An intelligent table processing system for a smart campus

Abstract

The application relates to the field of data processing, and provides an intelligent table processing system for a smart campus. The system comprises a knowledge construction module, which is used for ontology modeling and knowledge construction according to metadata and text data of a smart campus data center, so as to obtain a graph database and a vector knowledge base; a model configuration module, which is used for incremental pre-training based on business-related text materials, and supervised fine-tuning training based on artificial annotation data, so as to obtain large language models respectively meeting multiple different business scenarios; a fusion module, which is used for function fusion of a knowledge graph, a vector knowledge base and a large language model based on an M-RAG mode; and an application module, which is used for taking the functions provided by the fusion module as primary functions, and applying secondary functions to the current business scenario, wherein the secondary functions include one or a combination of data filling, data review and process approval. The application provides an intelligent table processing mode for various business scenarios of a campus.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to an intelligent form processing system for a smart campus. BACKGROUND

[0002] Teachers and students need to fill in many forms in teaching, scientific research, management and other aspects. The filling work is under great pressure and consumes a lot of time. In addition to some basic information such as name, gender, ID number, date of birth, education, work experience, scientific research projects, papers, teaching work and the like, which need to be collected from multiple places, they also need to be repeatedly filled in different forms according to different dimensions. Similarly, the work pressure of functional departments is also very large. When the functional departments need to complete a reporting task, not only do they need to manually build a reporting form according to the reporting form, but also need to manually associate the form fields with the database fields and clarify the data basis logic. How to use artificial intelligence technology to realize intelligent recognition and construction of forms, intelligent association of form fields and data, automatic filling of data and the like on the basis of the construction achievements of the smart campus data governance platform, better, faster and more accurately solve the pain points in the daily work and study of teachers and students, and more research and application are needed.

[0003] Therefore, the present application provides an intelligent form processing system for a smart campus to solve the above technical problems. SUMMARY

[0004] The present application aims to provide an intelligent form processing system for a smart campus, which can solve at least one of the above technical problems. The specific scheme is as follows:

[0005] According to the specific embodiment of the present application, in a first aspect, the present application provides an intelligent form processing system for a smart campus, comprising:

[0006] The knowledge construction module is configured to perform knowledge construction for a plurality of different business scenarios of the smart campus, the plurality of different business scenarios including a student financial aid application business scenario, a faculty and staff evaluation business scenario, a scientific research project declaration business scenario, and a title declaration business scenario, and the metadata used in different business scenarios is different. The model configuration module is configured to perform incremental pre-training based on business-related text materials, and supervised fine-tuning training based on artificially annotated data, to obtain a large language model that meets a plurality of different business scenarios. The fusion module is configured to implement a hybrid retrieval enhanced M-RAG mode based on fusion, and to perform functional fusion of the knowledge graph, the vector knowledge base, and the large language model. The functional fusion is configured to enable the large language model to refer to the graph database and the vector knowledge base to analyze and evaluate form fields of a current business scenario and recommend content, and to support the large language model to perform an intelligent business consultation function. The application module is configured to use the function provided by the fusion module as a primary function, and to apply a secondary function to the current business scenario, the secondary function including one or a combination of data reporting, data review, and process approval. The fusion module includes a vectorization unit, a search selection unit, a knowledge reordering unit, and an answer generation unit. The vectorization unit is configured to receive a specific input question for form field evaluation, to perform vectorization processing on the specific input question, or to perform feature analysis on the specific input question based on the large language model, to perform vectorization processing on entities or relationships obtained by the feature analysis, and to obtain a text vector. The search selection unit is configured to refer to the text vector, to perform knowledge search in the vector database and the graph database respectively, and to obtain knowledge text segments. The knowledge search in the graph database includes searching for the text vector in the graph database, obtaining the first m nodes with the highest similarity to the text vector, expanding each of the m nodes to a second neighborhood to obtain m subgraphs corresponding to the m nodes one by one, and processing each of the subgraphs into a knowledge text segment based on the triple relationship stored in the knowledge graph. The knowledge search in the vector database includes searching for the text vector in the vector database to obtain n knowledge text segments. The knowledge reordering unit is configured to perform mixing and reordering of the knowledge text segments based on a reordering model, and to obtain the first k knowledge text segments in the reordered knowledge text segments, where k < m + n. The answer generation unit is configured to construct a prompt word template based on the specific input question and the k knowledge text segments, and to assist in prompting the primary function and the secondary function based on the prompt word template.

[0007] In an implementation, the supervised fine-tuning training based on the artificially labeled data is used to supervise fine-tuning of the large language model in the selection of the first k knowledge text segments by artificially selected or modified data as artificially labeled data during use of the system.

[0008] In an implementation, the knowledge construction module comprises: a metadata extraction unit configured to extract metadata including data sets, data subsets, data classes, and data subclasses from a data center; an ontology modeling unit configured to define entity types and relationships of entity types, and construct a knowledge graph; a knowledge storage unit configured to store the knowledge graph in the form of triples, and store the knowledge graph using a database; and a dynamic updating unit configured to update the knowledge graph according to user feedback and new data collection.

[0009] In an implementation, the ontology modeling unit comprises: an entity type definition subunit configured to define entity types, the entity types including data items, data subclasses, data classes, data subsets, data sets, and / or data item standard types; and a relationship definition subunit configured to define relationships, the relationships including entity inclusion logic, entity English names, data standard sources, and / or data types.

[0010] In an implementation, the knowledge storage unit comprises: a triple construction subunit configured to construct triples, the triples used to represent relationships between different entities; and a database storage subunit configured to store the triples in a database.

[0011] In an implementation, the dynamic updating unit comprises: a feedback receiving subunit configured to receive feedback information of a user; and an updating logic subunit configured to update entities and relationships in the knowledge graph according to the feedback information.

[0012] In an implementation, the model configuration module comprises: a baseline model selection unit configured to select a baseline model; a data cleaning unit configured to clean business text materials; and a pre-training unit configured to incrementally pre-train the baseline model to adapt to a business scenario.

[0013] In an implementation, the pre-training unit comprises: a text material collating subunit configured to collate the cleaned business text materials; and a model incremental training subunit configured to incrementally train the baseline model based on the collated business text materials.

[0014] In an implementation, the vectorization unit comprises: a question analysis subunit configured to analyze a question input by a user; and a feature extraction subunit configured to extract feature entities or relationships from the question, and perform vectorization processing on the extracted structure.

[0015] In an implementation, the answer generation unit comprises: a text conversion subunit configured to convert the subgraph into text knowledge; and an answer synthesis subunit configured to input the text knowledge and the question into a large language model to obtain an answer output by the large language model.

[0016] Compared with the prior art, the above scheme of the embodiments of the present application has at least the following beneficial effects:

[0017] The present application provides an intelligent form processing system for a smart campus, which can significantly improve the form processing efficiency and accuracy in various business scenarios of the smart campus. Specifically, the knowledge construction module constructs knowledge for different business scenarios (such as student financial aid application, faculty evaluation, scientific research project declaration, and title declaration). The system can flexibly process the corresponding metadata according to different scenarios, enhancing the flexibility and pertinence of data processing. Meanwhile, the model configuration module performs incremental pre-training based on business-related text data to generate a large language model that meets specific business scenarios, enabling the system to understand complex professional terms and business rules and enhancing business adaptability and intelligence. The fusion module fuses the knowledge graph and the large language model function through M-RAG technology, enabling the large language model to analyze and evaluate form fields and recommend content using the knowledge graph, and supporting intelligent business consultation, thereby improving the automation level and user experience of the system. Finally, the application module applies the functions provided by the fusion module to specific business scenarios (such as data reporting, data review, and process approval), simplifying the business process and improving work efficiency. Therefore, the system not only reduces manual intervention but also improves the accuracy and efficiency of data processing. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A block composition diagram of an intelligent form processing system for a smart campus is shown;

[0019] Figure 2 A block composition diagram of a knowledge construction module is shown;

[0020] Figure 3 A subunit composition diagram of an ontology modeling unit is shown;

[0021] Figure 4 A subunit composition diagram of a knowledge storage unit is shown;

[0022] Figure 5 A subunit composition diagram of a dynamic updating unit is shown;

[0023] Figure 6 A block composition diagram of a model configuration module is shown;

[0024] Figure 7A subunit composition block diagram of a pre-training unit is shown;

[0025] Figure 8 A unit composition block diagram of a fusion module is shown;

[0026] Figure 9 A subunit composition block diagram of a vectorization unit is shown;

[0027] Figure 10 A subunit composition block diagram of an answer generation unit is shown;

[0028] Figure 11 A function display diagram of an intelligent one-table platform is shown;

[0029] Figure 12 A schematic diagram of a knowledge graph is shown;

[0030] Figure 13 A schematic diagram of an updated knowledge graph is shown;

[0031] Figure 14 A logical implementation schematic diagram of M-RAG is shown. DETAILED DESCRIPTION

[0032] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0033] The terms used in the embodiments of the present application are only for the purpose of describing the specific embodiments, and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. "Multiple" generally includes at least two.

[0034] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0035] It should be understood that, although the terms first, second, third, etc. can be employed in this application embodiment, these are described not to limit the scope of the application. These terms are only used to distinguish one description from another. For example, a first can be termed a second, and similarly, a second can be termed a first, without departing from the scope of the application.

[0036] Depending on the context, the word "if" as used herein can be interpreted to mean "when" or "while" or "in response to determining" or "in response to detecting." Similarly, the phrase "if it is determined" or "if [a stated condition or event] is detected" can be interpreted to mean "upon determining" or "in response to determining" or "upon detecting [the stated condition or event]" or "in response to detecting [the stated condition or event]."

[0037] Teachers and students need to fill in many forms in teaching, scientific research, management and other aspects. The filling work is under great pressure and consumes a lot of time. In addition to some basic information such as name, gender, ID number, date of birth, education, work experience, scientific research projects, papers, teaching work and the like, they also need to be collected from multiple places and repeatedly filled in different forms according to different dimensions. Similarly, the work pressure of functional departments is also very large. Whenever the functional department needs to complete a reporting task, it not only needs to manually build a reporting form according to the reporting form, but also needs to manually associate the form field with the database field and clarify the data basis logic. How to use artificial intelligence technology to realize intelligent recognition and construction of forms, intelligent association of form fields and data, automatic filling of data and the like on the basis of the construction achievements of the smart campus data governance platform, better, faster and more accurately solve the pain points in the daily work and study of teachers and students, more research and application are needed.

[0038] In a scheme in the related art, a data reporting system can be constructed based on a business leading mode of a form engine and a process engine. The form engine and the process engine are used to implement the construction of forms and the construction of processes according to business characteristics, and data is provided through a data governance platform. However, the technical ability and business ability required for business process production are high, the development cycle is long, and data management is not convenient. In another scheme in the related art, a corresponding reporting system can be developed for a specific data reporting scene based on software development technology on the basis of a data governance platform, such as a high-base table data reporting platform and a teaching basic state database reporting platform. However, in this way, a long development cycle is required, which is equivalent to developing a new information management system, and it is difficult to reuse, with high cost.

[0039] Therefore, the present application provides an intelligent form processing system for a smart campus to solve the above technical problems.

[0040] Optional embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0041] The embodiments provided in the present application are an embodiment of an intelligent form processing system for a smart campus. The following will be described in detail Figure 1 The embodiments of the present application will be described in detail.

[0042] Figure 1 A block diagram of the module composition of an intelligent form processing system for a smart campus is shown, as Figure 1 The system includes a knowledge construction module 11, a model configuration module 12, a fusion module 13, and an application module 14.

[0043] The knowledge construction module 11 is used to perform ontology modeling and knowledge construction according to the metadata and text data of the smart campus data center, to obtain a graph database and a vector knowledge base. The graph database is constructed based on the knowledge graph obtained from the metadata and text data, and the vector knowledge base is obtained based on text segmentation and vectorization processing of the text data. The knowledge construction module 11 is used to perform knowledge construction for a plurality of different business scenarios of the smart campus, including a student scholarship application business scenario, a faculty assessment business scenario, a scientific research project declaration business scenario, and a title declaration business scenario. Different business scenarios use different metadata.

[0044] The model configuration module 12 is used to perform incremental pre-training based on business-related text materials, and supervised fine-tuning training based on artificial annotation data, to obtain a large language model that meets a plurality of different business scenarios.

[0045] The fusion module 13 is used to implement a hybrid retrieval enhanced M-RAG method based on fusion, to functionally fuse the knowledge graph, the vector knowledge base, and the large language model. The functional fusion satisfies the large language model referring to the graph database and the vector knowledge base to analyze and evaluate the form fields of the current business scenario and recommend the content, and supports the large language model to perform an intelligent business consultation function.

[0046] The application module 14 is used to take the functions provided by the fusion module 13 as a primary function, to apply a secondary function to the current business scenario, the secondary function including one or a combination of data reporting, data review, and process approval.

[0047] In the present application, the business-related text materials refer to various text information related to specific business scenarios, which include but are not limited to the following contents.

[0048] Regulations and rules: various regulations and rules, process specifications, etc. of the school.

[0049] Table templates: templates of various forms required for different businesses, including faculty evaluation forms, research project application forms, and title evaluation forms.

[0050] Guidelines and manuals: manuals or guidelines for filling out forms.

[0051] Historical records: historical records of faculty filling out forms, past research project application materials, and historical archives of title evaluation.

[0052] Academic literature: academic papers and research reports related to research projects and academic achievements.

[0053] Training materials: relevant materials for faculty training, including training course outlines and lecture notes.

[0054] Meeting minutes: minutes of meetings related to faculty evaluation, research project application, and title evaluation.

[0055] Announcements: announcements related to faculty, research projects, and title evaluation released by the school.

[0056] Internal communications: internal communications such as email exchanges and internal announcements.

[0057] In the above embodiments of the present application, these text materials can be used to incrementally pre-train the large language model by cleaning and managing them, so that it better meets the needs of various business scenarios in the smart campus, thereby improving the functional performance and service quality of the system.

[0058] Meanwhile, in the present application, the model configuration module is also used for supervised fine-tuning training based on artificial annotation data, which is used to supervise the fine-tuning of the large language model in the selection of the first k knowledge text segments by using data selected or modified by humans as artificial annotation data, and then providing the knowledge text segment closest to the artificial selection in multiple knowledge text segments, so that the system can provide more demand-oriented content when executing secondary function applications.

[0059] Figure 2 A unit composition block diagram of a knowledge construction module 11 is shown.

[0060] As shown in Figure 2 The knowledge construction module 11 includes a metadata extraction unit 111, an ontology modeling unit 112, a knowledge storage unit 113, and a dynamic updating unit 114.

[0061] The metadata extraction unit 111 is used to extract metadata such as data sets, data subsets, data classes, and data subclasses from the data center.

[0062] The ontology modeling unit 112 is configured to define entity types and relationships between the entity types, and to construct the knowledge graph.

[0063] The knowledge storage unit 113 is configured to store the knowledge graph in the form of triples, and to store the knowledge graph using a database.

[0064] The dynamic updating unit 114 is configured to update the knowledge graph according to user feedback and new data collection.

[0065] In the present application, the knowledge construction module 11 realizes accurate construction and continuous optimization of the knowledge graph through meticulous metadata extraction, ontology modeling, knowledge storage and dynamic updating. The metadata extraction unit 111 can extract structured metadata from the data center, providing a reliable data foundation for subsequent ontology modeling. The ontology modeling unit 112 defines entity types and relationships between the entity types, and constructs the knowledge graph, so that the system can understand the logical relationships between data and enhance intelligent processing capabilities. The knowledge storage unit 113 stores the knowledge graph in the form of triples, and stores the knowledge graph using a database, which facilitates the query and management of the knowledge graph and improves the efficiency of data processing. The dynamic updating unit 114 updates the knowledge graph according to user feedback information and new data collection, ensuring the real-time and accuracy of the knowledge graph. In summary, the knowledge construction module 11 can effectively construct and maintain the knowledge graph, and provides a solid data foundation for intelligent processing of the system.

[0066] In the present application, the metadata may, for example, include data sets, data subsets, data classes, data subclasses, data items, data item standard types, data item types, constraint types, value spaces, authoritative departments, and detailed descriptions. For example, the metadata extracted by the present application is different for different business scenarios, and examples are shown as follows.

[0067] For the student financial aid application business scenario, the data set can be a student data set, the data subset can be a student aid data subset, the data class can be a student basic information class, the data subclass can be a financial aid application information subclass, the data item can be a student type, a research direction, etc., the data item standard type can be a standard definition such as a student type, the data item type can be a character type (C) such as a student type, the constraint type can be whether a student type is mandatory (M), the value space can be specific options (undergraduate, graduate, etc.) for a student type, the authoritative department can be a student aid office, and the detailed description can be the specific meaning of a student type,

[0068] For the business scenario of staff assessment, the data set can be a staff data set, the data subset can be a staff assessment data subset, the data class can be a staff basic information class, the data subclass can be a staff assessment information subclass, the data item can be the name, work number, duty, post responsibility completion, teaching quality evaluation, scientific research achievement, etc. of the staff, the data item standard type can be the standard definition of the staff assessment result, the data item type can be the character type (C) or numerical type (N) of the staff assessment result, the constraint type can be whether the staff assessment result is mandatory (M), the value space can be the specific options (excellent, good, qualified, unqualified, etc.) of the staff assessment result, the authoritative department can be the personnel department or the teaching affairs department, and the detailed description can be the specific meaning of the staff assessment result.

[0069] For the business scenario of scientific research project declaration, the data set can be a scientific research project data set, the data subset can be a scientific research project declaration data subset, the data class can be a scientific research personnel basic information class, the data subclass can be a scientific research project declaration information subclass, the data item can be the teacher's name, research direction, scientific research project name, project fund budget, expected achievement, etc., the data item standard type can be the definition of the scientific research project type, the data item type can be the character type (C) of the scientific research project name, the constraint type can be whether the scientific research project name is mandatory (M), the value space can be the options (national, provincial and ministerial, school, etc.) of the scientific research project type, the authoritative department can be the scientific research department or the science and technology department, and the detailed description can be the definition of the scientific research project type.

[0070] For the business scenario of title declaration, the data set can be a staff title data set, the data subset can be a title declaration data subset, the data class can be a staff title information class, the data subclass can be a title declaration information subclass, the data item can be the education, work experience, professional ability, academic achievement (such as the number of published papers, award-winning situation, etc.) of the staff, the data item standard type can be the definition of the title level, the data item type can be the character type (C) of the education, the constraint type can be whether the education is mandatory (M), the value space can be the options (assistant teacher, lecturer, associate professor, professor, etc.) of the title level, the authoritative department can be the personnel department, and the detailed description can be the definition of the title level.

[0071] Figure 3 A subunit composition block diagram of the ontology modeling unit 112 is shown.

[0072] As shown in the example, Figure 3 The ontology modeling unit 112 includes an entity type definition subunit 112a and a relationship definition subunit 112b.

[0073] An entity type defining subunit 112a is configured to define entity types, including data items, data sub-classes, data classes, data subsets, data sets, and / or data item standard types.

[0074] A relationship defining subunit 112b is configured to define relationships, including entity-to-entity inclusion logic, entity English names, data standard sources, and / or data types.

[0075] In the present application, the ontology modeling unit 112 ensures the clear structure and rigorous logic of the knowledge graph by defining entity types and relationships. The entity type defining subunit 112a defines entity types, including data items, data sub-classes, data classes, data subsets, data sets, data item standard types, etc., so that the system can accurately describe the structure and classification of data. The relationship defining subunit 112b defines the relationships between entities, such as inclusion logic, entity English names, data standard sources, data types, etc., to ensure that the logical relationship of the knowledge graph is clear and easy to understand and use. Therefore, the ontology modeling unit 112 can construct a knowledge graph with clear structure and rigorous logic, providing a reliable guarantee for subsequent knowledge graph applications.

[0076] Figure 4 A subunit composition block diagram of a knowledge storage unit 113 is shown.

[0077] As shown in the example, Figure 4 The knowledge storage unit 113 includes a triple construction subunit 113a and a database storage subunit 113b.

[0078] The triple construction subunit 113a is configured to construct triples, which are used to represent relationships between different entities.

[0079] The database storage subunit 113b is configured to store triples in a database.

[0080] In the present application, the knowledge storage unit 113 realizes effective storage and fast retrieval of the knowledge graph through triple construction and database storage. The triple construction subunit 113a constructs triples to represent relationships between different entities, allowing the knowledge graph to clearly express the connections between entities. The database storage subunit 113b stores triples in a database, allowing the knowledge graph to be quickly queried and accessed, improving data processing speed. Therefore, the knowledge storage unit 113 can effectively store and retrieve data in the knowledge graph, providing support for efficient operation of the system.

[0081] In the present application, the database storage subunit 113b is used for vector knowledge base construction. Among them, the main steps of the vector knowledge base are to clean the relevant unstructured documents, such as removing irrelevant characters, pictures, etc., to divide the text according to a certain format, such as paragraphs, periods, etc., to vectorize the text through an embedding model, such as word2vec, Glove, ELMo, etc., and store the vector in the vector database.

[0082] Figure 5 A subunit composition block diagram of the dynamic updating unit 114 is shown.

[0083] As shown in the example, Figure 5 The dynamic updating unit 114 includes a feedback receiving subunit 114a and an updating logic subunit 114b.

[0084] The feedback receiving subunit 114a is configured to receive feedback information of a user.

[0085] The updating logic subunit 114b is configured to update entities and relationships in the knowledge graph according to the feedback information.

[0086] In the present application, the dynamic updating unit 114 realizes dynamic updating of the knowledge graph through user feedback information and new data collection, ensuring the timeliness and accuracy of the knowledge graph. The feedback receiving subunit 114a receives feedback information of a user, so that the system can timely obtain opinions and suggestions of the user. The updating logic subunit 114b updates entities and relationships in the knowledge graph according to the feedback information, ensuring continuous improvement and perfection of the knowledge graph. Therefore, the dynamic updating unit 114 can continuously optimize the knowledge graph, keeping it closely matched with the actual business scenario.

[0087] Figure 6 A unit composition block diagram of the model configuration module 12 is shown.

[0088] As shown in the example, Figure 6 The model configuration module 12 includes a baseline model selection unit 121, a data cleaning unit 122, and a pre-training unit 123.

[0089] The baseline model selection unit 121 is configured to select a baseline model.

[0090] The data cleaning unit 122 is configured to clean business text materials.

[0091] The pre-training unit 123 is configured to incrementally pre-train the baseline model to adapt to the business scenario.

[0092] In the present application, the model configuration module 12 can adapt the large language model to the specific business scenario through the selection of the baseline model, data cleaning and pre-training, thereby improving the practicability of the model. The baseline model selection unit 121 selects a suitable baseline model, thereby laying a foundation for subsequent pre-training. The data cleaning unit 122 cleans the business text data, thereby removing noise data and improving the quality of the data. The pre-training unit 123 incrementally pre-trains the baseline model, thereby adapting it to the specific business scenario and enhancing the business adaptability of the model. Therefore, the model configuration module 12 can generate a large language model that meets the specific business scenario, thereby providing technical support for the intelligentization of the system.

[0093] Figure 7 A subunit composition block diagram of the pre-training unit 123 is shown.

[0094] As shown in the example, Figure 7 The pre-training unit 123 includes a text data collation subunit 123a and a model incremental training subunit 123b.

[0095] The text data collation subunit 123a is configured to collate the cleaned business text data.

[0096] The model incremental training subunit 123b is configured to incrementally train the baseline model based on the collated business text data.

[0097] In the present application, the pre-training unit 123 can make the large language model better understand and process the business text data through text data collation and model incremental training, thereby improving the understanding ability and accuracy of the model. The text data collation subunit 123a collates the cleaned business text data, thereby making the data more standardized and orderly. The model incremental training subunit 123b incrementally trains the baseline model based on the collated business text data, thereby improving the business processing ability of the model. Therefore, the pre-training unit 123 can further optimize the large language model, thereby better serving the specific business scenario.

[0098] First, a baseline large model is selected, and a part of the data is manually annotated to perform supervised fine-tuning training on the baseline large training model, thereby obtaining a large language model that meets the business scenario of the university.

[0099] Manual annotation: In the present application, the standard and accurate question and answer pairs are manually annotated according to the business of each functional department, such as

[0100] Q: Question: What aspects are usually included in the evaluation of the counselor assessment?

[0101] A: The counselor assessment usually includes the following aspects:

[0102] Basic qualities: This includes the political quality, moral quality, professional ethics of the counselors, which is the foundation of doing a good job of counselors.

[0103] Work responsibility fulfillment: This mainly assesses the fulfillment of the counselors' work responsibilities in ideological education, daily management of students, construction of study style, mental health education, and employment guidance. This involves whether the counselors can effectively carry out theme class meetings, heart-to-heart talks, academic guidance, and career planning activities.

[0104] Work performance: The work results of the counselors are assessed through quantitative indicators, such as student satisfaction survey results, student award-winning situations, and study style construction effectiveness. These indicators can objectively reflect the work effectiveness and contribution of the counselors.

[0105] Plus-minus items: Additional work performance or special circumstances of the counselors are given plus or minus points. For example, the counselors can be given points for participating in student work theory research, publishing relevant papers, or obtaining honors; while mistakes in work, violations of discipline, etc. will be deducted.

[0106] One-vote veto items: For serious violations of school regulations or professional ethics, such as corporal punishment of students, acceptance of students' property, etc., the one-vote veto system is implemented, directly canceling their examination qualifications.

[0107] Supervision training: Based on artificially annotated data, the LoRA (Low-Rank Adaptation of Large Language Models) fine-tuning technology is used to fine-tune the large language model, obtaining a large language model that understands education business. LoRA is a low-order adaptive fine-tuning technology for large language models. Its basic principle is to freeze the pre-trained model weight parameters, and in the case of freezing the original model parameters, by adding additional network layers to the model and only training these newly added network layer parameters.

[0108] Figure 8 A unit composition block diagram of a fusion module 13 is shown.

[0109] As shown in Figure 8 the fusion module 13 includes a vectorization unit 131, a search selection unit 132, a knowledge reordering unit 133, and an answer generation unit 134.

[0110] The vectorization unit 131 is configured to receive a specific input question for form field evaluation, perform vectorization processing on the specific input question, or perform feature analysis on the specific input question based on a large language model, to perform vectorization processing on entities or relationships obtained by the feature analysis, and obtain a text vector.

[0111] The search selection unit 132 is configured to perform knowledge search in the vector database and the graph database respectively by referring to the text vector, and obtain knowledge text segments. The knowledge search in the graph database includes: searching the text vector in the graph database, obtaining the first m nodes with the highest similarity to the text vector, expanding each of the m nodes to a two-level neighborhood, obtaining m subgraphs corresponding to the m nodes respectively, and processing each subgraph into a knowledge text segment based on the triple relationship stored in the knowledge graph. The knowledge search in the vector database includes: searching the text vector in the vector database, and obtaining n knowledge text segments.

[0112] The knowledge reordering unit 133 is configured to reorder the knowledge text segments based on a reordering model, and obtain the first k knowledge text segments from the reordered knowledge text segments. Wherein, k < m + n.

[0113] The answer generation unit 134 is configured to construct a prompt word template based on the specific input question and the k knowledge text segments, and assist in prompting the primary function and the secondary function based on the prompt word template.

[0114] In the present application, the fusion module 13 realizes the effective fusion of the knowledge graph and the large language model through feature analysis, search selection, subgraph acquisition and answer generation, and improves the intelligent consulting ability of the system. The vectorization unit 131 performs feature analysis on the user input question, so that the system can understand the key elements of the question. The search selection unit 132 searches for nodes related to the question in the knowledge graph and finds the potential answer to the question. The knowledge reordering unit 133 expands the neighborhood of the searched nodes to form a subgraph, which further enriches the answer to the question. The answer generation unit 134 converts the subgraph into text knowledge and inputs the text knowledge and the question into the large language model to obtain the answer output by the large language model, thereby realizing intelligent consulting. Therefore, the fusion module 13 can effectively combine the knowledge graph and the large language model to provide accurate intelligent consulting services.

[0115] Figure 9 A subunit composition block diagram of the vectorization unit 131 is shown.

[0116] As shown in the example, Figure 9 The vectorization unit 131 includes a question analysis subunit 131a and a feature extraction subunit 131b.

[0117] The question analysis subunit 131a is configured to analyze the user input question.

[0118] The feature extraction subunit 131b is configured to extract feature entities or relationships from the question.

[0119] In the present application, the vectorization unit 131 can accurately understand the user's question through question analysis and feature extraction, thereby improving the accuracy of question processing. The question analysis subunit 131a analyzes the user's input question, so that the system can identify the core of the question. The feature extraction subunit 131b extracts feature entities or relationships from the question, so that the system can grasp the key points of the question. Therefore, the vectorization unit 131 can accurately understand the user's question and provide reliable information for subsequent question processing.

[0120] Figure 10 A subunit composition block diagram of an answer generation unit 134 is shown.

[0121] As shown in the example, Figure 10 The answer generation unit 134 includes a text conversion subunit 134a and an answer synthesis subunit 134b.

[0122] The text conversion subunit 134a is configured to convert the subgraph into text knowledge.

[0123] The answer synthesis subunit 134b is configured to input the text knowledge and the question into a large language model to obtain an answer output by the large language model.

[0124] In the present application, the answer generation unit 134 realizes the conversion from the knowledge graph to the text answer through text conversion and answer synthesis, thereby improving the readability and accuracy of the answer. The text conversion subunit 134a converts the subgraph into text knowledge, so that the information in the knowledge graph is presented in the form of text. The answer synthesis subunit 134b inputs the text knowledge and the question into the large language model to obtain the answer output by the large language model, thereby realizing the conversion from the knowledge graph to the text answer. Therefore, the answer generation unit 134 can convert the information in the knowledge graph into a text answer that is easy to understand, thereby improving the readability and accuracy of the answer.

[0125] In the present application, Figures 1 to 10 The corresponding module composition block diagram, unit composition block diagram, and subunit composition block diagram are shown respectively. For each block diagram, the components in the diagram are connected as a whole, which represents that the components can be connected by wired or wireless connection, and the function interaction and cooperation are completed.

[0126] The intelligent form processing system for smart campus provided by the present application can present entities in the form of a platform. The following example of the present application is an intelligent form platform.

[0127] Figure 11 A functional display diagram of an intelligent form platform is shown.

[0128] The metadata of the smart campus data middle platform is extracted, and the index entity and the corresponding relationship of the unstructured data stored in the data middle platform are extracted to support the construction of the knowledge graph. The vector knowledge base is also constructed based on the integrated non-structured documents of the data middle platform. At the same time, the artificial labeled education business data is used to train and fine-tune the construction of the education big model, and through the fusion of the education big model, the vector knowledge base and the vector knowledge base, the hybrid retrieval enhancement generation is realized, and the intelligent one form function is realized, including the analysis and evaluation function of the form field, the form field matching recommendation function, the automatic generation of the form field, the intelligent business consulting function, etc., so that the constructor can quickly understand which function department the field belongs to, the meaning of the field, the calculation logic, which field in the data middle platform the field matches, and the automatic filling of the field data. At the same time, the constructor can feed back the form content, and based on the feedback content, the knowledge graph is updated. The intelligent one form platform also has system management functions such as data source management, user management, permission management, template management, and process tool capabilities such as process arrangement, process distribution, and user data confirmation. Through the intelligent one form platform constructed based on M-RAG, the technical threshold and business threshold of the function department teachers constructing the form can be reduced, and the form construction efficiency can be greatly improved. At the same time, the data (stored in the data middle platform) already existing between teachers and students does not need to be filled in repeatedly.

[0129] For example, the construction of the knowledge graph includes the steps of ontology modeling, knowledge construction, knowledge storage, and knowledge updating. This application takes student data as an example and describes each step in detail.

[0130] In the embodiments of the application, ontology modeling of the knowledge graph can be carried out, and based on the construction of the smart campus data middle platform, metadata such as data set, data subset, data class, and data subclass can be extracted, and index entities and attribute data such as explanation and description, and calculation formula in the related evaluation documents can be extracted by artificial means. The construction of the knowledge graph is realized based on the ontology, and the knowledge updating is dynamically realized.

[0131] Ontology modeling: Since the construction of the knowledge graph in the application is mainly based on the metadata of the data middle platform and the documents containing index data, the entity types mainly include data item, data subclass, data class, data subset, data set, data item standard type, data item type, data item length, data item constraint type, data item value space, data item belonging to authoritative department, and data item detailed description. The relationships mainly include derivation, inclusion, English name, data standard source, data type, data length, constraint, value space, authoritative data source department, and explanation description. For student data, the entity and relationship examples are shown in Table 1.

[0132] Table 1

[0133]

[0134] Knowledge construction: according to the metadata of the extraction data platform, according to the modeling of ontology (entity, relationship) to construct knowledge graph, realize the visualization display and query of knowledge. For example, Table 1 shows the corresponding description of entity 1 and entity 2 and the entity relationship. In a similar way to Table 1, an example of a knowledge graph is shown as Figure 12 , in which different entities and entity relationships are indicated. Taking one branch as an example, starting from the student data subset, it can be known that the student data subset contains the student status data class, the student status data class contains the student status basic data subclass table, and the student status basic data subclass table contains the student category, and so on, until the last entity indicated by the arrow, completing the indication of each entity relationship corresponding to the student data subset.

[0135] Knowledge storage: based on the modeling of ontology, the constructed knowledge is stored, mainly in the form of triples, such as (student status data class, contains, student status basic data subclass table), (student status data class, contains, student status change data subclass table), etc. The database is used for storage, and the database can be selected as Neo4j. At the same time, the entity and the relationship will be vectorized, and then stored in the database in the form of graph relationship.

[0136] Knowledge update: according to the feedback of the intelligent one-table platform for the content of the form and the data collection of the form, etc., the dynamic update of the knowledge graph is realized. For example, in the student scholarship application business scenario, it is required to fill in the student type, but the knowledge stored in the knowledge graph is the student category, so an additional knowledge is added to the student category, that is, the alias of the student category is the student type; if the research direction information of the student is collected, a data item is added to the student status basic data subclass table, that is, the student status basic data subclass table contains the research direction. In addition, with the update of the metadata in the data platform, the knowledge graph will also be dynamically updated and iterated, for example, Figure 12 The knowledge graph shown in Figure 4 is updated based on the above embodiment, and is updated to the knowledge graph shown in Figure 13 .

[0137] In this application, the mixed search enhancement generation mode is adopted, which can make the large model generate more accurate and efficient results when filling in the table. In the smart campus scenario, there are often scenes that need to mix multiple knowledge in the database. If only vector search is used, the text knowledge base needs to contain comprehensive and large amounts of similar semantic question and answer to help the large model understand. If only graph search is used, some simple scenes that can be completed by manual simple optimization will be wasted due to the overhead of the basic unit (token) of the text, and the efficiency will even be reduced after review. Therefore, the M-RAG (mixed retrieval and enhanced generation) mode is constructed to effectively improve the analysis and evaluation of form fields, the recommendation of form fields, the automatic filling of form fields, and the intelligent consultation of business in the intelligent form. For example, the implementation logic of M-RAG is as follows.

[0138] Vectorization: When the user inputs a question, there are two processing methods, one is to directly perform vectorization processing, and the other is to use a large language model to analyze the characteristics of the question, extract feature entities or relationships from the question, and perform vectorization processing on the extracted features. The embedding model used for this vectorization must be the same as the embedding model used in the construction process of the graph database and the vector database, such as word2vec, Glove, and ELMo. Among them, word2vec is a model for generating word vectors, which maps words to a high-dimensional vector space, so that words with similar semantics or context are also close in the vector space. GloVe is a static word embedding method that represents words as fixed-length vectors. ELMo uses a bidirectional LSTM (Long Short-Term Memory Network) to capture the semantics of words in different contexts to generate context-dependent word embeddings.

[0139] Knowledge search: Based on the first step of question vectorization, knowledge search is performed in the vector database and the graph database, respectively. Specifically, the text vector after feature extraction is searched in the graph database to obtain the m most similar nodes, and each node is expanded to a 2-level neighborhood to obtain m subgraphs. Each subgraph is flattened into a knowledge segment according to the triplets stored in the knowledge graph, and m knowledge text segments are obtained. Directly vectorized questions are searched in the vector database to obtain the n most similar text segments

[0140] Knowledge reordering: Based on a reordering model such as bge-reranker-large, the m knowledge segments obtained from the graph database and the n text segments obtained from the vector database are mixed and reordered to obtain the top k knowledge segments. The bge-reranker-large is a ranking model that is specifically used to reorder the list of multiple text blocks (or nodes) retrieved, so that the ranking is more matched to the relevance of the user input question.

[0141] Hint Template: Construct a hint template by combining the initial question with the k knowledge segments obtained. Specifically, it is: "Specific Question" + k knowledge segments. Based on the given k knowledge segments, analyze the fields in the question and determine whether the result can be calculated based on the existing data platform. If it can, generate the corresponding SQL code for calculation and run it in the database of the data platform to obtain the field value. Otherwise, output the meaning and description of the field.

[0142] Output: Based on the prompt template defined in the previous step, the output result is as follows.

[0143] Figure 14 A schematic diagram of the logic implementation of M-RAG is shown.

[0144] For example, such as Figure 14 As shown, the data platform integrates structured and unstructured data. For unstructured data, text cleaning and knowledge extraction are performed first, followed by the construction of a knowledge graph through text segmentation and other steps. Entities and relationships in the knowledge graph are vectorized and stored in the graph database, while unstructured text fragments are also converted into vectors and stored in the vector database. When a user inputs a question, the system first performs feature extraction and vectorization on the question, and then retrieves the most relevant subgraphs (such as sub-graphs) from both the graph database and the vector database. Figure 1 The search results are processed into subgraphs (e.g., subgraphs 1 to n) and text fragments. These results are reordered, and the top k subgraphs are selected. Next, the user's question is combined with these k subgraphs to form a prompt template, which is then fed into the large language model for inference. Finally, a response is generated based on the model's output, achieving intelligent business consultation. Throughout the process, the large language model incorporates knowledge acquired from the data platform to provide more accurate answers. Form field evaluation is equivalent to consulting the large model with specific input questions, such as "What fields are included in the form?", "Which departments do these fields belong to?", and "What do these fields mean?". Form field matching is equivalent to consulting the large model with specific input questions, such as "Which fields in the form match the fields in the data platform? Please recommend some?".

[0145] In this application, the intelligent one-form platform can be applied to various business scenarios, including student grant application, faculty and staff assessment, scientific research project application, and professional title application. The following examples illustrate the application effects for some business scenarios.

[0146] Staff assessment business scenario: Staff assessment is a routine work every year. For colleges and universities, the number of staff is large, and currently only paper forms can be used to complete the assessment. Staff assessment is divided into three steps: the first step is for staff to fill out the form, the second step is for the unit to review, and the third step is for the personnel department to approve. The construction of the staff assessment form and the workload of the staff filling out the form are the largest. Through the intelligent one-form platform, personnel department teachers can quickly and easily construct the staff assessment form, and associate the data in the data center to the assessment form, such as teachers' personnel, scientific research, and teaching information, reducing the staff's reporting content. For staff, only a small amount of data fields need to be filled out, greatly improving the efficiency of staff.

[0147] Scientific research project declaration business scenario: There are a lot of repeated content in the scientific research project declaration materials, such as teachers' basic information, previous leading or participating scientific research projects, etc. Through the intelligent one-form platform, scientific research department teachers can quickly and easily construct the scientific research project declaration form, and associate the data in the data center to the declaration form, such as teachers' name, research direction, scientific research project, and scientific research achievement, reducing the teachers' reporting content. For teachers, only the project and content of scientific research declaration need to be filled out, greatly improving the efficiency of teachers.

[0148] Title declaration business scenario: The title evaluation process is a systematic and rigorous process, aiming to comprehensively evaluate the professional ability and work performance of the declarer. Title evaluation requires the declarer to submit data such as education, work experience, professional ability, and performance results. The corresponding declaration materials usually include personal resume, work summary, performance proof (such as project report, award certificate, etc.), academic achievement (such as paper, application, etc.), and other relevant proof materials. The relevant data and materials may be scattered in different business system databases, so the intelligent one-form platform can be used to construct the form according to the requirements of the title evaluation declaration materials, and the table fields in the data center are associated through artificial intelligence to realize automatic filling of data, reducing the workload of the function department teachers and staff, and greatly improving the work experience. The title evaluation process is a systematic and rigorous process, aiming to comprehensively evaluate the professional ability and work performance of the declarer.

[0149] Other business scenarios: The intelligent one-form platform is not only suitable for colleges and universities, but also for vocational schools, K12, and even enterprises and other industries. As long as it involves the construction of forms, the association of form fields, and data reporting, the intelligent one-form platform mentioned in this application can be used to realize intelligentization, reduce personnel workload, and greatly improve work efficiency.

[0150] The present application is based on the construction of a knowledge graph and the training and fine-tuning of a large language model. Through the M-RAG technology, the knowledge graph and the large model technology are fused to support the construction of an intelligent one-table platform. Not only can it solve the repeated table filling of teachers and students in various scenarios and improve the experience of teachers and students, but it can also break down the barriers between technical and business teachers and improve the convenience of the form construction process.

[0151] Although the operations are described in a particular, sequential order, it should be understood that the ordering of the operations can change, some operations can be performed in parallel, and some operations can be performed at alternate times, without changing the overall function of the method. In some embodiments, the operations can be performed in parallel.

[0152] Any of the steps, operations, or procedures described herein can be performed or implemented using one or more hardware or software modules, alone or in combination with other devices. In one embodiment, a software module is implemented using a computer program product comprising a computer-readable medium containing computer program code, which can be executed by a computer processor for performing any or all of the steps, operations, or procedures described.

[0153] The foregoing description of the implementation of the present application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise form disclosed, and various modifications and variations can be possible according to the teaching above, which can be practiced or implemented in various ways. These examples were chosen and described in order to explain the principles of the application and its practical application so that others skilled in the art can utilize the application in various embodiments and with various modifications as are suited to the particular use contemplated.

[0154] It should be further understood that, unless otherwise specified, "connected" includes both direct and indirect connections, and that various elements can be directly or indirectly connected.

[0155] It should be further understood that, unless otherwise specified, "connected" includes both direct and indirect connections, and that various elements can be directly or indirectly connected.

[0156] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the field of this application that are not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0157] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

[0158] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An intelligent form processing system for smart campuses, characterized in that, The system includes: The knowledge construction module is used to perform ontology modeling and knowledge construction based on metadata and text data from the smart campus data center, in order to obtain a graph database and a vector knowledge base. The graph database is constructed from a knowledge graph based on metadata and text data, and the vector knowledge base is obtained based on text segmentation and vectorization of text data. The knowledge construction module is used to construct knowledge for various business scenarios of the smart campus, including student scholarship application, faculty and staff assessment, scientific research project application, and professional title application. Different business scenarios use different metadata. The model configuration module is used for incremental pre-training based on business-related text data and supervised fine-tuning training based on manually labeled data to obtain large language models that meet various different business scenarios. The fusion module is used to implement the hybrid retrieval enhancement M-RAG method based on fusion, which integrates knowledge graphs, vector knowledge bases and large language models. The functional fusion enables the large language model to refer to the graph database and vector knowledge base to analyze and evaluate form fields in the current business scenario and recommend content, and supports the large language model to perform intelligent business consultation functions. The application module is used to apply the functions provided by the integration module as primary functions to secondary functions for the current business scenario. Secondary functions include data entry, data verification, and one or a combination of process approval. The fusion module includes a vectorization unit, a search and selection unit, a knowledge reordering unit, and an answer generation unit. The vectorization unit is used to receive a specific input question for form field evaluation, perform vectorization processing on the specific input question, or perform feature analysis on the specific input question based on the large language model, so as to vectorize the entities or relations obtained by the feature analysis to obtain text vectors; The search selection unit is used to perform knowledge searches in both the vector database and the graph database, referring to the text vector, to obtain knowledge text segments. Specifically, the knowledge search in the graph database includes: searching the text vector in the graph database, obtaining the top m nodes with the highest similarity to the text vector, expanding each of the m nodes to a second-level neighborhood to obtain m subgraphs corresponding to the m nodes, and processing each subgraph into a knowledge text segment based on the triplet relationships stored in the knowledge graph. The knowledge search in the vector database includes: searching the text vector in the vector database to obtain n knowledge text segments. The knowledge reordering unit is used to mix and reorder the knowledge text segments based on the reordering model, and to obtain the first k knowledge text segments from the reordered knowledge text segments; where k < m + n. The answer generation unit constructs a prompt word template based on the specific input question and the k knowledge text segments, and guides the large language model to determine whether the result of the form field can be calculated based on the k knowledge text segments. If it can be calculated, it generates SQL code to obtain the result of the form field, and provides auxiliary prompts for the first-level function and the second-level function based on the prompt word template. The supervised fine-tuning training based on manually labeled data in the model configuration module is used to supervise and fine-tune the selection of the large language model in the first k knowledge text segments during the use of the system, using manually selected or modified data as manually labeled data.

2. The system according to claim 1, characterized in that, The knowledge building module includes: The metadata extraction unit is used to extract metadata from the data center, including datasets, data subsets, data classes, and data subclasses. The ontology modeling unit is used to define entity types and relationships between entity types, and to build a knowledge graph. The knowledge storage unit is used to store the knowledge graph in the form of triples and uses a database for storage; The dynamic update unit is used to update the knowledge graph based on user feedback and new data collection.

3. The system according to claim 2, characterized in that, The ontology modeling unit includes: The entity type definition subunit is used to define entity types, which include data items, data subclasses, data classes, data subsets, datasets, and / or standard data item types. The relation definition subunit is used to define relations, which include the inclusion logic between entities, the English names of the entities, the data standard source, and / or the data type.

4. The system according to claim 2, characterized in that, The knowledge storage unit includes: Triples are used to construct subunits, which are used to represent relationships between different entities. The database storage subunit is used to store triples in the database.

5. The system according to claim 2, characterized in that, The dynamic update unit includes: The feedback receiving subunit is used to receive user feedback information; The update logic subunit is used to update entities and relationships in the knowledge graph based on feedback information.

6. The system according to claim 1, characterized in that, The model configuration module includes: The baseline model selection unit is used to select the baseline model; The data cleaning unit is used to clean business text data; The pre-training unit is used to incrementally pre-train the baseline model to adapt it to the business scenario.

7. The system according to claim 6, characterized in that, The pre-trained units include: The text document processing subunit is used to process and clean up the business text documents. The incremental training subunit is used to incrementally train the baseline model based on the processed business text data.

8. The system according to claim 1, characterized in that, Vectorization units include: The question parsing subunit is used to parse the question input by the user; The feature extraction subunit is used to extract feature entities or relationships from a problem.

9. The system according to claim 1, characterized in that, The answer generation unit includes: The text transformation subunit is used to transform subgraphs into textual knowledge. The answer synthesis subunit is used to input textual knowledge and questions into the large language model and obtain the answer output by the large language model.

Images