Teaching link implicit structure intelligent identification method and system based on retrieval enhancement generation

Abstract

The invention relates to the cross technical field of artificial intelligence natural language processing and education, and discloses an intelligent recognition method and system for a teaching link implicit structure generated by retrieval enhancement, and the method comprises the steps: constructing a structured teaching link knowledge base which comprises a plurality of teaching link names and a set of corresponding content fragments, generating an embedded vector for each content segment; generating a query embedding vector based on the input disordered teaching text, and retrieving an initial candidate fragment set from the knowledge base according to the similarity between the query embedding vector and the embedding vector; on the basis of the initial candidate fragment set, using a cross encoder model to calculate a correlation score of query and each candidate fragment, and reordering the initial candidate fragment set on the basis of the correlation scores to obtain a final candidate fragment set; and according to the final candidate fragment set and a preset enhancement prompt, carrying out teaching link segmentation and sequence recovery by using a lightweight large language model, and generating an ordered teaching link sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence natural language processing and education, and particularly relates to a retrieval-enhanced generated teaching link implicit structure intelligent identification method and system. BACKGROUND

[0002] With the development of educational informatization, digital teaching materials are increasingly popular. These materials contain rich but usually implicit teaching structures, including teaching goals, knowledge flow, and semantic dependency relationships between teaching stages. These structures are crucial for understanding teaching logic and implementing adaptive learning systems, but they are not explicitly annotated in most texts, especially in lesson plan analysis. Different teachers may use different link names, different content division standards, and even subjective differences in understanding teaching logic for the same teaching content. This high level of professionalism and subjectivity makes teaching link division a typical open-domain task, and traditional closed classification methods are not suitable. Existing natural language processing methods have made progress in analyzing short educational content, such as question answering and summary generation. However, these methods often struggle to capture subtle inter-stage relationships and implicit teaching processes when dealing with long teaching documents. Traditional methods excel in knowledge-intensive tasks, but their ability to model fuzzy boundary definitions and sequential dependency relationships in teaching links is limited.

[0003] Therefore, the existing technology has the following limitations: 1) it is difficult to effectively identify and segment the implicit teaching link boundaries in long teaching texts. 2) it lacks the ability to model and recover the logical sequential dependency relationships between teaching links. 3) the decision-making process of the model is not transparent, and the credibility and explainability are insufficient in educational application scenarios. 4) lightweight large language models directly processing such complex tasks have problems such as insufficient semantic understanding depth and limited long text processing capability, resulting in incomplete or disordered results. Using existing large language models will consume a lot of computing resources. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a retrieval-enhanced generated teaching link implicit structure intelligent identification method that can more accurately identify teaching stage names, segment content paragraphs, and restore their correct order consistent with teaching logic, while improving the explainability of the model output and its practicality in educational scenarios.

[0005] To achieve the above purpose, the present application provides a retrieval-enhanced generated teaching link implicit structure intelligent identification method, comprising the following steps:

[0006] Construct a structured teaching segment knowledge base, which includes a set of teaching segment names and corresponding content segments, and generate an embedding vector for each content segment;

[0007] Based on the input unordered teaching text, a query embedding vector is generated, and an initial set of candidate segments is retrieved from the knowledge base according to the similarity between the query embedding vector and the embedding vector.

[0008] Based on the initial candidate fragment set, a cross-encoder model is used to calculate the relevance score between the query and each candidate fragment, and the initial candidate fragment set is reordered based on the relevance scores to obtain the final candidate fragment set.

[0009] Based on the final candidate segment set and the preset enhancement prompts, a lightweight large language model is used to segment and restore the order of teaching segments, generating an ordered sequence of teaching segments.

[0010] Preferably, the steps for constructing the knowledge base include:

[0011] The lesson plan text is processed into fragment pairs of teaching segment names and corresponding content segments to construct an initial knowledge base;

[0012] Use an embedding model to generate an embedding vector for each content fragment in the initial knowledge base;

[0013] Based on the embedding vector, a vector index for each content fragment is constructed to form the final knowledge base.

[0014] Preferably, the step of retrieving the initial candidate fragment set includes:

[0015] Use an embedding model to convert unordered input instructional text into query embedding vectors;

[0016] Calculate the cosine similarity between the query embedding vector and the embedding vectors of each content segment in the structured teaching knowledge base;

[0017] An initial set of candidate segments is selected based on the cosine similarity.

[0018] Preferably, the step of reordering the initial candidate fragment set includes:

[0019] For each candidate fragment in the initial candidate fragment set, the query and candidate fragment are concatenated and input into the cross-encoder model to obtain a relevance score;

[0020] The candidate segment set is sorted according to the relevance score to obtain the final candidate segment set;

[0021] Select the top few segments from the final candidate segment set as the output.

[0022] Preferably, the steps for generating the teaching sequence include:

[0023] The final candidate segment set is structured and extracted to generate intermediate output for teaching segmentation;

[0024] The design includes enhanced prompts that contain common process templates and sequence restoration rules for teaching procedures;

[0025] The intermediate outputs and enhanced prompts are input into a lightweight large language model to generate the final result of restoring the order of the teaching steps.

[0026] Preferably, the method further includes an evaluation step for quantitatively evaluating the generated teaching segment segmentation and sequence restoration results. The evaluation dimensions include content completeness, sequence consistency, segment boundary accuracy, and name standardization.

[0027] Preferably, the formula for calculating the content integrity is:

[0028] ;

[0029] Where N represents the total number of segments in the standard answer; This represents the content of the i-th standard answer segment. represents the content of the j-th predicted segment; ROUGE-L represents the similarity calculation function based on the longest common subsequence.

[0030] Preferably, the formula for calculating the sequential consistency is:

[0031] ;

[0032] Where M represents the number of predicted segments; pos k This indicates the position of the k-th predicted segment in the standard answer. I is an indicator function that takes the value 1 when the condition is met, and 0 otherwise.

[0033] The present invention also provides an intelligent recognition system for implicit structure of teaching links in retrieval enhancement generation. The system is used to implement the above method and includes: a construction module, a retrieval module, a sorting module and a generation module.

[0034] The construction module is used to build a structured teaching segment knowledge base, which includes a set of teaching segment names and corresponding content segments, and generates an embedding vector for each content segment.

[0035] The retrieval module is used to generate a query embedding vector based on the input unordered teaching text, and retrieve an initial set of candidate segments from the knowledge base based on the similarity between the query embedding vector and the embedding vector.

[0036] The sorting module is used to calculate the relevance score between the query and each candidate segment based on the initial candidate segment set using a cross-encoder model, and to re-sort the initial candidate segment set based on the relevance score to obtain the final candidate segment set.

[0037] The generation module is used to segment and restore the order of teaching segments using a lightweight large language model based on the final candidate segment set and preset enhancement prompts, thereby generating an ordered sequence of teaching segments.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention significantly improves the overall performance of teaching process structure recognition by integrating multi-stage retrieval strategies and domain knowledge-enhanced generation. The method employs a two-stage retrieval mechanism of "coarse retrieval + fine ranking," effectively combining the efficiency of vector retrieval with the deep semantic understanding capabilities of cross-encoders. It exhibits higher accuracy and robustness when processing diverse and highly specialized lesson plan texts. By injecting structured hints containing teaching process rules into a lightweight large language model, prior knowledge from the teaching domain is explicitly introduced, enabling the model to recover the stage sequence that conforms to teaching logic, thus overcoming the shortcomings of traditional methods in modeling sequential dependencies. The entire process is clear and transparent, with traceable retrieval sources, significantly enhancing the interpretability and credibility of the model output. This method can greatly improve the performance of small language models on this complex task, enabling them to achieve near- or even surpass the performance of large models in resource-constrained environments, significantly reducing the cost of technology implementation. Simultaneously, by introducing an automated multi-dimensional evaluation system, a "generation-evaluation" technical closed loop is formed, providing a scientific basis for model iterative optimization and greatly enhancing the system's practicality and reliability. Attached Figure Description

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

[0041] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention;

[0043] Figure 3 A schematic diagram illustrating the workflow of the knowledge base construction module for structured teaching activities provided in this embodiment of the invention;

[0044] Figure 4A schematic diagram illustrating the workflow of the first-round retrieval module provided in an embodiment of the present invention;

[0045] Figure 5 A schematic diagram illustrating the workflow of the secondary precise reordering module provided in an embodiment of the present invention;

[0046] Figure 6 This is a schematic diagram illustrating the workflow of the lightweight large language model enhancement generation module provided in an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Example 1:

[0050] This embodiment provides a method for intelligent recognition of implicit structures in teaching segments generated through retrieval enhancement. This method is used to identify and restore the order of teaching segments in unordered teaching text. The method can be defined as follows:

[0051] ;

[0052] Here, Task represents the task objective; This indicates that the unordered text D is segmented into teaching segments, resulting in unordered names of the teaching segments r. i and content c i R represents the set of teaching segments whose order is restored from the partitioned unordered set to an ordered set. j and c j A collection of content; the method flow is as follows Figure 1 As shown, it includes the following steps:

[0053] S1. Construct a knowledge base for structured teaching components.

[0054] The names of teaching segments and their corresponding content paragraphs are extracted from high-quality teaching texts to construct a structured knowledge base of teaching segments, and a vectorized index is generated for each segment.

[0055] Specifically, the collected high-quality lesson plan texts are processed into fragment pairs of [teaching segment name, corresponding content], and a knowledge base KB is constructed, the expression of which is:

[0056] ;

[0057] Where, r i c represents the name of the i-th teaching segment, such as "Introduction to New Lesson", "Knowledge Explanation", "Classroom Exercises", etc. i This represents the detailed text content corresponding to the i-th teaching segment; n represents the total number of segments in the knowledge base. The knowledge base KB serves as the source of professional knowledge for model retrieval and provides a standardized reference for dividing teaching segments.

[0058] Vectorization and indexing are used to transform textual knowledge into vector forms that can be efficiently computed and retrieved. For each content fragment c in the knowledge base... i Generate embedding vector e ci Its expression is:

[0059] ;

[0060] in, This refers to the BGE embedding model, a lightweight text embedding model that converts text into embedding vectors in a high-dimensional space; e ci Represents content fragment c i The vector representation obtained after calculation by the BGE model has a dimension that is a hyperparameter. Through vectorization, semantically similar text content will be closer together in the vector space.

[0061] Constructing vector indexes for content fragments Its expression is:

[0062] ;

[0063] The index It stores vectors of all content fragments in the knowledge base, which are used to support subsequent fast vector similarity retrieval.

[0064] S2. Based on the input unordered teaching text, generate a query embedding vector, and retrieve an initial candidate fragment set from the knowledge base according to the similarity between the query embedding vector and the embedding vector.

[0065] This step involves inputting unordered teaching text that implicitly lacks the names of teaching segments. Through semantic vector similarity calculation, it initially retrieves relevant teaching data from the knowledge base. One candidate teaching segment.

[0066] First, the unordered instructional text input by the user is converted into a vector through query vectorization. The input unordered instructional text D is used as the query q, and it is converted into a query embedding vector e using the same BGE model as the knowledge base. qIts expression is:

[0067] ;

[0068] Among them, e q It is a vector representation of the query text q, and its dimension is the same as e. ci same.

[0069] Then, similarity calculation and coarse retrieval are performed to quickly filter candidate segments from the knowledge base that are semantically similar to the query. The query vector e is calculated. q With the vector e of each content fragment in the knowledge base ci cosine similarity Its expression is:

[0070] ;

[0071] in, The range of values ​​is A larger value indicates greater semantic similarity. This calculation utilizes vector indexing. To achieve efficient retrieval.

[0072] Based on the similarity score, return the set C of the top k most relevant candidate segments, expressed as:

[0073] ;

[0074] in, This means selecting the k segments with the highest similarity; k is a parameter, denoted as In this embodiment, it is preferably set to 10; C is the set of candidate fragments output in the first round of retrieval, which provides the basis for the fine sorting in the next stage.

[0075] S3. Based on the initial candidate fragment set, use a cross-encoder model to calculate the relevance score between the query and each candidate fragment, and reorder the initial candidate fragment set based on the relevance score to obtain the final candidate fragment set.

[0076] First, deep semantic matching is performed to accurately evaluate the relevance of the query to each candidate fragment. This is then applied to the candidate fragment set obtained in the first round of retrieval. Query q with each candidate fragment After concatenation, the data is input into a cross-encoder model to obtain the relevance score of each segment. Its expression is:

[0077] ;

[0078] in, This indicates a text concatenation operation; It is a cross-encoder model that can perform deep, global semantic interaction understanding on concatenated text pairs and output an accurate relevance score. The higher the value, the better the candidate fragment. The stronger the correlation with query q.

[0079] Next, a fine sorting and output process is performed to reorder the candidate set based on the deep semantic matching score and output the most relevant fragments. This is done according to the score. Sort the candidate set C in descending order to obtain the final sorted set of fragments. Its expression is:

[0080] .

[0081] from The top K segments are selected as the output of the exact sorted sequence. K is a hyperparameter, denoted as K. Considering that a typical lesson plan contains approximately 3 to 7 teaching segments, this embodiment... The parameter is preferably set to 5 or 6. It is a high-quality, highly relevant set of candidate fragments used in the final generation stage.

[0082] S4. Based on the final candidate segment set and the preset enhancement prompts, use a lightweight large language model to segment and restore the order of teaching segments, and generate an ordered sequence of teaching segments.

[0083] This step, based on the reordered segments and preset enhancement prompts, drives a lightweight large language model to perform two subtasks: identifying and outputting the name of the teaching segment corresponding to each segment, and restoring the correct logical order of all teaching segments to generate a structured sequence of teaching segments.

[0084] First, structured extraction is performed, organizing the finely sorted candidate fragments into an intermediate format that the model can process. This involves processing the precisely sorted fragment set. Perform structured and precise extraction to generate intermediate outputs for segmenting teaching processes. Its expression is:

[0085] ;

[0086] in, This indicates that the name of the teaching segment (r) is extracted from the candidate segment. i and content c i They are then organized into a structured form and used as one of the inputs to a large language model.

[0087] Hints are provided to guide the large language model in generating structured output that conforms to pedagogical logic. Enhanced hints are designed to explicitly include the following:

[0088] ① A common teaching process template, such as: "Introduction → Explanation → Practice → Summary → Homework".

[0089] ②The content of the fragment to be processed .

[0090] ③ Clear sequence restoration rules require the model to identify the stage name corresponding to each paragraph and sort the stages according to the teaching logic.

[0091] Will and Input a lightweight large language model to generate the final teaching sequence recovery result. Its expression is:

[0092] ;

[0093] in, This represents a lightweight, large language model. The final output is an ordered set. The teaching segment name is r j Standards, content c j The division was accurate and the order conformed to the teaching logic, thus restoring the unordered text D to a structured and orderly teaching process.

[0094] This embodiment also includes an evaluation step for quantitatively evaluating the generated teaching segment segmentation and sequence restoration results. The evaluation dimensions include content completeness, sequence consistency, segment boundary accuracy, and name standardization.

[0095] In this embodiment, content completeness assessment is used to evaluate the extent to which the output content covers the standard answer. Its calculation formula is as follows:

[0096] ;

[0097] Where N represents the total number of segments in the standard answer; This represents the content of the i-th standard answer segment. represents the content of the j-th predicted segment; ROUGE-L represents the similarity calculation function based on the longest common subsequence.

[0098] Sequential consistency assessment is used to evaluate whether the order of output steps conforms to the logic of the standard teaching process. Its calculation formula is:

[0099] ;

[0100] Where M represents the number of prediction stages; This indicates the position of the k-th prediction step in the standard answer; I is an indicator function, which is defined when the condition... The value is 1 if the condition is true, and 0 otherwise. Denominator This represents the number of the most likely correct order pairs. This metric explicitly models and validates the sequence dependencies between instructional steps.

[0101] The accuracy assessment of teaching segment boundaries is used to evaluate the precision of the defined boundaries of teaching segments. The calculation formula is as follows:

[0102] ;

[0103] Where P0 represents the total number of predicted boundary points. This indicates the character position of the p-th predicted boundary point; Character positions representing standard boundary points; functions This formula is used to calculate the character distance between two boundary points. The formula uses a maximum tolerance window of 50 characters to accommodate reasonable deviations in boundary annotations through a progressive scoring method.

[0104] The name standardization assessment evaluates the degree of professionalism and semantic matching between the generated process names and standard names. The calculation formula is as follows:

[0105]

[0106] Where P represents the number of predicted stage names; N0 represents the number of stage names in the standard answer; This represents the name of the p-th prediction stage. This is a fuzzy string matching function, and its value range is... In the formula It is a quantitative balancing factor used to prevent inflated indicators due to excessive segmentation. This indicator effectively overcomes the assessment challenges arising from differences in naming habits among different teachers.

[0107] Further optimization in this embodiment of the invention involves constructing a structured knowledge base of teaching processes to provide the model with professional and traceable domain knowledge references; employing a two-stage retrieval strategy of 'coarse vector retrieval + fine sorting by cross-encoder' to balance retrieval efficiency with the accuracy of deep semantic matching; and finally, by injecting enhanced hints based on prior knowledge of the teaching process, guiding the lightweight large language model to complete process identification and sequence recovery. This effectively solves the core problems of blurred process boundaries and difficulty in capturing logical order in long teaching texts, significantly improving the accuracy, consistency, and professionalism of the segmentation results.

[0108] Example 2:

[0109] like Figure 2 As shown, this embodiment provides a retrieval-enhanced teaching segment implicit structure intelligent recognition system, including: a structured teaching segment knowledge base construction module, used to extract fine-grained [teaching segment name, content] fragment pairs from high-quality lesson plan texts, construct a structured knowledge base, and use the BGE embedding model to generate vector representations for each content fragment, and finally construct a vectorized index that supports efficient retrieval, providing a professional and traceable source of domain knowledge for subsequent retrieval stages.

[0110] The first-round retrieval module receives unordered teaching text input, converts it into query vectors using the BGE model, calculates the cosine similarity between the query vector and all fragment vectors in the knowledge base, and returns the top results based on the similarity score. The most relevant candidate fragment set is used to complete a coarse search based on semantic vector similarity, which effectively overcomes the shortcomings of small LLMs in terms of insufficient understanding in professional teaching fields.

[0111] The secondary precise re-ranking module is used to perform deep semantic matching and fine-ranking on the candidate fragment set obtained in the first round of retrieval, using a cross-encoder model. This module concatenates the query with each candidate fragment and inputs it into the model to obtain a precise relevance score. Based on this score, the candidate set is re-ranked, selecting the most relevant fragments. This significantly improves the quality and relevance of candidate fragments.

[0112] The lightweight large language model enhancement generation module is used to drive the lightweight large language model to perform two core tasks based on high-quality segments after fine sorting and pre-set structured enhancement prompts containing teaching process rules: accurately identifying the teaching segment name corresponding to each segment, and restoring the correct order of all teaching segments in accordance with teaching logic, and finally outputting a structured and ordered sequence of teaching segments.

[0113] The evaluation module is used to perform multi-dimensional quantitative evaluation of the teaching segmentation and sequence recovery results output by the large language model enhancement generation module. This module receives the final output of the generation module. In addition to standard answer data, the system comprehensively evaluates the results using four built-in automated assessment metrics—content completeness, sequence consistency, accuracy of process boundaries, and naming standardization—and outputs quantitative scoring reports, providing data support for the objective measurement and continuous optimization of model performance.

[0114] Furthermore, by Figure 3 As shown, the working process of the knowledge base construction module in structured teaching includes:

[0115] First, the collected high-quality lesson plan texts were preprocessed, uniformly converted to TXT format, and the core text content after the "teaching process" section was extracted. Then, based on the inherent logical units of the teaching activities, each lesson plan was manually or semi-automatically segmented and labeled into multiple pairs of fragments [teaching segment name, corresponding content], such as ["Introduction to the new lesson," "Students, today we will learn..."], ["Knowledge explanation," "First, let's look at Theorem 1..."], etc., thus constructing a structured knowledge base KB. Next, the BGE embedding model was used to analyze each content fragment c in the knowledge base. i Vectorize it to generate its corresponding dense vector representation e ci Finally, the vector e of all content fragments is... ci Collect and construct a vector index The index is organized using an efficient similarity search algorithm, providing fast vector query capabilities for subsequent first-round retrieval.

[0116] Furthermore, by Figure 4 As shown, the working process of the first-round retrieval module includes:

[0117] The system receives unordered, stage-name-hidden instructional text D as the query q from user input. First, it calls the BGE embedding model to convert the query text q into a query vector e. q Then, the module accesses the vector index. Calculate the query vector e q With each fragment vector e in the index ci cosine similarity After calculation, the system sorts all candidate segments in descending order based on their similarity scores. Finally, the module uses preset hyperparameters... (Usually set to 10), extract the top k most relevant segments from the sorted list to form the candidate segment set C for the first round of retrieval, and output it to the next module.

[0118] Furthermore, by Figure 5 As shown, the working process of the secondary precise reordering module includes:

[0119] Receive the set of candidate fragments from the first round of retrieval. For each candidate fragment in the set This is then concatenated with the original query q to form a complete text pair. This text pair is then input into a pre-trained cross-encoder model. The cross-encoder, through its deep attention mechanism, performs global, fine-grained semantic interaction computation on all lexical units between the query and candidate segments, ultimately outputting an accurate relevance score. iAfter all candidate segments have had their scores calculated, this module will then use the scores to... i Re-sort set C in descending order to obtain the finely sorted set. Finally, based on another parameter (Usually set to 5 or 6), from Selected from These fragments serve as high-quality, highly relevant reference content that is ultimately sent to the generation module.

[0120] Furthermore, by Figure 6 As shown, the working process of the lightweight large language model enhancement generation module includes:

[0121] First, the final fragment set after fine sorting... The process involves using a structured, precise extraction function. The names of the teaching segments and their core content in the segment are organized into a standardized intermediate representation. At the same time, the system will load a predefined enhanced hint. The prompt template clearly outlines common teaching flows (such as "Introduction → New Lesson → Consolidation → Summary → Homework") and includes clear instructions requiring the model to complete two sub-tasks: "Step Identification" and "Sequence Restoration." Subsequently, and Together, these elements form the contextual prompts, which are then input into the specified lightweight large language model. Based on its internal knowledge and logical reasoning capabilities, the LLM, guided by enhanced prompts, generates the final structured output. That is, an ordered sequence containing standardized step names and corresponding content. This allows for the identification and reconstruction of the implicit structure of the entire teaching text.

[0122] Furthermore, the evaluation module's workflow includes:

[0123] This module first receives the final output from the large language model enhancement generation module. and the corresponding standard answers Subsequently, the module executes the computational tasks of four evaluators in parallel: content completeness evaluation, which assesses the completeness of content coverage by calculating the ROUGE-L similarity between the predicted content and the standard answer content; sequence consistency evaluation, which assesses the logical rationality of the sequence by comparing the consistency between the predicted sequence and the standard sequence; sequence boundary accuracy evaluation, which assesses the accuracy of sequence division by calculating the character distance between the predicted boundary point and the standard boundary point (with a tolerance window of 50 characters); and name normalization evaluator, which assesses the semantic similarity between the predicted sequence name and the standard name using a fuzzy string matching algorithm. Finally, the module summarizes the scores from the four dimensions to generate a comprehensive evaluation report, completing an automated, multi-dimensional quantitative analysis of the overall system output quality.

[0124] During the experimental phase, to verify the effectiveness of this model, we conducted a comprehensive experimental evaluation. Regarding the dataset, we selected a batch of high-quality real-world lesson plan documents, converted their formats and extracted their content, and used a semi-automatic process of "large-scale initial screening + manual review" for annotation, constructing a structured knowledge base containing 236 segments [teaching segment name, content].

[0125] Furthermore, in terms of evaluation indicators, we designed a comprehensive evaluation system with four dimensions: content completeness, sequential consistency, accuracy of process boundaries, and standardization of names. We adopted fuzzy matching and tolerance mechanisms to adapt to the complexity of teaching scenarios.

[0126] Furthermore, the experiment was conducted on four NVIDIA RTX 4090 GPUs. We selected three lightweight models—DeepSeek-R1-Distill-Qwen-7B, Internlm2_5-7b-chat-1m, and Qwen3-4B—as the core test subjects, and used PIKE-RAG and SELF-RAG as strong baselines for comparison.

[0127] Furthermore, in the main experimental section, the results show that the lightweight model using the RAGIS framework significantly outperforms its version using only prompt words in all metrics. Moreover, the combination of Qwen3-4B+RAGIS even surpasses several mainstream models with larger parameter sets (DeepSeek-V3-0324, Moonshot-v1-128k) in overall performance, demonstrating the powerful enhancement capabilities of this framework.

[0128] Furthermore, in the ablation experiment, we verified that both the first-round retrieval and the second-round precise re-ranking modules significantly improved the final performance of the model and were indispensable, thus confirming the necessity of the two-stage retrieval architecture design.

[0129] Furthermore, in the parameter experiment section, we focused on exploring the impact of the knowledge base fragment content organization format (i.e., "stage name + first sentence" versus "stage name + complete paragraph") on performance. This is related to... Figure 2 The knowledge base module supports dynamic selection of retrieval methods, which complements the design. Experiments show that the Qwen3-4B model is more suitable for the "section name + complete paragraph" pattern (optimal parameters are...). , DeepSeek-R1-Distill-Qwen-7B performs better in the "section name + first sentence" mode (optimal parameters are...). , This indicates that it has a stronger ability to capture concise and key information. This proves that the model can flexibly adapt to the optimal knowledge retrieval and utilization strategy according to the characteristics of different backend LLMs.

[0130] This invention presents a teaching segment implicit structure recognition model based on multi-stage retrieval enhancement, aiming to address the core challenges of blurred boundaries and difficulty in capturing sequential dependencies in lengthy teaching texts. The model comprises four modules: a structured teaching segment knowledge base construction module, a first-round retrieval module, a secondary precise reordering module, and a lightweight large language model enhancement generation module. The structured teaching segment knowledge base construction module deconstructs domain knowledge into fine-grained fragments and constructs vector indexes, providing professional references for the model. The first-round retrieval module performs efficient coarse retrieval based on vector similarity, initially screening relevant candidates. The secondary precise reordering module utilizes a cross-encoder for deep semantic matching and fine reordering, ensuring high quality and high relevance of candidate fragments. The lightweight large language model enhancement generation module guides the model to accurately identify segments and restore logical order by injecting enhancement prompts based on prior knowledge of the teaching process. This multi-stage retrieval enhancement-based teaching segment implicit structure recognition model, through the combination of two-stage retrieval and enhancement generation, enables the model output to possess high accuracy, strong logic, and good interpretability, providing a reliable technical foundation for automated lesson plan analysis and intelligent teaching assistance.

[0131] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for intelligent recognition of implicit structures in teaching segments generated through retrieval enhancement, characterized in that, Includes the following steps: Construct a structured teaching segment knowledge base, which includes a set of teaching segment names and corresponding content segments, and generate an embedding vector for each content segment; Based on the input unordered teaching text, a query embedding vector is generated, and an initial set of candidate segments is retrieved from the knowledge base according to the similarity between the query embedding vector and the embedding vector. Based on the initial candidate fragment set, a cross-encoder model is used to calculate the relevance score between the query and each candidate fragment, and the initial candidate fragment set is reordered based on the relevance scores to obtain the final candidate fragment set. Based on the final candidate segment set and the preset enhancement prompts, a lightweight large language model is used to segment and restore the order of teaching segments, generating an ordered sequence of teaching segments.

2. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 1, characterized in that, The steps for constructing the knowledge base include: The lesson plan text is processed into fragment pairs of teaching segment names and corresponding content segments to construct an initial knowledge base; Use an embedding model to generate an embedding vector for each content fragment in the initial knowledge base; Based on the embedding vector, a vector index for each content fragment is constructed to form the final knowledge base.

3. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 2, characterized in that, The step of retrieving the initial candidate fragment set includes: Use an embedding model to convert unordered input instructional text into query embedding vectors; Calculate the cosine similarity between the query embedding vector and the embedding vectors of each content segment in the structured teaching knowledge base; An initial set of candidate segments is selected based on the cosine similarity.

4. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 3, characterized in that, The steps for reordering the initial candidate fragment set include: For each candidate fragment in the initial candidate fragment set, the query and candidate fragment are concatenated and input into the cross-encoder model to obtain a relevance score; The candidate segment set is sorted according to the relevance score to obtain the final candidate segment set; Select the top few segments from the final candidate segment set as the output.

5. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 4, characterized in that, The steps to generate a sequence of teaching activities include: The final candidate segment set is structured and extracted to generate intermediate output for teaching segmentation; The design includes enhanced prompts that contain common process templates and sequence restoration rules for teaching procedures; The intermediate outputs and enhanced prompts are input into a lightweight large language model to generate the final result of restoring the order of the teaching steps.

6. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 1, characterized in that, The method also includes an evaluation step for quantitatively evaluating the generated teaching segment segmentation and sequence restoration results. The evaluation dimensions include content completeness, sequence consistency, segment boundary accuracy, and name standardization.

7. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 6, characterized in that, The formula for calculating the completeness of the content is: ； Where N represents the total number of segments in the standard answer; This represents the content of the i-th standard answer segment. represents the content of the j-th predicted segment; ROUGE-L represents the similarity calculation function based on the longest common subsequence.

8. The method for intelligent recognition of implicit structure in teaching segments generated by retrieval enhancement according to claim 6, characterized in that, The formula for calculating the sequential consistency is: ； Where M represents the number of predicted segments; pos k This indicates the position of the k-th predicted segment in the standard answer. I is an indicator function that takes the value 1 when the condition is met, and 0 otherwise.

9. A retrieval-enhanced teaching process implicit structure intelligent recognition system, the system being used to implement the method described in any one of claims 1-8, characterized in that, include: The module consists of a construction module, a retrieval module, a sorting module, and a generation module. The construction module is used to build a structured teaching segment knowledge base, which includes a set of teaching segment names and corresponding content segments, and generates an embedding vector for each content segment. The retrieval module is used to generate a query embedding vector based on the input unordered teaching text, and retrieve an initial set of candidate segments from the knowledge base based on the similarity between the query embedding vector and the embedding vector. The sorting module is used to calculate the relevance score between the query and each candidate segment based on the initial candidate segment set using a cross-encoder model, and to re-sort the initial candidate segment set based on the relevance score to obtain the final candidate segment set. The generation module is used to segment and restore the order of teaching segments using a lightweight large language model based on the final candidate segment set and preset enhancement prompts, thereby generating an ordered sequence of teaching segments.

Images