student's mastery of each knowledge point, as well as the relationships between knowledge points and the learning path. Learning reports are presented using a variety of elements, including text, formulas, images, and tables, so that students can clearly and comprehensively understand their learning status.

[0255] Based on the learning report, personalized learning feedback content can be further generated, mainly including course recommendations, exercise recommendations and video recommendations.

[0256] Course Recommendations: We conduct an in-depth analysis of the knowledge points and weak links in students' knowledge graphs, and recommend relevant course content to students based on the correlation and learning order between the knowledge points in the knowledge graph data. These courses will not only cover students' weak points, but also help them build a complete knowledge system and deepen their understanding of relevant knowledge. For example, if a student shows a poor grasp of the function knowledge graph in the mathematics subject, we recommend a series of courses from basic function concepts to complex function applications, including function graph drawing, function property analysis, and the application of functions in practical problems.

[0257] Exercise recommendations: For the knowledge points marked in the knowledge graph data where students frequently make mistakes, suitable exercises can be carefully selected from the massive exercise library. These exercises will be classified and sorted according to multiple dimensions such as difficulty, question type, and knowledge point relevance to form a set of targeted exercise sets. By practicing these exercises, students can consolidate weak knowledge points and improve their problem-solving and knowledge application abilities. For example, for the knowledge points in mechanics in the physics subject where students frequently make mistakes, the system will recommend a series of exercises ranging from the application of simple mechanics principles to complex mechanics comprehensive problems, covering a variety of question types such as multiple-choice questions, fill-in-the-blank questions, and calculation questions.

[0258] Video Recommendations: Based on students' learning progress and interests, and taking into account the characteristics and importance of each knowledge point in the knowledge graph data, relevant video courses can be recommended. These video courses not only explain the knowledge points in detail, but also combine real-life examples and problem-solving techniques to present the learning content in a vivid and visual way, enhancing students' learning experience and comprehension. For example, for a knowledge point in a specific historical period in the history subject, the system will recommend a video course that includes vivid historical stories, expert interpretations, and cultural relics displays to stimulate students' learning interest.

[0259] Users can interact with the generated recommendations in the target feedback interface. For example, students can select recommended exercises and practice them on the target feedback page. After students select exercises, the target application can record the students' practice process and results, including answering time, answer accuracy, reasons for errors, etc. At the same time, the target application can provide timely feedback on answering results and detailed analysis based on the students' practice, helping students understand their learning effects and existing problems. In addition, students' practice data will also be updated in real time to the knowledge graph data, so that the target application can further adjust the learning recommendation strategy and provide students with more accurate and personalized learning support.

[0260] The following will take the autonomous learning and feedback platform as the target application, and elaborate on the application of the data processing method provided by this application in actual scenarios, covering the entire process of the platform processing student input data, generating feedback, and further optimizing learning support based on feedback data.

[0261] For example, Figure 6 As shown, a student inputs a picture of a page in the textbook and a text description "I didn't understand this in class" into the self-learning and feedback platform. The self-learning and feedback platform can generate a message of encouragement, "You are really great. You took the initiative to learn the knowledge points you didn't understand. This is the beginning of progress!" (i.e., one implementation method of the second positive evaluation content).

[0262] The system then understands user intent and identifies domain categories, invoking a physics processing model to generate and process the student's input data. Since only one processing model is involved, its inference results can be used as the target response. This inference result can include: related knowledge points, related physics knowledge points, real-life examples, recommended learning materials, and more.

[0263] Feedback generated through student interaction within the self-learning and feedback platform generates learning reports, updates knowledge graph data, and the first positive evaluation, such as, "Congratulations on completing this study! As long as you study hard, all this physics knowledge will be no problem. Keep up the good work! Your learning report has been generated, and the learning graph has been updated. Click to view it." This not only acknowledges the student's efforts and progress but also reminds them to review their learning reports and updated knowledge graph to further understand their learning progress. As a means of implementing goal-oriented information, learning reports and learning graphs can guide students in targeted follow-up learning.

[0264] For example, student Xiao Ming is doing his math homework and encounters a complex algebra problem. Xiao Ming tries to solve it on his own, but finds himself stuck at a certain step and is unsure whether he made a mistake.

[0265] Xiao Ming takes a photo of his handwritten math problem and uploads it to the self-learning and feedback platform. The platform uses multimodal LLM and image analysis technology to accurately recognize the handwritten problem and formulas, and converts them into a unified format for subsequent analysis.

[0266] The autonomous learning and feedback platform analyzes the identified problem, accurately identifying it as a math problem and the student's intention to solve it. Based on this identification, the multimodal LLM (one implementation of the processing model) for the mathematics discipline is invoked to solve the problem.

[0267] During the problem-solving process, the multimodal LLM in mathematics discovered that Xiao Ming had made a small error in a calculation step, resulting in an inaccurate result. The system further identified the root cause of Xiao Ming's error and specifically pointed out the specific calculation step that was faulty. The system also provided Xiao Ming with a detailed explanation of the problem-solving steps, helping him understand the correct approach.

[0268] To help Xiao Ming strengthen his understanding of relevant knowledge points, the self-learning and feedback platform recommends relevant exercises and videos based on Xiao Ming's weaknesses. These recommended resources are targeted and can meet Xiao Ming's learning needs.

[0269] With the help of the platform, Xiao Ming was able to solve the problem in a short time, understand the reasons for his mistakes, and improve his math ability through personalized learning resources.

[0270] For example, student Xiao Zhang is preparing to take the final history exam, but he finds that his memory of certain historical events is vague, especially the changes of dynasties and major historical events in ancient China.

[0271] Xiao Zhang enters his understanding of a historical question into the autonomous learning and feedback platform, such as "Why did the Tang Dynasty become a powerful empire?" The autonomous learning and feedback platform uses a multimodal LLM to perform grammatical analysis, keyword extraction, and semantic understanding on the text input, generating target input data in a unified format.

[0272] The autonomous learning and feedback platform performs semantic analysis on target input data in a unified format to determine that the student's intention is to answer questions and analyze the reasons, and that the domain categories are history and economics.

[0273] In order to provide comprehensive and accurate answers, the autonomous learning and feedback platform simultaneously calls on the multimodal LLM of history and economics to provide answers and takes them into comprehensive consideration.

[0274] During the answering process, it was discovered that Xiao Zhang had overlooked two important factors: political reform and economic development in the early Tang Dynasty. This error was then pointed out. Simultaneously, the autonomous learning and feedback platform generated a detailed learning report based on Xiao Zhang's mistakes and weaknesses, using a dynamically updated knowledge graph. This report highlighted Xiao Zhang's shortcomings in the ancient history section, providing him with a clear direction for his studies.

[0275] To help Xiao Zhang conduct targeted review, the self-learning and feedback platform provides personalized recommendations for relevant review resources, such as the video course "Politics and Economic Reforms in the Tang Dynasty" and its homework exercises. Furthermore, within the learning report and feedback page generation module, the system automatically generates detailed reports to help Xiao Zhang understand his review progress.

[0276] Xiao Zhang was able to conduct targeted review on his weak points through the study reports and review resources provided by the platform. Through personalized learning resources, Xiao Zhang improved his history knowledge and was better prepared for the final exam.

[0277] The detailed description of the above scenarios demonstrates that the data processing method provided in this application has broad application prospects and significant advantages in autonomous learning and feedback platforms. This method can provide personalized learning support and feedback based on students' actual needs and learning situations, effectively improving their learning outcomes and efficiency.

[0278] Next, a data processing device provided by the present application is introduced. The data processing device introduced below and the data processing method introduced above can be referenced to each other.

[0279] Reference Figure 7 The data processing device may include: an intention and domain category recognition module 100, a scheduling module 200 and a collaborative answer module 300.

[0280] The intent and domain category identification module 100 is used to identify the domain category to which the target input data belongs and the user intent represented in response to obtaining target input data input into a target application, wherein the domain category is determined based at least on the attributes of the target content in the target input data.

[0281] The scheduling module 200 is used to call at least one processing model based on the domain category and the user intention to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories.

[0282] The collaborative answering module 300 is configured to process the inference result of the at least one processing model into a target response result for the target input data.

[0283] The scheduling module 200 invokes at least one processing model based on the domain category and the user intention to generate and process the target input data, which may specifically include at least one of the following:

[0284] In the case of identifying that the domain category is unique, calling a first processing model that matches the domain category from a target model library based on the user intention to generate and process the target input data;

[0285] In the case of recognizing that the domain category is not unique, calling a plurality of processing models corresponding to the non-unique domain categories from a target model library based on the user intention to generate and process the target input data;

[0286] The target model library is deployed on a local end of the electronic device or on a target processing device connected to the electronic device.

[0287] The scheduling module 200 calls multiple processing models corresponding to the non-unique domain categories to generate and process the target input data, which may include:

[0288] Determining contribution weights of the multiple processing models based on the user intent and / or target reference data, wherein the target reference data can represent the contribution degree of the content of each attribute in the target input data to the user intent;

[0289] Based on the contribution weights, the multiple processing models are called in parallel or serially to generate and process the attribute contents of the target input data respectively; or

[0290] The plurality of processing models are called based on the contribution weights to process the target input data into fused input data which is input into a second processing model for generation processing.

[0291] The scheduling module determines the contribution weights of the multiple processing models based on the user intention and / or target reference data, which may specifically include at least one of the following:

[0292] Determining contribution weights of the multiple processing models based on a contribution ratio of each attribute content in the target input data to the user intention;

[0293] Determining contribution weights of the plurality of processing models based on the degree of association between each attribute content and the context in the target input data;

[0294] Determining contribution weights of the multiple processing models based on the attribute categories and proportions of the attribute contents in the target input data;

[0295] The contribution weights of the multiple processing models are determined based on the degree of matching between the attribute content in the target input data and the user portrait information of the target user.

[0296] The intent and domain category identification module 100 may include:

[0297] The multimodal data processing module is used to process the target input data in a unified format when the target input data includes multiple data in different modalities.

[0298] The intent classification module is used to perform semantic analysis on the target input data in a unified format to determine the user intent represented.

[0299] The domain category classification module is configured to determine the domain category to which the target input data belongs based on the category attribute and / or semantic attribute of the target content and the context of the target input data.

[0300] The collaborative answering module processes the inference result of the at least one processing model into a target response result for the target input data, which may specifically include at least one of the following:

[0301] Performing weighted fusion processing on the respective inference results based on the contribution weights corresponding to the multiple processing models to obtain a target response result for the target input data;

[0302] Obtain user portrait data of the target user, and based on the user portrait data, process the inference results of multiple processing models into target feedback content and output it to the target feedback interface, wherein the user portrait data at least includes interaction data between the target user and the at least one processing model.

[0303] The data processing device may further include:

[0304] The incentive mechanism module is used to generate corresponding first positive evaluation content based on the feedback data of the target user regarding the target response result when the feedback data is obtained.

[0305] The first positive evaluation content is different from the second positive evaluation content generated for the target input data, and the first positive evaluation content includes target guidance information.

[0306] The data processing device may further include:

[0307] A dynamic knowledge graph update module is used to obtain interaction data between a target user and the target application, and update the knowledge graph data of the target user based on the interaction data;

[0308] The target application is an application that can call the at least one processing model to execute a target processing function, and the interaction data includes data as input to the processing model and feedback data for the inference result output by the processing model.

[0309] The data processing device may further include:

[0310] A feedback page generation module is used to generate a corresponding target feedback interface based on the feedback data obtained from the target user regarding the target response result, and the target feedback interface provides operation controls or recommended content for updating the knowledge graph data of the target user.

[0311] In another embodiment of the present application, an electronic device is provided, comprising at least one processor and at least one processing model capable of running on the processor, wherein the processing model can be called by a target application to perform at least one of the following:

[0312] In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intent represented, wherein the domain category is determined based on at least an attribute of target content in the target input data;

[0313] Based on the domain category and the user intention, calling at least one processing model to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories;

[0314] The inference result of the at least one processing model is processed into a target response result for the target input data.

[0315] It should also be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines.

[0316] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course can also be implemented by special hardware including application-specific integrated circuits, special CPUs, special memories, special components, etc. In general, all functions performed by computer programs can be easily implemented with corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits or special circuits, etc. However, for the present application, software program implementation is a better implementation method in most cases. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer's floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., and includes a number of instructions to enable a computer device (which can be a personal computer, training equipment, or network equipment, etc.) to execute the methods described in each embodiment of the present application.

[0317] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0318] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, a computer, a training device or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website, a computer, a training device or a data center. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium can be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

Claims

1. A data processing method, comprising: In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intent represented, wherein the domain category is determined based on at least an attribute of target content in the target input data; Based on the domain category and the user intention, calling at least one processing model to generate and process the target input data, wherein the calling strategy of the processing model is related to the number of the domain categories; The inference result of the at least one processing model is processed into a target response result for the target input data.

2. The data processing method according to claim 1, wherein the generating and processing of the target input data by calling at least one processing model based on the domain category and the user intention comprises at least one of the following: In the case of identifying that the domain category is unique, calling a first processing model that matches the domain category from a target model library based on the user intention to generate and process the target input data; In the case of recognizing that the domain category is not unique, calling a plurality of processing models corresponding to the non-unique domain categories from a target model library based on the user intention to generate and process the target input data; in, The target model library is deployed on a local end of the electronic device or on a target processing device connected to the electronic device.

3. The data processing method according to claim 2, wherein: Calling a plurality of processing models corresponding to the non-unique domain categories to generate and process the target input data includes: Determining contribution weights of the multiple processing models based on the user intent and / or target reference data, wherein the target reference data can represent the contribution degree of the content of each attribute in the target input data to the user intent; Based on the contribution weights, the multiple processing models are called in parallel or serially to generate and process the attribute contents of the target input data respectively; or The plurality of processing models are called based on the contribution weights to process the target input data into fused input data which is input into a second processing model for generation processing.

4. The data processing method according to claim 3, wherein: Determining contribution weights of the plurality of processing models based on the user intention and / or target reference data includes at least one of the following: Determining contribution weights of the multiple processing models based on a contribution ratio of each attribute content in the target input data to the user intention; Determining contribution weights of the plurality of processing models based on the degree of association between each attribute content and the context in the target input data; Determining contribution weights of the multiple processing models based on the attribute categories and proportions of the attribute contents in the target input data; The contribution weights of the multiple processing models are determined based on the degree of matching between the attribute content in the target input data and the user portrait information of the target user.

5. The data processing method according to claim 1, wherein the step of identifying the domain category to which the target input data belongs and the user intent represented comprises: In a case where the target input data includes multiple data of different modalities, performing unified format processing on the target input data; Perform semantic analysis on target input data in a unified format to determine the user intent represented; The domain category to which the target input data belongs is determined according to the category attribute and / or semantic attribute of the target content and the context of the target input data.

6. The data processing method according to claim 1, wherein processing the inference result of the at least one processing model into a target response result for the target input data comprises at least one of the following: Performing weighted fusion processing on the respective inference results based on the contribution weights corresponding to the multiple processing models to obtain a target response result for the target input data; Obtain user portrait data of the target user, and based on the user portrait data, process the inference results of multiple processing models into target feedback content and output it to the target feedback interface, wherein the user portrait data at least includes interaction data between the target user and the at least one processing model.

7. The data processing method according to claim 1, further comprising: When feedback data of the target user regarding the target response result is obtained, generating corresponding first positive evaluation content based on the feedback data; The first positive evaluation content is different from the second positive evaluation content generated for the target input data, and the first positive evaluation content includes target guidance information.

8. The data processing method according to claim 1 or 7, further comprising: Obtaining interaction data between a target user and the target application, and updating the knowledge graph data of the target user based on the interaction data; The target application is an application that can call the at least one processing model to execute a target processing function, and the interaction data includes data as input to the processing model and feedback data for the inference result output by the processing model.

9. The data processing method according to claim 1 or 7, further comprising: When feedback data of the target user regarding the target response result is obtained, a corresponding target feedback interface is generated based on the feedback data, and the target feedback interface provides operation controls or recommended content for updating the knowledge graph data of the target user.

10. An electronic device comprising at least one processor and at least one processing model capable of running on the processor, wherein the processing model can be called by a target application to perform at least one of the following: In response to obtaining target input data input into a target application, identifying a domain category to which the target input data belongs and a user intent represented, wherein the domain category is determined based on at least an attribute of target content in the target input data; Based on the domain category and the user intention, at least one processing model is called to generate and process the target input data, wherein: The calling strategy of the processing model is related to the number of the domain categories; The inference result of the at least one processing model is processed into a target response result for the target input data.

Images