Method and system for generating content interaction of educational data report based on search enhancement

Abstract

The application relates to the technical field of artificial intelligence and discloses an education data report content interaction method and system based on retrieval enhancement generation, which comprises the following steps: using a large language model to determine whether a natural language question is an education field question, and if so, performing semantic analysis to generate a structured query instruction. When the question involves cross-document correlation analysis, a structured semantic index library is retrieved to generate a retrieval result set; if policy correlation analysis is involved, a policy knowledge graph is matched by combining semantic similarity calculation and entity linking technology, and cross-modal fusion processing is performed to obtain the retrieval result set; and the retrieval result set is input into a retrieval enhancement generation model, and an education field language model is called to generate an analysis report. The application can solve the problems of inaccurate intent recognition, low cross-document analysis efficiency and the inability to dynamically combine the latest policies in the traditional interaction mode, significantly improve the efficiency and quality of education field data report interaction, and optimize the user interaction experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for interactive content of educational data reports generated based on retrieval enhancement. Background Technology

[0002] In data reporting interaction scenarios within the education sector, traditional question-and-answer methods have numerous limitations. Under current technology, multiple education data reports are typically stored independently in unstructured form, lacking semantic indexes. Cross-document knowledge retrieval relies on manual intervention, resulting in low efficiency in knowledge reuse and difficulty in connecting historical analysis logic. When faced with complex issues, such as a user inquiring about the correlation between the changing trends in graduate employment directions over the past three years and regional industrial policies, this traditional approach requires manual review of each report, filtering fragmented information through keywords, and then integrating and analyzing the data—a time-consuming process. For general questions related to higher education management, existing methods also lack structured storage of management knowledge systems, leading to answers based on preset templates or scattered data, failing to dynamically update with the latest policy documents and industry trends. Furthermore, traditional interaction models only recognize user intent at a superficial semantic level, making it difficult to distinguish between different types of questions. This results in irrelevant questions consuming system resources, and delayed responses or inaccurate answers to critical questions. Furthermore, the response generation technology for unfused retrieval enhancement relies on a single data source, which limits its depth and accuracy when multi-source correlation analysis is involved. The interaction process is cumbersome, requiring users to clarify their needs multiple times, and there is a lack of effective guidance for irrelevant questions.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for interactive educational data report content generated based on retrieval enhancement.

[0005] Firstly, the present invention provides a method for interactive content of educational data reports generated based on retrieval enhancement, the technical solution of which is as follows:

[0006] The system receives a natural language question input by the user and uses a large language model to determine whether the natural language question is an educational question, and obtains the determination result.

[0007] When the judgment result is yes, the natural language question is semantically parsed through the large language model to generate a structured query instruction containing question type labels, core entities and association analysis dimensions;

[0008] When the question type tag of the structured query instruction is cross-document association analysis, a pre-built structured semantic index is retrieved based on the core entity and the association analysis dimension to generate a search result set; when the question type tag of the structured query instruction is policy association analysis, the structured semantic index is retrieved based on the core entity and the association analysis dimension, semantic similarity is calculated on the search results to obtain semantic similarity calculation results, and the core entity is used to match nodes in the real-time accessed policy knowledge graph through entity linking technology to generate entity linking results, and the semantic similarity calculation results and entity linking results are subjected to cross-modal fusion processing to obtain the search result set;

[0009] The search result set is input into the search enhancement generation model, which then calls a language model specific to the education field to fuse the search result set and generate an analysis report containing data source citation information and dynamic visualization charts.

[0010] The beneficial effects of the interactive method for educational data report content based on retrieval enhancement of the present invention are as follows:

[0011] The method of this invention can solve industry pain points such as inaccurate intent recognition, low efficiency of cross-document analysis, and inability to dynamically combine the latest policies in traditional interaction modes, significantly improving the efficiency and quality of data report interaction in the education field and optimizing the user interaction experience.

[0012] Secondly, this invention provides an interactive system for educational data report content generated based on retrieval enhancement, the technical solution of which is as follows:

[0013] It includes: a question judgment module, a semantic parsing module, a result generation module, and a content interaction module;

[0014] The problem judgment module is used to: receive natural language questions input by users, and determine whether the natural language questions are education-related questions through a large language model, and obtain a judgment result;

[0015] The semantic parsing module is used to: when the judgment result is yes, perform semantic parsing on the natural language question through the large language model to generate a structured query instruction containing question type labels, core entities and association analysis dimensions;

[0016] The result generation module is used to: when the question type tag of the structured query instruction is cross-document association analysis, retrieve a pre-built structured semantic index based on the core entity and the association analysis dimension, and generate a search result set; when the question type tag of the structured query instruction is policy association analysis, retrieve the structured semantic index based on the core entity and the association analysis dimension, perform semantic similarity calculation on the search results, obtain semantic similarity calculation results, and use the core entity to match nodes in the real-time accessed policy knowledge graph through entity linking technology to generate entity linking results, and perform cross-modal fusion processing on the semantic similarity calculation results and the entity linking results to obtain the search result set;

[0017] The content interaction module is used to: input the search result set into the search enhancement generation model, and use the search enhancement generation model to call a language model specifically for the education field to fuse the search result set and generate an analysis report containing data source citation information and dynamic visualization charts.

[0018] The beneficial effects of the interactive system for generating educational data reports based on retrieval enhancement according to the present invention are as follows:

[0019] The system of this invention can solve industry pain points such as inaccurate intent recognition, low efficiency of cross-document analysis, and inability to dynamically integrate the latest policies in traditional interaction modes, significantly improving the efficiency and quality of data report interaction in the education field and optimizing the user interaction experience.

[0020] Thirdly, the technical solution of an electronic device according to the present invention is as follows:

[0021] It includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the interactive method for generating educational data reports based on retrieval enhancement as described in this invention.

[0022] Fourthly, the technical solution of a computer-readable storage medium provided by the present invention is as follows:

[0023] The computer-readable storage medium stores instructions that, when read, cause the computer-readable storage medium to perform the steps of the interactive method for generating educational data reports based on retrieval enhancement according to the present invention.

[0024] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0025] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0026] Figure 1 This is a flowchart illustrating an embodiment of the interactive method for generating educational data reports based on retrieval enhancement according to the present invention.

[0027] Figure 2 This is a schematic diagram of an embodiment of an interactive system for generating educational data reports based on retrieval enhancement, according to the present invention.

[0028] Figure 3 This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation

[0029] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0030] Figure 1 The diagram illustrates a flowchart of an embodiment of an interactive method for educational data report content generated based on retrieval enhancement, provided by the present invention. This interactive method for educational data report content generated based on retrieval enhancement can be executed by electronic devices such as terminal devices or servers. Figure 1 As shown, it includes the following steps:

[0031] S1. Receive the natural language question input by the user, and determine whether the natural language question is an educational question through a large language model, and obtain the judgment result.

[0032] Natural language questions refer to everyday expressions input by users in text or speech form, such as: "Changes in the employment rate of graduates from University A in the Yangtze River Delta region over the past three years." Large language models are machine learning models that integrate textual features (such as "employment rate"), contextual features (such as the preceding question containing "talent cultivation"), and behavioral features (such as the high frequency of users clicking on education policies in their history) to classify domain intent. Education domain questions refer to specialized questions involving education management, policy analysis, or institutional operations.

[0033] It should be noted that when the natural language question is not related to the education field, the process will terminate and return to prompt the user to modify the natural language question they entered.

[0034] S2. When the judgment result is yes, the natural language question is semantically parsed through the large language model to generate a structured query instruction containing question type labels, core entities and association analysis dimensions.

[0035] Here, "problem type label" refers to a classification label that identifies the type of problem analysis. "Core entity" refers to the key educational object that serves as the subject of analysis in a natural language problem (e.g., University A, University B, etc.). "Relationship analysis dimension" refers to the implicit time, theme, or policy-related analysis direction within the problem. "Structured query command" refers to a machine-parseable command that includes the problem type label, core entity, and relationship analysis dimension.

[0036] S3A. When the question type tag of the structured query instruction is cross-document association analysis, a pre-built structured semantic index library is retrieved based on the core entity and the association analysis dimension to generate a search result set.

[0037] Cross-document association analysis refers to an analysis type that performs correlation retrieval on multiple unstructured educational data reports. A structured semantic index library refers to a pre-built database that establishes semantic indexes on educational data reports.

[0038] S3B. When the question type label of the structured query instruction is policy association analysis, the structured semantic index is retrieved based on the core entity and the association analysis dimension. The semantic similarity of the retrieval results is calculated to obtain the semantic similarity calculation result. The core entity is then matched with nodes in the real-time accessed policy knowledge graph using entity linking technology to generate entity linking results. The semantic similarity calculation result and the entity linking results are then fused across modally to obtain the retrieval result set.

[0039] Policy correlation analysis refers to question types that require correlation analysis in conjunction with a policy knowledge graph. Semantic similarity calculation results refer to the set of quantified similarity values ​​between semantic fragments of policy text and query instructions. Entity linking technology refers to the matching technique that maps core entities to nodes in the knowledge graph. A policy knowledge graph is a graph-structured database that stores policy nodes, attributes, and relationships. Entity linking results refer to the set of matched target nodes and their structured attribute data in the policy knowledge graph. The retrieval result set refers to the set of relevant data extracted from a structured semantic index or the policy knowledge graph.

[0040] S4. Input the search result set into the search enhancement generation model, and use the search enhancement generation model to call the education-specific language model to fuse the search result set and generate an analysis report containing data source citation information and dynamic visualization charts.

[0041] Among them, the retrieval-enhanced generative model refers to an analysis report generation engine that integrates retrieval results with a language model. The education-specific language model refers to a natural language generation model pre-trained on education-related corpora. The analysis report refers to a structured output document that includes at least data source citation information and dynamic visualization charts. Data source citation information refers to the policy version number and document identifier information indicating the data source. Dynamic visualization charts refer to data charts generated in real time based on an education management indicator coordinate system.

[0042] The technical solution of this embodiment can solve industry pain points such as inaccurate intent recognition, low efficiency of cross-document analysis, and inability to dynamically combine the latest policies in the traditional interaction mode, significantly improving the efficiency and quality of data report interaction in the education field and optimizing the user interaction experience.

[0043] In one alternative approach, a large language model is used to determine whether the natural language problem is an educational domain problem, and the determination result includes:

[0044] Based on the large language model, the text feature vector of the natural language question is extracted, the current session context feature vector associated with the natural language question is captured, the behavioral feature vector corresponding to the natural language question in the user operation log is parsed, and the domain keyword matching degree of the natural language question is calculated in combination with a predefined education domain-specific dictionary.

[0045] Among them, the text feature vector refers to the numerical representation of the natural language problem after vectorization. The current conversation context feature vector refers to the numerical representation of the user's current dialogue state. The user operation log refers to the data set recording the user's historical interaction operations. The behavior feature vector refers to the numerical representation of behavioral patterns extracted from the user operation log. The predefined education domain-specific dictionary refers to a customized thesaurus containing professional terms in education management. The domain keyword matching degree refers to the score of the overlap between the natural language problem and the terminology in the education domain-specific dictionary.

[0046] Specifically: ① Based on the large language model, a text feature vector extraction operation is performed. This operation inputs the natural language question into a pre-trained word embedding layer and generates a text feature vector through weighted averaging of word vectors. ② Based on the large language model, a context feature capture operation is performed. This operation acquires the continuous dialogue sentences preceding the natural language question in the current dialogue window, encodes the dialogue state through a bidirectional long short-term memory network, and outputs the current conversation context feature vector. ③ Based on the large language model, a behavioral feature parsing operation is performed. This operation queries the user operation logs for operation records of users with the same natural language question within the most recent preset time period, counts the click frequency of policy-related questions, the number of chart downloads, and the frequency of historical query keywords, and generates a behavioral feature vector through normalization. ④ Based on the large language model, a domain matching degree calculation operation is performed. This operation compares the segmented natural language question with a predefined educational domain-specific dictionary, and calculates the ratio of the total number of matching terms to the total number of words in the question as the domain keyword matching degree.

[0047] Multimodal judgment features are generated by fusing the text feature vector, the context feature vector, and the behavior feature vector, and semantic relevance score is calculated based on the matching degree between the multimodal judgment features and the domain keywords.

[0048] Among them, multimodal judgment features refer to the joint representation that integrates text feature vectors, context feature vectors, and behavioral feature vectors. Semantic relevance score refers to a numerical score that quantifies the relevance of natural language problems to the education field.

[0049] Specifically: ① The text feature vector, context feature vector, and behavioral feature vector are input into a pre-trained multimodal feature fusion layer for feature concatenation and dimensionality reduction, generating a multimodal judgment feature vector with uniform dimensions. ② The domain keyword matching degree is normalized to generate a scalar matching degree coefficient. ③ The multimodal judgment feature vector is input into a fully connected layer for linear transformation, outputting an initial semantic relevance value. Simultaneously, the matching degree coefficient is input into a preset weight adjustment function to generate a dynamic weighting factor. ④ The initial semantic relevance value and the dynamic weighting factor are multiplied by a dot product to obtain the semantic relevance score. It should be noted that the dynamic weighting factor increases exponentially with the matching degree coefficient, ensuring that the matching degree of educational domain terms has a non-linear enhancement effect on the final score.

[0050] When the semantic relevance score is greater than a first threshold, the natural language question is determined to be an education domain question; when the semantic relevance score is less than a second threshold, the natural language question is determined to be an irrelevant question; when the semantic relevance score is between the first threshold and the second threshold, the user is guided to confirm whether the natural language question is an education domain question based on an interactive interface.

[0051] The first threshold refers to the lower limit of the semantic relevance score for determining whether a question belongs to the education domain. The second threshold refers to the upper limit of the semantic relevance score for determining whether a question belongs to an unrelated domain. The interactive interface refers to a visual operation interface that guides users to confirm the domain to which the question belongs, for example: a pop-up dialog box asks "Is your question 'Cooperating Enterprises of University N' a question in the education domain?", and provides a [Yes / No] selection button.

[0052] Specifically: ① The semantic relevance score is compared numerically with a predefined first threshold. When the semantic relevance score is greater than the first threshold, the natural language question is determined to be an education domain question. ② The semantic relevance score is compared numerically with a predefined second threshold. When the semantic relevance score is less than the second threshold, the natural language question is determined to be an irrelevant question. ③ When the semantic relevance score is between the first and second thresholds, a domain confirmation request instruction (containing the text content of the natural language question and preset confirmation options) is generated through a built-in interactive interface. The domain confirmation request instruction is pushed to the user's display interface, and the domain determination result based on the user's feedback based on the confirmation options is received. The question domain status identifier is updated according to the domain determination result.

[0053] Among the above optional methods, a large language model is further used to extract text, context, and behavioral feature vectors and combine them with a domain dictionary to calculate the matching degree. Multimodal judgment features and semantic relevance scores are then fused to dynamically determine whether a question is an educational question, accurately distinguish between relevant and irrelevant questions, reduce the occupation of system resources by irrelevant questions, improve the response speed and answer accuracy of key questions, and enhance the user interaction experience.

[0054] In one alternative approach, the natural language question is semantically parsed using the large language model to generate a structured query instruction containing question type labels, core entities, and association analysis dimensions, including:

[0055] Based on the large language model, the core entity is obtained by performing education domain entity recognition on the natural language problem, the association analysis dimension is obtained by performing association relationship analysis on the natural language problem, and the problem type label is obtained by performing semantic structure recognition on the natural language problem.

[0056] The structured query instruction is generated based on the core entity, the association analysis dimension, and the question type label.

[0057] Specifically: ① Input the core entity, time-dimensional data of the association analysis dimension, topic-dimensional data, and issue type tags into the predefined instruction building engine. The instruction building engine maps the core entity to the main search key, converts the time-dimensional data into time-range filtering conditions, and parses the topic-dimensional data into a set of topic tags. ② Select the corresponding instruction template based on the issue type tag. When the issue type tag is for cross-document association analysis, the first template structure is called; when the issue type tag is for policy association analysis, the second template structure is called. The main search key, time-range filtering conditions, and topic tag set are populated according to the field rules of the selected template. ③ The structured instruction encoder performs standardized encapsulation on the populated template, generating machine-parseable JSON format instructions as structured query instructions. The JSON format instructions contain fixed fields: "entity" stores the core entity, "time_range" stores the time-range filtering conditions, "topic_tags" stores the topic tag set, and "analysis_type" stores the issue type tag value. The "time_range" and "topic_tags" fields together constitute the complete representation of the association analysis dimension in the JSON instruction.

[0058] Among the above optional methods, a large language model is further used to perform entity recognition, relational analysis, and semantic structure recognition in the education field on natural language questions, generating structured query instructions containing question type labels, core entities, and relational analysis dimensions. This transforms natural language questions into searchable structured instructions, improving the efficiency and accuracy of question parsing, laying the foundation for subsequent accurate retrieval and analysis, and optimizing the interaction process.

[0059] In one optional approach, based on the large language model, the core entity is obtained by performing educational domain entity recognition on the natural language problem; the association analysis dimension is obtained by performing relational analysis on the natural language problem; and the problem type label is obtained by performing semantic structure recognition on the natural language problem, including:

[0060] Based on the large language model, the text feature vector and the current conversation context feature vector are fused to obtain a fused feature vector. Based on a pre-built education domain entity library, entity matching and disambiguation are performed on the fused feature vector to output the core entity.

[0061] Specifically: ① The text feature vector and the current session context feature vector are input into a pre-trained feature fusion layer for vector concatenation. A fully connected neural network is then used to perform dimensionality reduction on the concatenated vector, generating a unified fusion feature vector. ② The cosine similarity between the fusion feature vector and the name vector of each entity in a pre-built education domain entity database is calculated. A set of candidate entities with similarity greater than a preset matching threshold is selected. When a candidate entity set contains entities with the same name, the contextual attribute features in the fusion feature vector are extracted. The edit distances to the registered region attribute and institution type attribute of the entity with the same name are calculated, and the entity with the smallest overall distance is selected as the entity to be disambiguated. ③ The attribute labels of the entity to be disambiguated are verified using preset entity attribute verification rules. When the matching degree between the contextual attribute features contained in the fusion feature vector and the institution code attribute of the entity to be disambiguated is less than a preset disambiguation threshold, the hierarchical retrieval mechanism of the entity database is activated. Backtracking matching is performed along the educational institution affiliation tree until the attribute matching conditions are met, and the core entity is output. The hierarchical retrieval mechanism uses a preset educational administrative affiliation graph as the backtracking path.

[0062] Based on the large language model, the semantic dependency structure in the natural language problem is analyzed to generate a preliminary association dimension. The historical interaction patterns in the behavioral feature vector are parsed to correct the preliminary association dimension and output the association analysis dimension.

[0063] Specifically: ① The core predicates and modification relationships in natural language problems are identified using a pre-built dependency parser in the large language model. Temporal adverbs are extracted and mapped to time dimension data, and object qualifiers are extracted and mapped to topic dimension data, combining these to generate a preliminary association dimension. ② The frequency distribution of historical query keywords and historical filtering condition preference parameters stored in the behavioral feature vector are extracted. ③ The similarity of the historical query keyword frequency distribution with the topic dimension data of the preliminary association dimension is compared. When the historical frequency of a specific topic keyword is greater than a preset topic threshold, the weight of that topic in the dimension is increased. ④ The historical filtering condition preference parameters are matched with the time range of the preliminary association dimension. When time windows overlap, the boundary precision of the time dimension is adjusted based on the preference parameters. ⑤ The corrected time dimension weights and topic dimension weights are balanced using a dimension weight normalization layer, outputting an association analysis dimension containing a standardized weight distribution. The time dimension data includes a start time point and an end time point, and the topic dimension data includes at least one topic label and its corrected weight coefficient.

[0064] Based on the large language model, and combined with the domain keyword matching degree and the multimodal judgment features, the problem type label is output through a preset educational problem type classifier.

[0065] Specifically: ① First, the domain keyword matching degree is input into the feature transformation layer for vectorization processing to generate a keyword matching feature vector. The keyword matching feature vector is then merged with the multimodal judgment feature input into the feature concatenation layer to generate a classification input feature vector. ② The classification input feature vector is input into a preset education problem type classifier. The classifier's multi-layer neural network performs a non-linear feature transformation, outputting a probability distribution vector containing probability values ​​for both cross-document association analysis and policy association analysis. ③ The problem type corresponding to the highest probability value in the probability distribution vector is selected as the problem type label. It should be noted that the output layer of the education problem type classifier uses a softmax function to generate the probability distribution, and the domain keyword matching degree uses a preset exponential weighting coefficient to non-linearly enhance the probability value of the education policy category.

[0066] Among the above-mentioned optional methods, the text and context feature vectors are further integrated through a large language model, and the core entities are output by entity matching and disambiguation in combination with an entity database in the education field; the semantic dependency structure and historical interaction patterns are analyzed to output the association analysis dimension; and the question type label is output by combining the domain keyword matching degree and multimodal judgment features. The key elements of natural language questions are accurately extracted, and structured query instructions are generated to improve the efficiency and accuracy of question parsing and enhance the accuracy and intelligence of education data report interaction.

[0067] In one optional approach, when the question type tag of the structured query instruction is cross-document association analysis, a pre-built structured semantic index is retrieved based on the core entity and the association analysis dimension to generate a retrieval result set, including:

[0068] Based on the core entity, entity anchoring index matching is performed to locate the set of target documents in the structured semantic index that correspond to the core entity.

[0069] Specifically: ① Input the core entity into a pre-built structured semantic index and perform index matching. Retrieve a candidate document set containing the core entity name using an inverted index mechanism. Calculate the entity mention frequency and position weight for each document in the candidate document set to generate an initial relevance score. ② When the initial relevance score is less than a preset anchoring threshold, extract the synonym set and abbreviation variations of the core entity. Perform an extended search in the entity alias field of the index to obtain a supplementary document set. ③ Combine the candidate document set and the supplementary document set into an intermediate document set. Remove duplicate document identifiers and retain documents in the intermediate document set whose relevance score is greater than a preset effective threshold. Output a target document set containing document identifiers and corresponding entity relevance scores. The entity relevance score is calculated by weighting the initial relevance score with fuzzy matching weights.

[0070] Multidimensional association index expansion is performed based on the association analysis dimension. A subset of documents in the target document set that meet the preset time window is extracted according to the time dimension in the association analysis dimension. Each semantic fragment in the document subset that is consistent with the topic tag is matched according to the topic dimension in the association analysis dimension.

[0071] Specifically: ① Extract the start and end times from the time dimension data in the association analysis dimension. Retrieve the publication date field from the document metadata of the target document set, and filter all documents whose publication dates fall within the range of the start and end times to generate a document subset that meets the preset time window. ② Parse the topic dimension data in the association analysis dimension to obtain the topic tag set and its weight coefficients. Traverse each document in the document subset, and match segments that match the topic tags in the pre-segmented semantic fragments of the document. Specifically, calculate the cosine similarity between the semantic fragment embedding vector and the topic tag vector. When the similarity is greater than the preset topic matching threshold, it is considered a match. ③ Output all successfully matched semantic fragments in the document subset and their respective document identifiers. Here, the semantic fragments are text blocks pre-generated by the document semantic segmentation algorithm, and each text block is labeled with the location information and topic tag attributes in the original document.

[0072] Based on the joint weights of the core entity and the association analysis dimension, all semantic segments are ranked for relevance, and the semantic segments with weight values ​​greater than the third threshold in the ranking results are selected to generate the initial retrieval set.

[0073] Specifically: ① The entity relevance score for each semantic segment is calculated by weighting the frequency and position of mention of the core entity in the corresponding document; simultaneously, the weight coefficient of the topic dimension in the association analysis is extracted as the topic relevance coefficient. ② The entity relevance score and the topic relevance coefficient are input into a preset weight fusion function to generate a joint weight value for each semantic segment. The weight fusion function uses the geometric mean of the entity relevance score and the topic relevance coefficient to achieve non-linear fusion. ③ All semantic segments are sorted in descending order according to their joint weight values ​​to generate a sorted result list. ④ Semantic segments in the sorted result list whose joint weight value is greater than a predefined third threshold are selected and grouped and aggregated according to their original document affiliation to generate an initial retrieval set. The third threshold is a configurable parameter that is dynamically adjusted based on historical retrieval accuracy.

[0074] The initial retrieval set is expanded by performing associated entity expansion using a pre-constructed knowledge graph in the education domain. Semantic fragments corresponding to extended entities that have a preset association with the core entity are added to the initial retrieval set to generate the retrieval result set.

[0075] Specifically: ① Locate the nodes corresponding to the core entities in the knowledge graph of the education domain, traverse directly connected nodes along the preset association edges to obtain the extended entity set. The preset associations include institutional affiliation, subject collaboration, and project management subordination. ② Retrieve semantic fragments corresponding to the extended entity set from the structured semantic index. Specifically, perform entity anchoring index matching for each extended entity and filter semantic fragments that meet the time window constraints and topic tag matching conditions in the association analysis dimension. ③ Perform a union operation between the filtered semantic fragments and the initial retrieval set, and remove duplicates using document identifiers and fragment position information to form the extended retrieval set. ④ Output a retrieval result set containing the content of the original initial retrieval set and the newly added extended entity semantic fragments. The association degree of the extended entity semantic fragment inherits the graph relationship weight value between its corresponding extended entity and the core entity.

[0076] Among the above optional methods, efficient retrieval of cross-document association analysis can be further achieved, improving the efficiency and quality of educational data report interaction and meeting the cross-document knowledge retrieval needs of complex problems.

[0077] In one optional approach, semantic similarity is calculated on the search results to obtain semantic similarity calculation results, including:

[0078] Each semantic fragment in the search results is converted into a corresponding policy semantic vector using a pre-trained vocabulary library specifically for education policies.

[0079] Specifically: ① Perform word segmentation on each semantic segment in the search results to generate a word sequence. ② Input each word into the education policy-specific word vector library and perform vector mapping to obtain the corresponding word vector. If a word does not exist in the word vector library, it is decomposed using a sub-word segmentation algorithm, and the mean of the sub-word vectors is obtained as a substitute. ③ Perform a weight allocation operation on each word vector in the word sequence, adjusting the word vector weights based on a pre-calculated inverse document frequency factor. The weight enhancement coefficient for specialized terminology words is a specified value from a pre-defined education policy terminology table. ④ Perform a weighted average calculation on all weighted word vectors to generate a fixed-dimensional policy semantic vector. This policy semantic vector retains the policy text feature distribution characteristics of the original semantic segments.

[0080] The entity vector of the core entity is fused with the dimension vector of the association analysis dimension to generate a policy query vector.

[0081] Specifically: ① First, the dimensional vector of the association analysis dimension is decomposed into a topic dimension vector and a time dimension vector. A pre-defined feature concatenation layer linearly concatenates the entity vector, topic dimension vector, and time dimension vector to form an intermediate fused vector. ② The intermediate fused vector is input into a pre-trained policy feature cross-network. This network learns the interaction relationship between entity features and dimensional features through fully connected layers, outputting a cross-feature vector. During the training phase, the policy feature cross-network uses educational policy question-and-answer samples to optimize feature interaction weights. ③ Layer normalization is performed on the cross-feature vector to generate a policy query vector with unified dimensions. The policy query vector simultaneously encodes the semantic attributes of core entities, topic association characteristics, and time constraint features.

[0082] Calculate the cosine similarity between the policy query vector and the policy semantic vector corresponding to each semantic segment, and generate a basic similarity score for each semantic segment.

[0083] Specifically: ① Calculate the dot product of the unit vector of the policy query vector and the unit vector of each policy semantic vector. This dot product result is the cosine similarity value. ② Map the cosine similarity value to a preset similarity scoring interval to generate a basic similarity score in the range [0,1]. When the dot product result is less than 0, it is forcibly mapped to 0 points.

[0084] It should be noted that the policy query vector is denoted as . , No. The policy semantic vector of each semantic fragment is: Basic similarity score The calculation expression is: ;in, The policy query vector is generated by fusing entity vectors and correlation analysis dimension vectors. Indicates the first The policy semantic vectors corresponding to each semantic fragment. The above expression calculates cosine similarity through vector normalization and dot product operation, and constrains the output to be a non-negative standardized score.

[0085] By using the education policy synonym conversion rule base, it is detected whether each semantic segment contains a synonymous term corresponding to the association analysis dimension. When a synonymous term is detected in any semantic segment, the basic similarity score corresponding to the semantic segment is corrected to obtain the corrected similarity score corresponding to the semantic segment, until the corrected similarity score corresponding to each semantic segment is obtained.

[0086] Specifically: ① Perform a synonym detection operation on each semantic segment. Scan the word sequence in the semantic segment using a pre-built education policy synonym conversion rule base. If a word belongs to the set of synonyms of standard terms in the rule base, it is considered a successful match, and the matched synonym and its preset weight value are recorded. ② Calculate the sum of the weight values ​​of all matching synonyms in the current semantic segment as the synonym impact factor. ③ Then input the synonym impact factor into a preset gain function to generate a non-linear correction amount. The gain function uses a logarithmic function to constrain the correction magnitude, ensuring that high-weight terms do not lead to over-correction. ④ Add the non-linear correction amount to the basic similarity score to generate a preliminary correction score. ⑤ Perform an upper limit truncation operation on the preliminary correction score. When the score exceeds the preset maximum similarity value, it is forcibly set to the maximum value, generating the corrected similarity score corresponding to the semantic segment. ⑥ Repeat steps ① to ⑤ until the corrected similarity score corresponding to each semantic segment is obtained.

[0087] It should be noted that the expression for the nonlinear correction is: The expression for correcting the similarity score is: ;in, Indicates the first The set of synonyms detected in a semantic segment Indicating synonyms Preset weight values ​​in the rule base; This represents the preset gain coefficient, used to control the scaling of the effect of synonyms; Indicates the first The basic similarity score of each semantic segment Indicates the first Nonlinear correction amount for each semantic segment This represents the preset maximum similarity threshold (with a value of 1.0). Indicates the first The corrected similarity score of each semantic segment.

[0088] Based on the historical attention weights contained in the behavioral feature vector, the modified similarity score corresponding to each semantic segment is dynamically weighted to generate the semantic similarity calculation result containing the similarity values ​​of all semantic segments.

[0089] Specifically: ① Extract historical attention weight parameters from the behavioral feature vector. These parameters include the historical click frequency distribution and historical query duration ratio of users for specific policy topics. ② Establish a topic weight mapping table, matching the policy topic tag corresponding to each semantic fragment with the historical attention weight parameters to obtain the initial attention value of the topic. ③ Input the initial attention value into a preset time decay function to generate a dynamic attention factor. The time decay function adjusts the weight intensity based on the most recent operation timestamp to ensure that recent operations receive higher weight. ④ Convert the dynamic attention factor into a standardized weighted coefficient using a sigmoid activation function, with the coefficient value constrained to the range [0.5, 1.5]. ⑤ Multiply the standardized weighted coefficient by the corrected similarity score of the corresponding semantic fragment to generate a weighted similarity value. ⑥ Repeat the above steps to process all semantic fragments, and output the semantic similarity calculation result containing the weighted similarity values ​​of all semantic fragments.

[0090] It should be noted that the expression for the dynamic attention factor is: The expression for the standardized weighting coefficients is: The expression for the weighted similarity value is: ;in, Indicates the first The initial attention value of each semantic fragment corresponding to the policy topic (calculated by weighting historical click frequency and query duration). Indicates the current timestamp. Indicates the first Each semantic fragment corresponds to the last operation timestamp of the policy topic. This represents the time decay coefficient (a preset positive constant). Indicates the first Each semantic fragment corresponds to a dynamic attention factor for policy topics. This represents the slope parameter of the S-shaped function (preset as a positive constant). This represents the median threshold for attention (a preset constant). Indicates the first The standardized weighted coefficients of each semantic fragment corresponding to the policy theme. Indicates the first The weighted similarity value of each semantic segment.

[0091] Among the above-mentioned optional methods, further precise measurement of the semantic similarity between search results and questions can improve the accuracy of policy correlation analysis, provide a reliable basis for generating high-quality education data reports, and enhance the professionalism and reliability of the interaction.

[0092] In one alternative approach, the core entity is used to match nodes in a real-time accessed policy knowledge graph via entity linking technology to generate entity linking results, including:

[0093] Based on the entity type and attribute characteristics of the core entities, multi-level node matching is performed in the policy knowledge graph.

[0094] Specifically: ① Select the starting matching level based on the entity type of the core entity. When the entity type is an educational institution, matching starts at the institution level; when the entity type is a policy or regulation, matching starts at the policy level. ② Perform attribute matching at the selected level, calculating the similarity between the core entity's attribute features and the attributes of knowledge graph nodes. This includes name string similarity, institution code consistency, and geographic coordinate distance calculation, generating a primary matching node set. ③ Perform a drill-down operation on the primary matching node set, traversing the preset hierarchical path of the knowledge graph to the sub-levels. Repeat the attribute matching operation in the next level node to generate a secondary matching node set. ④ Perform a cross-level node aggregation operation, merging the matching node results from all levels to generate a candidate node set containing the complete hierarchical path. The preset hierarchical path follows the tree-like organizational structure of the education administration system.

[0095] When multiple candidate nodes are matched, the node confidence score of each candidate node is calculated by combining the policy timeliness constraint in the association analysis dimension with the historical attention preference of the behavioral feature vector. The candidate nodes are then ranked according to their node confidence scores, and candidate nodes with confidence scores greater than a fourth threshold are selected as target policy nodes, and a node relationship topology graph is generated.

[0096] The node relationship topology includes each target policy node and its corresponding policy derivative nodes associated within a preset time interval.

[0097] Specifically: ① Obtain policy timeliness constraint parameters from the association analysis dimension. These parameters include the policy effective time interval requirement and the policy effectiveness decay coefficient. Simultaneously, analyze historical attention preference data from the behavioral feature vector. This data includes the historical click frequency distribution and historical query duration ratio of users for specific policy topics. ② Calculate the time validity score of the policy corresponding to each candidate node based on the policy timeliness constraint parameters. Specifically, if the policy release time of the candidate node is within the policy effective time interval, a basic timeliness score is assigned; if it exceeds the time interval, an exponential decay correction is applied based on the excess time according to the policy effectiveness decay coefficient to generate a time validity score. ③ Match historical attention preference data with the policy topic tags of the candidate nodes to extract the historical attention intensity value of the corresponding topic. Map the historical attention intensity value to a standardized attention coefficient using a preset preference transformation function. ④ Integrate the time validity score and the standardized attention coefficient to generate an initial confidence value. Specifically, a non-linear fusion is achieved using the weighted geometric mean of the time validity score and the standardized attention coefficient. Simultaneously, overlay the topological centrality index of the candidate node in the policy knowledge graph. This index is calculated by the normalized sum of the node's in-degree and out-degree. ⑤ The initial confidence value and the topological centrality index are linearly weighted and summed to generate the final node confidence score. The weighting is dynamically allocated according to the question type label, and the weight of the standardized attention coefficient is increased when the question type label is policy correlation analysis. ⑥ All candidate nodes are sorted in descending order according to the node confidence score. Candidate nodes with a node confidence score greater than the preset fourth threshold are selected as target policy nodes. ⑦ Using the target policy node as the root node, the directly related policy derivative nodes within the preset time interval are traversed along the relationship edges of the policy knowledge graph. The node attributes and relationship type data of the root node and derivative nodes are extracted to construct a network graph containing the node relationship topology. The selection of policy derivative nodes must satisfy the constraint that their release time overlaps with the effective period of the root node's policy.

[0098] Based on the structured attribute data stored in the nodes of the policy knowledge graph, the policy effectiveness identifier and policy impact factor corresponding to each target policy node are extracted to generate the structured entity link results.

[0099] Specifically: ① First, access the structured attribute dataset stored in the policy knowledge graph for the target policy node. This dataset includes the policy status field, effective time field, repeal time field, and impact factor index field. ② Parse the value of the policy status field. If the value is a valid status identifier and the current system time is within the interval defined by the effective time and repeal time fields, then set the policy validity identifier to a valid status. If the policy status field is a repeal status identifier or the current system time exceeds the repeal time field, then set the policy validity identifier to a repeal status. If the policy status field is a pending status identifier, perform a status linkage verification based on the revision records of the associated policy nodes, and update the policy validity identifier based on the verification results. ③ Extract the basic impact factor value from the impact factor index field. This value includes the number of policy coverage areas, the fiscal appropriation scale index, and the number of regulated institutions. Simultaneously, retrieve the policy derivative nodes associated with the target policy node within a preset time interval, and statistically analyze the policy level weight and revision frequency of the derivative nodes. ④ The basic impact factor value is weighted and adjusted according to the hierarchical weight of the policy-derived nodes to generate a primary impact factor; then, a dynamic adjustment coefficient corresponding to the policy revision frequency is superimposed, which is positively correlated with the revision frequency and subject to a preset upper limit. ⑤ The primary impact factor is multiplied by the dynamic adjustment coefficient to generate the final policy impact factor; if the policy effectiveness is marked as abolished, a time-based attenuation penalty is applied to the final policy impact factor, causing it to decrease according to a preset attenuation rate. ⑥ The unique identifier of the target policy node, the policy effectiveness identifier, the final policy impact factor, and the node's original attribute data are encapsulated to generate an entity link result conforming to a preset structured format; this result is stored in the form of a machine-resolvable data object, containing fixed fields to record the policy effectiveness status code and impact factor value.

[0100] Among the above-mentioned optional methods, we can further deepen the analysis of policy correlations, improve the quality and professionalism of education data report interaction, and help users grasp the policy correlations and impacts.

[0101] In one optional approach, the semantic similarity calculation result and the entity link result are subjected to cross-modal fusion processing to obtain the retrieval result set, including:

[0102] Based on the policy effectiveness identifier in the entity linking results, the similarity value in the semantic similarity calculation results is corrected for timeliness.

[0103] Specifically: ① Identify the policy validity identifier corresponding to each target policy node in the entity link results. This identifier includes three values: valid, repealed, or pending. Simultaneously, obtain the semantic similarity value of the semantic fragment associated with the target policy node from the semantic similarity calculation results. ② If the policy validity identifier is valid, check if there is a revised version of the policy node. If a revised version exists, calculate the version decay coefficient based on the time difference between the revised version's release time and the current query time. Multiply the original similarity value of the semantic fragment by the product of a preset gain factor and the version decay coefficient to generate a corrected similarity value. The gain factor is dynamically adjusted according to the number of policy coverage areas. ③ If the policy validity identifier is repealed, extract the policy repeal time point. Calculate the time interval (in days) between the repeal time point and the current query time. Perform a non-linear mapping on the time interval (in days) using a preset time decay function to generate a repeal decay coefficient. Multiply the original similarity value of the semantic fragment by the repeal decay coefficient to generate a corrected similarity value. The time decay function satisfies the characteristic that the coefficient approaches zero as the interval (in days) increases. ④ If the policy effectiveness is marked as pending, the revision-related nodes of that node in the policy knowledge graph are retrieved in conjunction with the policy knowledge graph; the proportion of valid state nodes among the related nodes is used as the state confidence level; the original similarity value is multiplied by the state confidence level and then a preset pending compensation value is added to generate the corrected similarity value; the pending compensation value is adjusted inversely according to the magnitude of the policy impact factor. ⑤ Boundary constraint processing is performed on the corrected similarity value to ensure that its value is within the preset similarity value range; the correction result is updated to the semantic similarity calculation result dataset to form a time-sensitive corrected similarity value sequence; this sequence retains the identification information of the original semantic fragment and the correction operation record.

[0104] A policy entity aggregation network is constructed based on the node relationship topology graph of the entity linking results. The target policy nodes and policy-derived nodes in the entity linking results are mapped to network nodes through the policy entity aggregation network, and the semantic fragments in the semantic similarity calculation results are mapped to associated text nodes.

[0105] Specifically: ① Parse the node relationship topology graph structure data of the entity linking results. This data includes the target policy node set, the policy derivative node set, and the relationship types between nodes. Using each target policy node as the root node, create a corresponding policy master node in the policy entity aggregation network, inheriting the attribute data of the original node. ② Traverse the policy derivative nodes directly connected to the root node in the node relationship topology graph, create derivative nodes in the policy entity aggregation network, and label their derivative hierarchy relationship with the root node. The attribute data of the derivative nodes is synchronously obtained from the policy knowledge graph. ③ Establish network connection edges between policy master nodes and derivative nodes. The weight value of this edge is determined according to the preset weight coefficient of the relationship type in the topology graph. If multiple derivative nodes have mutual relationships, secondary connection edges between derivative nodes are synchronously established. ④ Extract semantic fragment data from the semantic similarity calculation results, create an independent text node for each semantic fragment, and establish an association mapping between the text nodes and the policy master nodes in the policy entity aggregation network. Specifically, this is based on matching the policy version number corresponding to the semantic fragment with the unique identifier of the policy master node. ⑤ Simultaneously detect whether a text node involves the content of multiple policy-derived nodes. If content association exists, establish supplementary connection edges between the text node and the corresponding derived node. The initial weight value of these supplementary connection edges is set to the normalized result of the semantic fragment similarity value. ⑥ Perform topology optimization on the policy entity aggregation network, merge redundant nodes with the same attributes, and prune connection edges with weights less than a preset edge threshold. Generate a complete network structure containing policy master nodes, policy-derived nodes, associated text nodes, and weighted connection edges. This network structure is stored as a node-edge relationship dataset recognizable by a graph database.

[0106] The association strength value between policy nodes and text nodes in the policy entity aggregation network is calculated using a predefined multimodal alignment matrix. Text nodes with association strength values ​​greater than a fifth threshold are selected to generate a policy-related text set. This policy-related text set is then combined with the structured attribute data in the entity link results to generate the retrieval result set.

[0107] Among them, the correlation strength value integrates the weights of policy impact factors and the corrected similarity value.

[0108] Specifically: ① Load a predefined multimodal alignment matrix, which contains mapping parameters between policy attribute feature dimensions and text semantic feature dimensions; traverse each policy node and text node connection pair in the policy entity aggregation network, extracting the structured attribute feature vector of the policy node and the semantic feature vector of the text node. ② Input the attribute feature vector of the policy node into the policy feature projection layer of the matrix, outputting the dimensionality-reduced policy projection vector; simultaneously input the semantic feature vector of the text node into the text feature projection layer of the matrix, outputting the dimensionality-reduced text projection vector. ③ Calculate the dot product of the policy projection vector and the text projection vector to generate a basic association strength value; superimpose the influence factor weight coefficient of the policy node for weighted adjustment, which is obtained from the structured attribute data of the entity linking results. ④ Perform a standardization transformation on the weighted basic association strength value to make it fall within a preset numerical range; then multiply by the timeliness-corrected similarity value corresponding to the text node to generate the final association strength value; where, when the policy node is a derived node, it needs to be additionally multiplied by a preset derived level attenuation factor. ⑤ Compare the final association strength value with the preset fifth threshold, and filter all text nodes with association strength values ​​greater than the threshold; extract the complete semantic fragment content of these text nodes, group and aggregate them according to their respective policy nodes, and generate a policy-related text set. ⑥ Extract the structured attribute data of the target policy node and policy-derived nodes from the entity link results. This data includes policy effectiveness identifiers, impact factor values, and version information; match the structured attribute data with the policy-related text set according to node identifiers. ⑦ Encapsulate the matched semantic fragment content and the corresponding node's structured attribute data to generate a machine-parseable retrieval result set; this result set is stored using a nested data structure, with each policy node entry containing a set of related text fragments and their corresponding attribute fields; the retrieval result set retains the original text node's location traceability information and the policy node's effectiveness status identifier.

[0109] Among the above-mentioned optional methods, cross-modal fusion processing can be further implemented to improve the accuracy and comprehensiveness of search results, provide better input for the generation of educational data reports, and optimize the user interaction experience.

[0110] In one alternative approach, the education-specific language model is used for:

[0111] Based on the policy version number and document identifier extracted from the search results set, the data tracing reference information is generated. The policy impact factors and time-dimensional data in the search results set are mapped to a predefined education management indicator coordinate system to generate the dynamic visualization chart. The strength of the suggestion statement is adjusted by combining the historical adoption rate of the behavioral feature vector, generating decision recommendations. The analysis report is generated by integrating the policy background description, the data correlation analysis, and the decision recommendations.

[0112] Specifically: ① Parse the policy version number and document identifier data associated with each semantic fragment in the search results set, and perform matching verification in the pre-built source traceability metadata database; extract the full name, issuing agency, and effective date fields of the successfully matched policy documents, and generate standardized source traceability entries according to the preset citation format template; aggregate the source traceability entries corresponding to all semantic fragments to form a data source traceability citation information list, which is arranged in descending order of policy release time and marked with the source credibility level. ② Extract the numerical sequence of policy impact factors and corresponding time dimension data from the search results set; convert the time dimension data into the time axis scale of a predefined education management indicator coordinate system; at the same time, map the policy impact factor values ​​to the influence measurement axis of the indicator coordinate system through a linear interpolation algorithm; based on the mapped data point set, call the visualization engine to generate a dynamic line chart, which marks the policy revision event nodes in real time and supports axial zoom interaction; the chart color scheme is dynamically adjusted according to the policy effectiveness status. ③ Analyze the historical adoption rate parameter in the behavioral feature vector. This parameter records the historical adoption frequency and implementation feedback score of users for similar policy recommendations. Calculate the moving average of the historical adoption rate as the basic adjustment coefficient. Select the recommendation expression strength level according to the range of the basic adjustment coefficient. When the coefficient is greater than the adoption threshold, a strengthened expression mode is adopted; when the coefficient is less than the warning threshold, a conservative expression mode is adopted. Generate differentiated decision recommendation text by combining the changing trends of policy impact factors. ④ Integrate the three parts: policy background description, data correlation analysis, and decision recommendation. The policy background description section references the policy effectiveness identifier and impact factor analysis from the entity link results. The data correlation analysis section integrates semantic similarity calculation logic and cross-document retrieval conclusions. The decision recommendation section inserts dynamically generated recommendation text and strength identifiers. The three parts are structurally assembled according to a preset template through the report generation engine, and data source citation information footnotes and dynamic visualization charts are added to generate the final analysis report. This report retains machine-readable metadata tags for version tracking.

[0113] Among the above-mentioned optional methods, data traceability, visualization, and decision support can be further integrated to enhance the depth and value of education data reports and improve the professionalism and practicality of user interaction.

[0114] Figure 2 This diagram illustrates the structure of an embodiment of an interactive system 200 for generating educational data reports based on retrieval enhancement, provided by the present invention. Figure 2 As shown, the system 200 includes: a problem judgment module 210, a semantic parsing module 220, a result generation module 230, and a content interaction module 240;

[0115] The question judgment module 210 is used to: receive a natural language question input by the user, and determine whether the natural language question is an educational question through a large language model, and obtain a judgment result;

[0116] The semantic parsing module 220 is used to: when the judgment result is yes, perform semantic parsing on the natural language question through the large language model to generate a structured query instruction containing question type labels, core entities and association analysis dimensions;

[0117] The result generation module 230 is used to: when the question type tag of the structured query instruction is cross-document association analysis, retrieve a pre-built structured semantic index based on the core entity and the association analysis dimension, and generate a search result set; when the question type tag of the structured query instruction is policy association analysis, retrieve the structured semantic index based on the core entity and the association analysis dimension, perform semantic similarity calculation on the search results, obtain semantic similarity calculation results, and use the core entity to match nodes in the real-time accessed policy knowledge graph through entity linking technology to generate entity linking results, and perform cross-modal fusion processing on the semantic similarity calculation results and the entity linking results to obtain the search result set;

[0118] The content interaction module 240 is used to: input the search result set into the search enhancement generation model, and use the search enhancement generation model to call the education-specific language model to fuse the search result set and generate an analysis report containing data source citation information and dynamic visualization charts.

[0119] In an alternative embodiment, the problem-determining module 210 is specifically used for:

[0120] Based on the large language model, the text feature vector of the natural language question is extracted, the current session context feature vector associated with the natural language question is captured, the behavioral feature vector corresponding to the natural language question in the user operation log is parsed, and the domain keyword matching degree of the natural language question is calculated in combination with a predefined education domain-specific dictionary;

[0121] Multimodal judgment features are generated by fusing the text feature vector, the context feature vector, and the behavior feature vector, and a semantic relevance score is calculated based on the matching degree between the multimodal judgment features and the domain keywords.

[0122] When the semantic relevance score is greater than a first threshold, the natural language question is determined to be an education domain question; when the semantic relevance score is less than a second threshold, the natural language question is determined to be an irrelevant question; when the semantic relevance score is between the first threshold and the second threshold, the user is guided to confirm whether the natural language question is an education domain question based on an interactive interface.

[0123] In an alternative embodiment, the semantic parsing module 220 is specifically used for:

[0124] Based on the large language model, the core entity is obtained by performing educational domain entity recognition on the natural language problem, the association analysis dimension is obtained by performing association relationship analysis on the natural language problem, and the problem type label is obtained by performing semantic structure recognition on the natural language problem.

[0125] The structured query instruction is generated based on the core entity, the association analysis dimension, and the question type label.

[0126] In an alternative embodiment, the result generation module 230 is specifically used for:

[0127] Based on the large language model, the text feature vector and the current conversation context feature vector are fused to obtain a fused feature vector. Based on a pre-built education domain entity library, entity matching and disambiguation are performed on the fused feature vector to output the core entity.

[0128] Based on the large language model, the semantic dependency structure in the natural language problem is analyzed to generate a preliminary association dimension. The historical interaction patterns in the behavioral feature vector are parsed to correct the preliminary association dimension and output the association analysis dimension.

[0129] Based on the large language model, and combined with the domain keyword matching degree and the multimodal judgment features, the problem type label is output through a preset educational problem type classifier.

[0130] In an alternative embodiment, the result generation module 230 is specifically used for:

[0131] Based on the core entity, perform entity anchoring index matching to locate the set of target documents in the structured semantic index library that correspond to the core entity;

[0132] Multidimensional association index expansion is performed based on the association analysis dimension. A subset of documents in the target document set that meet the preset time window is extracted according to the time dimension in the association analysis dimension. Each semantic fragment in the document subset that is consistent with the topic tag is matched according to the topic dimension in the association analysis dimension.

[0133] Based on the joint weight of the core entity and the association analysis dimension, all semantic segments are ranked by relevance, and the semantic segments with weight values ​​greater than the third threshold in the ranking results are selected to generate an initial retrieval set.

[0134] The initial retrieval set is expanded by performing associated entity expansion using a pre-constructed knowledge graph in the education domain. Semantic fragments corresponding to extended entities that have a preset association with the core entity are added to the initial retrieval set to generate the retrieval result set.

[0135] In an alternative embodiment, the result generation module 230 is specifically used for:

[0136] Each semantic fragment in the search results is converted into a corresponding policy semantic vector using a pre-trained term vector library specifically for education policies.

[0137] The entity vector of the core entity is fused with the dimension vector of the association analysis dimension to generate a policy query vector;

[0138] Calculate the cosine similarity between the policy query vector and the policy semantic vector corresponding to each semantic segment, and generate a basic similarity score for each semantic segment;

[0139] By using the education policy synonym conversion rule base, it is detected whether each semantic segment contains a synonymous term corresponding to the association analysis dimension. When a synonymous term is detected in any semantic segment, the basic similarity score corresponding to the semantic segment is corrected to obtain the corrected similarity score corresponding to the semantic segment, until the corrected similarity score corresponding to each semantic segment is obtained.

[0140] Based on the historical attention weights contained in the behavioral feature vector, the modified similarity score corresponding to each semantic segment is dynamically weighted to generate the semantic similarity calculation result containing the similarity values ​​of all semantic segments.

[0141] In an alternative embodiment, the result generation module 230 is specifically used for:

[0142] Based on the entity type and attribute characteristics of the core entities, multi-level node matching is performed in the policy knowledge graph;

[0143] When multiple candidate nodes are matched, the node confidence of each candidate node is calculated by combining the policy timeliness constraint in the association analysis dimension with the historical attention preference of the behavioral feature vector.

[0144] Candidate nodes are sorted according to the node confidence level, and candidate nodes with confidence values ​​greater than the fourth threshold are selected as target policy nodes and a node relationship topology graph is generated; wherein, the node relationship topology graph includes each target policy node and the corresponding policy derivative nodes associated within a preset time interval;

[0145] Based on the structured attribute data stored in the nodes of the policy knowledge graph, the policy effectiveness identifier and policy impact factor corresponding to each target policy node are extracted to generate the structured entity link results.

[0146] In an alternative embodiment, the result generation module 230 is specifically used for:

[0147] Based on the policy effectiveness identifier in the entity linking results, the similarity value in the semantic similarity calculation results is corrected for its timeliness.

[0148] Based on the node relationship topology graph of the entity linking results, a policy entity aggregation network is constructed. Through the policy entity aggregation network, the target policy nodes and policy-derived nodes in the entity linking results are mapped to network nodes, and the semantic fragments in the semantic similarity calculation results are mapped to associated text nodes.

[0149] The association strength value between policy nodes and text nodes in the policy entity aggregation network is calculated using a predefined multimodal alignment matrix; wherein the association strength value integrates the weight of the policy influence factor and the corrected similarity value;

[0150] Text nodes with a correlation strength value greater than the fifth threshold are selected to generate a policy-related text set. The policy-related text set is then combined with the structured attribute data in the entity link results to generate the retrieval result set.

[0151] In one alternative approach, the education-specific language model is used for:

[0152] Based on the policy version number and document identifier extracted from the search result set, the data tracing reference information is generated;

[0153] The policy impact factors and time dimension data in the search results set are mapped to a predefined education management indicator coordinate system to generate the dynamic visualization chart;

[0154] By combining the historical adoption rate of the aforementioned behavioral feature vectors, the strength of the suggestion statement is adjusted to generate decision suggestions;

[0155] The analysis report is generated by integrating the policy background description, the data correlation analysis, and the decision recommendations.

[0156] It should be noted that the beneficial effects of the retrieval-enhanced educational data report content interaction system 200 provided in the above embodiments are the same as those of the retrieval-enhanced educational data report content interaction method described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0157] The educational data report content interaction system 200 based on retrieval enhancement generation of the present invention can be a computer program (including program code) running on a computer device. For example, the educational data report content interaction system based on retrieval enhancement generation of the present invention is an application software that can be used to execute the corresponding steps in the educational data report content interaction method based on retrieval enhancement generation of the present invention.

[0158] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned interactive methods for educational data report content generated based on retrieval enhancement. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the interactive method for educational data report content generated based on retrieval enhancement according to any embodiment of the present invention by calling the computer program.

[0159] In one alternative embodiment, an electronic device is provided, such as Figure 3 As shown, Figure 3 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0160] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0161] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0162] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0163] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0164] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0165] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described interactive methods for generating educational data reports based on retrieval enhancement.

[0166] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0167] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for interactive content of educational data reports generated based on retrieval enhancement, characterized in that, include: The system receives a natural language question input by the user and uses a large language model to determine whether the natural language question is an educational question, and obtains the determination result. When the judgment result is yes, the natural language question is semantically parsed through the large language model to generate a structured query instruction containing question type labels, core entities and association analysis dimensions; When the question type tag of the structured query instruction is cross-document association analysis, a pre-built structured semantic index is retrieved based on the core entity and the association analysis dimension to generate a search result set; when the question type tag of the structured query instruction is policy association analysis, the structured semantic index is retrieved based on the core entity and the association analysis dimension, semantic similarity is calculated on the search results to obtain semantic similarity calculation results, and the core entity is used to match nodes in the real-time accessed policy knowledge graph through entity linking technology to generate entity linking results, and the semantic similarity calculation results and entity linking results are subjected to cross-modal fusion processing to obtain the search result set; The search result set is input into the search enhancement generation model, which then calls a language model specifically designed for the education field to fuse the search result set and generate an analysis report containing data source citation information and dynamic visualization charts. The natural language problem is determined using a large language model to determine whether it falls under the educational domain, and the determination result includes: Based on the large language model, the text feature vector of the natural language question is extracted, the current session context feature vector associated with the natural language question is captured, the behavioral feature vector corresponding to the natural language question in the user operation log is parsed, and the domain keyword matching degree of the natural language question is calculated in combination with a predefined education domain-specific dictionary; Multimodal judgment features are generated by fusing the text feature vector, the context feature vector, and the behavior feature vector, and a semantic relevance score is calculated based on the matching degree between the multimodal judgment features and the domain keywords. When the semantic relevance score is greater than a first threshold, the natural language question is determined to be an education domain question; when the semantic relevance score is less than a second threshold, the natural language question is determined to be an irrelevant question; when the semantic relevance score is between the first threshold and the second threshold, the user is guided to confirm whether the natural language question is an education domain question based on an interactive interface.

2. The interactive method for educational data report content generated based on retrieval enhancement according to claim 1, characterized in that, The natural language question is semantically parsed using the large language model to generate structured query instructions containing question type labels, core entities, and association analysis dimensions, including: Based on the large language model, the core entity is obtained by performing educational domain entity recognition on the natural language problem, the association analysis dimension is obtained by performing association relationship analysis on the natural language problem, and the problem type label is obtained by performing semantic structure recognition on the natural language problem. The structured query instruction is generated based on the core entity, the association analysis dimension, and the question type label.

3. The interactive method for educational data report content generated based on retrieval enhancement according to claim 2, characterized in that, Based on the large language model, the core entity is obtained by performing educational domain entity recognition on the natural language problem; the association analysis dimension is obtained by performing relational analysis on the natural language problem; and the problem type label is obtained by performing semantic structure recognition on the natural language problem, including: Based on the large language model, the text feature vector and the current conversation context feature vector are fused to obtain a fused feature vector. Based on a pre-built education domain entity library, entity matching and disambiguation are performed on the fused feature vector to output the core entity. Based on the large language model, the semantic dependency structure in the natural language problem is analyzed to generate a preliminary association dimension. The historical interaction patterns in the behavioral feature vector are parsed to correct the preliminary association dimension and output the association analysis dimension. Based on the large language model, and combined with the domain keyword matching degree and the multimodal judgment features, the problem type label is output through a preset educational problem type classifier.

4. The interactive method for educational data report content generated based on retrieval enhancement according to claim 2 or 3, characterized in that, When the question type tag of the structured query instruction is cross-document association analysis, a pre-built structured semantic index is retrieved based on the core entity and the association analysis dimension to generate a search result set, including: Based on the core entity, perform entity anchoring index matching to locate the set of target documents in the structured semantic index library that correspond to the core entity; Multidimensional association index expansion is performed based on the association analysis dimension. A subset of documents in the target document set that meet the preset time window is extracted according to the time dimension in the association analysis dimension. Each semantic fragment in the document subset that is consistent with the topic tag is matched according to the topic dimension in the association analysis dimension. Based on the joint weight of the core entity and the association analysis dimension, all semantic segments are ranked by relevance, and the semantic segments with weight values ​​greater than the third threshold in the ranking results are selected to generate an initial retrieval set. The initial retrieval set is expanded by performing associated entity expansion using a pre-constructed knowledge graph in the education domain. Semantic fragments corresponding to extended entities that have a preset association with the core entity are added to the initial retrieval set to generate the retrieval result set.

5. The interactive method for educational data report content generated based on retrieval enhancement according to claim 4, characterized in that, Semantic similarity is calculated on the search results to obtain the semantic similarity calculation results, including: Each semantic fragment in the search results is converted into a corresponding policy semantic vector using a pre-trained term vector library specifically for education policies. The entity vector of the core entity is fused with the dimension vector of the association analysis dimension to generate a policy query vector; Calculate the cosine similarity between the policy query vector and the policy semantic vector corresponding to each semantic segment, and generate a basic similarity score for each semantic segment; By using the education policy synonym conversion rule base, it is detected whether each semantic segment contains a synonymous term corresponding to the association analysis dimension. When a synonymous term is detected in any semantic segment, the basic similarity score corresponding to the semantic segment is corrected to obtain the corrected similarity score corresponding to the semantic segment, until the corrected similarity score corresponding to each semantic segment is obtained. Based on the historical attention weights contained in the behavioral feature vector, the modified similarity score corresponding to each semantic segment is dynamically weighted to generate the semantic similarity calculation result containing the similarity values ​​of all semantic segments.

6. The interactive method for educational data report content generated based on retrieval enhancement according to claim 5, characterized in that, Using the core entity, entity linking technology is employed to match nodes in the real-time accessed policy knowledge graph, generating entity linking results, including: Based on the entity type and attribute characteristics of the core entities, multi-level node matching is performed in the policy knowledge graph; When multiple candidate nodes are matched, the node confidence of each candidate node is calculated by combining the policy timeliness constraint in the association analysis dimension with the historical attention preference of the behavioral feature vector. Candidate nodes are sorted according to the node confidence level, and candidate nodes with confidence values ​​greater than the fourth threshold are selected as target policy nodes and a node relationship topology graph is generated; wherein, the node relationship topology graph includes each target policy node and the corresponding policy derivative nodes associated within a preset time interval; Based on the structured attribute data stored in the nodes of the policy knowledge graph, the policy effectiveness identifier and policy impact factor corresponding to each target policy node are extracted to generate the structured entity link results.

7. The interactive method for educational data report content generated based on retrieval enhancement according to claim 6, characterized in that, The semantic similarity calculation results and the entity link results are fused across modally to obtain the retrieval result set, including: Based on the policy effectiveness identifier in the entity linking results, the similarity value in the semantic similarity calculation results is corrected for its timeliness. Based on the node relationship topology graph of the entity linking results, a policy entity aggregation network is constructed. Through the policy entity aggregation network, the target policy nodes and policy-derived nodes in the entity linking results are mapped to network nodes, and the semantic fragments in the semantic similarity calculation results are mapped to associated text nodes. The association strength value between policy nodes and text nodes in the policy entity aggregation network is calculated using a predefined multimodal alignment matrix; wherein the association strength value integrates the weight of the policy influence factor and the corrected similarity value; Text nodes with a correlation strength value greater than the fifth threshold are selected to generate a policy-related text set. The policy-related text set is then combined with the structured attribute data in the entity link results to generate the retrieval result set.

8. The interactive method for educational data report content generated based on retrieval enhancement according to claim 7, characterized in that, The education-specific language model is used for: Based on the policy version number and document identifier extracted from the search result set, the data tracing reference information is generated; The policy impact factors and time dimension data in the search results set are mapped to a predefined education management indicator coordinate system to generate the dynamic visualization chart; By combining the historical adoption rate of the aforementioned behavioral feature vectors, the strength of the suggestion statement is adjusted to generate decision suggestions; The analysis report is generated by integrating the policy background description, the data correlation analysis, and the decision recommendations.

9. A system for interactive educational data report content based on retrieval enhancement, characterized in that, include: The module includes a problem identification module, a semantic parsing module, a result generation module, and a content interaction module. The problem judgment module is used to: receive natural language questions input by users, and determine whether the natural language questions are education-related questions through a large language model, and obtain a judgment result; The semantic parsing module is used to: when the judgment result is yes, perform semantic parsing on the natural language question through the large language model to generate a structured query instruction containing question type labels, core entities and association analysis dimensions; The result generation module is used to: when the question type tag of the structured query instruction is cross-document association analysis, retrieve a pre-built structured semantic index based on the core entity and the association analysis dimension, and generate a search result set; when the question type tag of the structured query instruction is policy association analysis, retrieve the structured semantic index based on the core entity and the association analysis dimension, perform semantic similarity calculation on the search results, obtain semantic similarity calculation results, and use the core entity to match nodes in the real-time accessed policy knowledge graph through entity linking technology to generate entity linking results, and perform cross-modal fusion processing on the semantic similarity calculation results and the entity linking results to obtain the search result set; The content interaction module is used to: input the search result set into the search enhancement generation model, and use the search enhancement generation model to call the education-specific language model to fuse the search result set and generate an analysis report containing data source citation information and dynamic visualization charts; The problem judgment module is specifically used for: Based on the large language model, the text feature vector of the natural language question is extracted, the current session context feature vector associated with the natural language question is captured, the behavioral feature vector corresponding to the natural language question in the user operation log is parsed, and the domain keyword matching degree of the natural language question is calculated in combination with a predefined education domain-specific dictionary; Multimodal judgment features are generated by fusing the text feature vector, the context feature vector, and the behavior feature vector, and a semantic relevance score is calculated based on the matching degree between the multimodal judgment features and the domain keywords. When the semantic relevance score is greater than a first threshold, the natural language question is determined to be an education domain question; when the semantic relevance score is less than a second threshold, the natural language question is determined to be an irrelevant question; when the semantic relevance score is between the first threshold and the second threshold, the user is guided to confirm whether the natural language question is an education domain question based on an interactive interface.

Images