RAG-based large language model learning path dynamic generation optimization method

Abstract

The invention discloses an RAG-based large language model learning path dynamic generation optimization method, and relates to the technical field of data processing, and the method comprises the steps: obtaining a predetermined learning path in a predetermined learning outline; obtaining an RAG model; performing information retrieval analysis on the target dynamic learning portrait to obtain a target real-time learning state, and performing correlation comparison analysis on the target real-time learning state and a predetermined knowledge base to obtain a candidate knowledge sequence; and performing knowledge fusion analysis on the target predetermined knowledge and the candidate knowledge sequence to obtain a target knowledge sequence, and dynamically adjusting the predetermined learning path in combination with the target period to generate a target dynamic learning path of the target student. The technical problems that in the prior art, intelligent dynamic adjustment of a follow-up learning path cannot be conducted according to the real-time learning condition of a student, so that the pertinence of the learning process of the student is not high, the learning quality and effect are not ideal, and the efficiency is low are solved, and the purposes of improving the individuation and flexibility of the learning path and improving the learning efficiency are achieved. And the overall learning effect of the student is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a RAG-based large language model learning path dynamic generation optimization method. Background Art

[0002] With the rapid development of artificial intelligence and online education technology, personalized learning and adaptive learning path generation have been widely used in the field of education. However, with the increasing complexity and scale of learning content, traditional learning path setting methods face many challenges. Traditional learning paths are often static and cannot be dynamically adjusted according to the real-time learning situation of students, resulting in low learning efficiency and failure to meet personalized learning needs. Summary of the invention

[0003] The present application provides a RAG-based large language model learning path dynamic generation optimization method, which is used to solve the technical problem that the existing technology is unable to perform intelligent dynamic adjustment of subsequent learning paths according to the real-time learning situation of the students, resulting in the students' learning process being not targeted, the learning quality and effect being unsatisfactory, and the efficiency being low.

[0004] In view of the above problems, the present application provides a RAG-based large language model learning path dynamic generation optimization method.

[0005] The present application provides a RAG-based large language model learning path dynamic generation optimization method, the method comprising: Acquire a predetermined learning path in a predetermined learning outline; acquire a RAG model, wherein the RAG model includes a multimodal retriever and an integrated fusion device; perform multimodal information retrieval analysis on the target dynamic learning portrait of the target learner through the multimodal retriever to obtain a target real-time learning state; perform a correlation comparison analysis between the target real-time learning state and a predetermined knowledge base to obtain a candidate knowledge sequence; perform a knowledge fusion analysis on the target predetermined knowledge of the target period in the predetermined learning path and the candidate knowledge sequence through the integrated fusion device to obtain a target knowledge sequence; dynamically adjust the predetermined learning path in combination with the target knowledge sequence and the target period to generate a target dynamic learning path for the target learner.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application obtains a predetermined learning path in a predetermined learning outline; obtains a RAG model, the RAG model includes a multimodal retriever and an integrated fusion device; performs multimodal information retrieval analysis on the target dynamic learning portrait of the target student through the multimodal retriever to obtain the target real-time learning status; performs a correlation comparison analysis on the target real-time learning status and the predetermined knowledge base to obtain a candidate knowledge sequence; performs a knowledge fusion analysis on the target predetermined knowledge of the target period in the predetermined learning path and the candidate knowledge sequence through the integrated fusion device to obtain a target knowledge sequence; dynamically adjusts the predetermined learning path in combination with the target knowledge sequence and the target period to generate a target dynamic learning path for the target student. The present invention solves the technical problem that the prior art cannot perform intelligent dynamic adjustment of subsequent learning paths according to the real-time learning situation of students, resulting in the lack of pertinence in the students' learning process, unsatisfactory learning quality and effect, and low efficiency. The present invention obtains a predetermined learning path and a RAG model, uses a multimodal retriever to analyze the dynamic learning portrait of the students, obtains the real-time learning status, and then compares the learning status with the predetermined knowledge base to generate a candidate knowledge sequence. The predetermined learning path and the candidate knowledge are fused through an integrated fuser to generate a target knowledge sequence, and the learning path is dynamically adjusted to finally generate a personalized dynamic learning path, thereby achieving the technical effect of improving the personalization and flexibility of the learning path and improving the overall learning effect of the students. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0008] Figure 1 A schematic diagram of a flow chart of a RAG-based large language model learning path dynamic generation optimization method provided in an embodiment of the present application; Figure 2 A schematic diagram of the process of obtaining a candidate knowledge sequence in the RAG-based large language model learning path dynamic generation optimization method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0009] The present application provides a RAG-based large language model learning path dynamic generation optimization method to solve the technical problem that the existing technology is unable to perform intelligent dynamic adjustment of subsequent learning paths according to the real-time learning situation of students, resulting in the lack of pertinence of the students' learning process, unsatisfactory learning quality and effect, and low efficiency. By obtaining a predetermined learning path and a RAG model, a multimodal retriever is used to analyze the dynamic learning portrait of the students, the real-time learning status is obtained, and then the learning status is compared with the predetermined knowledge base to generate a candidate knowledge sequence. The predetermined learning path and candidate knowledge are fused through an integrated fuser to generate a target knowledge sequence, and the learning path is dynamically adjusted to finally generate a personalized dynamic learning path, thereby achieving the technical effect of improving the personalization and flexibility of the learning path and improving the overall learning effect of the students.

[0010] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0011] It should be noted that any variations of the terms "include" and "have" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or inherent to these processes, methods, products or devices.

[0012] like Figure 1 As shown, the present application provides a RAG-based large language model learning path dynamic generation optimization method, the method comprising: Step S100: Obtain a predetermined learning path in a predetermined learning outline.

[0013] In the embodiment of the present application, a set learning outline, i.e., a predetermined learning outline, is first loaded from a preset database. The predetermined learning outline is planned in advance based on the student's learning objectives, course requirements, and knowledge areas, and contains all necessary learning content, modules, and knowledge points. The predetermined learning outline is then parsed to identify and extract the various learning modules, chapters, and knowledge points therein, and arranged in a logical order.

[0014] After parsing the planned learning outline, the planned learning path is extracted, which contains the learning content that students need to complete according to the planned plan, and is planned according to the order of learning, difficulty level and progress requirements.

[0015] Step S200: Acquire a RAG model, wherein the RAG model includes a multimodal retriever and an integrated fuser.

[0016] In an embodiment of the present application, a RAG model is first extracted from a predefined model library. The core of the RAG model includes a multimodal retriever and an integrated aggregator. Among them, RAG (Retrieval-Augmented Generation) technology is a technical method that combines retrieval and generation, which aims to enhance the ability of large language models (LLM) in processing knowledge-intensive tasks.

[0017] The main function of the multimodal retriever is to retrieve information from multiple data sources, covering data in different modalities such as text, images, audio and video. The multimodal retriever is used to perform cross-domain analysis in multimodal information and find knowledge points related to the student's current learning status. Through the multimodal retriever, the RAG model can not only perform single text information retrieval, but also extract useful information from video or audio.

[0018] The function of the integrated aggregator is to integrate and analyze the various information obtained by the multimodal retriever. The integrated aggregator deeply integrates the content from different modalities through semantic analysis and feature extraction to ensure that the generated knowledge content has logical consistency and coherence. For example, the integrated aggregator integrates different information on the same topic extracted from text and video into a complete knowledge sequence, making the generated learning path or knowledge recommendation more targeted and practical.

[0019] By acquiring and loading the RAG model and combining the functions of the multimodal retriever and integrated aggregator, intelligent dynamic learning path generation can be achieved, providing students with comprehensive and personalized learning support.

[0020] Step S300: Perform multimodal information retrieval and analysis on the target dynamic learning profile of the target student through the multimodal retriever to obtain the target real-time learning status.

[0021] In the embodiment of the present application, multi-dimensional data of the target student in the learning process is first collected, including the target student's historical learning records, completed courses, learning time, mastery of learning content, and test or quiz results. Based on the collected data, a target dynamic learning profile of the target student is constructed.

[0022] Next, a multimodal information retrieval analysis is performed on the target dynamic learning portrait of the target student through a multimodal retriever. First, the knowledge content that the target student has learned is identified, which includes the target learned text sequence and the target learned video sequence. Then, the text content of multiple lessons that have been learned is retrieved through the text retrieval channel of the multimodal retriever to extract the first target learned content corresponding to the lesson, such as course notes, book chapters and other text materials. At the same time, multiple course videos that the student has watched are retrieved through the non-text retrieval channel to extract the second target learned content related to these courses, such as important knowledge points and explanation content in the video. After completing the retrieval of text and video content, the first target learned content and the second target learned content are integrated and analyzed, that is, the text information and video information are integrated to generate the target real-time learning status of the target student.

[0023] Furthermore, the method provided in the application embodiment also includes: The multimodal retriever includes a text retrieval channel and a non-text retrieval channel; the target dynamic learning portrait is analyzed to obtain the target learned knowledge of the target student, wherein the target learned knowledge includes a target learned text sequence and a target learned video sequence; content retrieval is performed on multiple text segments corresponding to multiple lessons in the target learned text sequence through the text retrieval channel to obtain a first target learned content; content retrieval is performed on multiple videos corresponding to multiple courses in the target learned video sequence through the non-text retrieval channel to obtain a second target learned content; the first target learned content and the second target learned content are used as the real-time learning status of the target.

[0024] In the embodiment of the present application, the multimodal retriever is composed of a text retrieval channel and a non-text retrieval channel, which are used to comprehensively analyze the student's learning data. Specifically, the text retrieval channel is responsible for extracting and analyzing the target learned text sequence from the student's target dynamic learning portrait, and the non-text retrieval channel is used to analyze the target learned video sequence. The multimodal retriever constructs the student's target learned knowledge through the text retrieval channel and the non-text retrieval channel.

[0025] In the specific operation, the content of multiple lessons in the target learned text sequence is first retrieved through the text retrieval channel. Text retrieval uses natural language processing technology, such as TF-IDF, to perform vectorized analysis on the text content, extract the core knowledge points that the students have learned in multiple lessons, and generate the first target learned content.

[0026] At the same time, the content of multiple course videos in the target learned video sequence is retrieved through the non-text retrieval channel. By using video content analysis technology, such as automatic speech recognition, the audio part of the video is converted into text through automatic speech recognition technology, and then the converted text is analyzed using the same TF-IDF method as the text retrieval channel to extract the knowledge points learned by the students from the video, and generate the second target learned content.

[0027] Finally, the first target learned content obtained from the text retrieval channel is integrated with the second target learned content obtained from the non-text retrieval channel to obtain the target real-time learning status of the target learner.

[0028] Furthermore, the method provided in the application embodiment also includes: The non-text retrieval channel stores a format processing strategy; matching a first video corresponding to a first course among the multiple videos, the first course being any one of the multiple courses; parsing the first video according to the format processing strategy to obtain a first video parsing result, the first video parsing result including a first image sequence and a first audio; converting the first audio into a first text, and forming a formatted text with the first included text of the first image in the first image sequence; performing content retrieval on the formatted text to obtain the second target learned content.

[0029] In an embodiment of the present application, a format processing strategy is stored in the non-text retrieval channel, and its function is to convert the audio and image content in the video into a searchable text format.

[0030] Specifically, first, a keyword matching method is used to find the first video corresponding to the course that the student is studying from multiple course videos. When performing keyword matching, the metadata of the video, such as the course title or course number, is retrieved and compared to find the video that best matches the student's learning content. The first course is any one of the multiple courses that the student is studying. After matching, the first video is parsed in terms of audio and image timing according to the format processing strategy. In audio parsing, automatic speech recognition (ASR) technology is used to convert the audio part of the video into text. ASR technology directly converts the voice into text by analyzing the voice content in the audio, and generates a first text corresponding to the audio. This part of the text reflects the language content described in the video. At the same time, the image part of the video is parsed, and OCR technology, such as Tesseract, is used to analyze the video frame by frame to extract the text content in the image frame. Through this process, the text information displayed in the video frame is extracted to generate a first image timing corresponding to the image timing, which records the text extracted from each frame of the image and its time information. After generating the first text and the first image sequence, the audio text and the image text are integrated through the timestamp alignment method. According to the timestamp of the video frame, the audio content and the text information in the image frame are accurately matched in time and integrated into a structured format text.

[0031] Finally, the generated formatted text is retrieved, and the TF-IDF method is used to perform word segmentation and word frequency statistics on the content in the formatted text, and extract keywords related to the student's learning progress. This step identifies the knowledge points that the student has learned through the video, and finally generates the second target learned content, that is, the learning results obtained by the student by watching the video.

[0032] Step S400: performing correlation comparison analysis on the target real-time learning state and a predetermined knowledge base to obtain a candidate knowledge sequence.

[0033] In the embodiment of the present application, the predetermined knowledge base is a knowledge system covering multiple fields, including multiple knowledge points and conceptual structures, and these fields include different disciplines and topics.

[0034] When generating candidate knowledge sequences for students, a correlation comparison analysis is performed to compare the student's target real-time learning status with the knowledge points in the predetermined knowledge base to determine which contents in the knowledge base are related to the knowledge currently mastered by the student. Specifically, the student's learning status and the knowledge points in the knowledge base are first converted into vector form, and then the similarity between the student's learning status and each knowledge point in the knowledge base is calculated using the cosine similarity method. Based on the results of this correlation analysis, the knowledge points most relevant to the student's current learning are screened out from the predetermined knowledge base to generate a candidate knowledge sequence.

[0035] Further, such as Figure 2 As shown, the method provided by the application embodiment also includes: Extract the first predetermined text of the first predetermined knowledge in the predetermined knowledge base, and obtain the first predetermined word vector of the first predetermined text; obtain the target real-time word vector of the target real-time learning state; when the first cosine similarity between the first predetermined word vector and the target real-time word vector reaches a first similarity threshold, add the first predetermined knowledge to the candidate knowledge sequence; wherein the calculation expression of the first cosine similarity between the first predetermined word vector and the target real-time word vector is as follows: ; refers to the first predetermined word vector With the target real-time word vector The first cosine similarity, refers to the first predetermined word vector With the target real-time word vector The dot product of and They refer to the first predetermined word vectors The Euclidean norm of the target real-time word vector The Euclidean norm of , and They refer to the first predetermined word vectors With the target real-time word vector The elements, refers to the first predetermined word vector With the target real-time word vector Dimension.

[0036] In an embodiment of the present application, a first predetermined knowledge is first randomly extracted from a predetermined knowledge base by a random selection method, and a first predetermined text related to the knowledge point is obtained, wherein the first predetermined text is a specific content description of the first predetermined knowledge. Then, a word vector model such as Word2Vec is used to process the first predetermined text to generate a first predetermined word vector related to the first predetermined knowledge. At the same time, the target real-time learning status of the trainee is obtained, and the target real-time learning status represents the knowledge currently mastered by the trainee. For comparative analysis, the same word vector model is used to convert text data in the trainee's learning status, such as study notes, completed course content, etc., into vectors to generate a target real-time word vector.

[0037] Then, the similarity between the first predetermined word vector and the target real-time word vector is calculated by cosine similarity. The specific calculation process is as follows: ; in, refers to the first predetermined word vector With the target real-time word vector The first cosine similarity, refers to the first predetermined word vector With the target real-time word vector The dot product of and They refer to the first predetermined word vectors The Euclidean norm of the target real-time word vector The Euclidean norm of , and They refer to the first predetermined word vectors With the target real-time word vector The elements, refers to the first predetermined word vector With the target real-time word vector If the calculated cosine similarity reaches the preset first similarity threshold, that is, the preset similarity threshold, it indicates that the predetermined knowledge point has a high correlation with the student's learning status, and the first predetermined knowledge is added to the candidate knowledge sequence.

[0038] Repeat the above process, continuously randomly extract other knowledge points from the predetermined knowledge base, and calculate their similarity with the student's learning status in turn. When all knowledge points have been calculated for similarity, the traversal of the predetermined knowledge base is completed, and finally a candidate knowledge sequence containing multiple related knowledge points is generated.

[0039] Furthermore, the method provided in the application embodiment also includes: The target real-time learning state is divided to obtain N learned contents of N periods, wherein the N periods correspond to N weight coefficients, and 2≤N<4, and N is an integer; a first period among the N periods is extracted, and a first learned content of the first period is matched among the N learned contents, wherein the first learned content includes multiple contents of multiple nodes; the multiple nodes are randomly collected to obtain a first node set, and a first content set corresponding to the first node set is matched among the multiple contents; a fusion analysis is performed on the first content set to obtain a first content vector, and a first target content vector is obtained by weighting the first weight coefficient; and the first target content vector is analyzed to obtain the target real-time word vector.

[0040] In an embodiment of the present application, the time series division technology is first used to divide the student's target real-time learning state, and the entire learning process is divided into N cycles, 2≤N<4. Each cycle represents the knowledge content mastered by the student in different time periods. For each cycle, a preset weight coefficient is assigned to it. These weight coefficients are pre-set and reflect the degree of influence of each cycle on the current learning state. Usually, the earlier learning stages have less impact, so the weight coefficient is lower; and the most recent learning stage has a greater impact on the current learning state, so the weight coefficient is higher. For example, the weight is set to a coefficient of 0.2 for the first cycle, a coefficient of 0.3 for the second cycle, and a weight of 0.5 for the most recent cycle.

[0041] Next, the first cycle is extracted from the divided N cycles, which represents the knowledge mastered by the students in the earliest learning stage. The first learned content related to the first cycle is extracted through the time-stamped data extraction method. These contents are composed of multiple nodes, each of which represents the unit learned by the students in the course or class. Then these nodes are randomly collected, and some nodes are randomly selected from multiple classes or courses through random sampling technology to form a first node set. Based on the node labels, such as course numbers or class names, the first content sets related to these nodes are matched, indicating the specific content learned by the students at these nodes.

[0042] After obtaining the first content set, use a word vector model such as Word2Vec to convert these contents into vector representations for fusion analysis. By mapping the text content of each node into a vector, a comprehensive vector representing the learner in the first learning stage is generated, namely the first content vector. Next, the first content vector is weighted according to the preset weight coefficient. Since the impact of the first cycle is small, its weight coefficient is set to 0.2. The weighted result of the cycle is calculated by the weighted summation method, namely the first target content vector.

[0043] Repeat this process and weight the content vector of each cycle. Assume that the weight coefficient of the second cycle is 0.3 and the weight of the most recent cycle is 0.5. The weighted results of each cycle will generate the second target content vector and the third target content vector in turn. Finally, the weighted results of all cycles are summarized and the target real-time word vector of the student is generated through weighted combination.

[0044] Step S500: performing knowledge fusion analysis on the target predetermined knowledge of the target cycle in the predetermined learning path and the candidate knowledge sequence through the integrated fuser to obtain a target knowledge sequence.

[0045] In an embodiment of the present application, when performing knowledge fusion analysis on the target predetermined knowledge of the target cycle in the predetermined learning path and the candidate knowledge sequence through an integrated fuser, the candidate knowledge sequence is first arranged in descending order by cosine similarity, and a candidate knowledge descending list is generated. Next, the first adjustment ratio is read, the target predetermined knowledge is adjusted, and some knowledge points are deleted to obtain a first adjustment result. Then, the first candidate knowledge is extracted from the candidate knowledge descending list and added to the adjustment result to generate a first knowledge fusion result. Next, the vector of the first knowledge fusion result is calculated by the word vector model, and the similarity is calculated with the target real-time word vector of the trainee to determine whether it has reached the second similarity threshold. If not, repeat the adjustment and fusion steps until the similarity requirement is met or the predetermined number of iterations is reached, and finally the fused target knowledge sequence is output.

[0046] Furthermore, the method provided in the application embodiment also includes: Step a: Arrange the candidate knowledge sequence in descending order of cosine similarity to obtain a candidate knowledge descending list; step b: Read the first adjustment ratio; step c: Delete and adjust the target predetermined knowledge based on the first adjustment ratio to obtain a first adjustment result; step d: Extract the first candidate knowledge in the candidate knowledge descending list based on the first adjustment ratio; step e: Add the first candidate knowledge to the first adjustment result to obtain a first knowledge fusion result; step f: Obtain the first fused word vector of the first knowledge fusion result; step g: Calculate the first fused cosine similarity between the first fused word vector and the target real-time word vector; step h: Determine whether the first fused cosine similarity reaches the second similarity threshold; step i: If not, repeat steps b to i until the predetermined number of iterations is reached, and output the knowledge fusion result at that time as the target knowledge sequence.

[0047] In an embodiment of the present application, the candidate knowledge sequence is first arranged in descending order of cosine similarity, and the candidate knowledge points are arranged from high to low according to the similarity to generate a descending list of candidate knowledge. Next, the preset first adjustment ratio is called, and the first adjustment ratio is used to determine the ratio of deletion and replacement of knowledge points in the predetermined learning path. The adjustment ratio is a built-in parameter used to find a balance between predetermined knowledge and candidate knowledge. For example, setting the adjustment ratio to 0.5 will delete 50% of the predetermined knowledge and replace it with candidate knowledge.

[0048] According to the first adjustment ratio read, the target predetermined knowledge is deleted. A certain proportion of knowledge points are deleted from the predetermined knowledge using a deletion algorithm based on the adjustment ratio to generate a first adjustment result. For example, if the adjustment ratio is 0.5, half of the predetermined knowledge points are deleted, and the remaining part is retained for subsequent fusion. Based on the same first adjustment ratio, the most relevant first candidate knowledge is extracted from the previously sorted candidate knowledge descending list. Next, the first candidate knowledge is added to the first adjustment result to obtain the first knowledge fusion result.

[0049] After the fusion is completed, a word vector model such as Word2Vec is used to convert the fused first knowledge fusion result into a numerical representation to generate a first fused word vector. Subsequently, the similarity between the first fused word vector and the target real-time word vector is calculated by cosine similarity to obtain the first fused cosine similarity. Next, it is determined whether the first fused cosine similarity reaches the second similarity threshold. The second similarity threshold is a pre-set similarity threshold used to determine whether the knowledge fusion result meets the learning requirements. If the calculated similarity reaches or exceeds this threshold, the fusion result is considered valid.

[0050] If the first fused cosine similarity does not reach the second similarity threshold, the aforementioned fusion process is repeated until a preset number of iterations is reached, and finally the knowledge fusion result at this time is used as the target knowledge sequence.

[0051] Step S600: Dynamically adjust the predetermined learning path in combination with the target knowledge sequence and the target period to generate a target dynamic learning path for the target student.

[0052] In an embodiment of the present application, the target knowledge sequence is first combined with the target cycle. The target knowledge sequence contains the knowledge points that the trainee currently needs to learn, and the target cycle represents the current stage of the trainee in the learning path. Next, the knowledge points in the predetermined learning path are matched with the knowledge points in the target knowledge sequence for similarity. Specifically, by comparing the semantic similarity of the two knowledge points, the similarity score between them is calculated. The matching degree of each knowledge point is quantified using cosine similarity. Based on the obtained similarity matching results, the order of the knowledge points in the predetermined learning path begins to be adjusted. According to the similarity, the knowledge points in the learning path are rearranged. Specifically, knowledge points with higher similarity are preferentially placed at the front of the learning path to ensure that the trainee learns the most relevant content first, and for those knowledge points with lower similarity or no longer relevant, they are moved back. After completing the order adjustment, the target dynamic learning path is generated by the direct path generation method.

[0053] Furthermore, the method provided in the application embodiment also includes: The predetermined learning outline also includes a predetermined learning acceptance and a predetermined learning goal; obtaining the target learning result of the target student based on the target dynamic learning path; judging whether the target learning result meets the predetermined learning acceptance; if so, issuing a quantitative evaluation instruction; based on the quantitative evaluation instruction, dynamically adjusting the target dynamic learning path in combination with the predetermined learning goal is evaluated.

[0054] In the embodiment of the present application, the predetermined learning outline also includes a predetermined learning acceptance and a predetermined learning goal. The predetermined learning acceptance refers to the minimum standard that the trainee must meet, which is measured by the trainee's academic performance, task completion or skill mastery level. The predetermined learning goal is the specific outcome that the trainee needs to achieve throughout the entire learning path.

[0055] The target learning results of the target students based on the target dynamic learning path are obtained through online tests, homework submissions or progress reports. These learning results reflect the specific achievements achieved by the students in the learning process, including the students' mastery of knowledge points, performance in completing tasks, etc. The target learning results of the target students are compared with the predetermined learning acceptance standards. If the students' learning results meet or exceed these acceptance standards, it means that the students have mastered the learning content and meet the predetermined learning acceptance. A quantitative evaluation instruction is issued through an automatic trigger mechanism. This instruction is based on the students' performance and collects the students' learning data, such as learning time, test scores, completion of learning tasks, etc. These data are converted into quantitative evaluation results through data analysis technology to evaluate the students' learning performance.

[0056] Finally, based on the quantitative evaluation results and in combination with the predetermined learning goals, the adjustment effect of the target dynamic learning path is evaluated. Analyze whether the trainees have achieved the predetermined learning experience goals, learning domain goals, and learning knowledge volume goals in the adjusted learning path. Specific evaluation methods include judging whether the trainees feel that their learning experience has improved in the adjusted learning path through learning feedback, such as satisfaction surveys or learning participation data. For example, trainees may report that the adjusted learning tasks are easier, less stressful, or more interesting. At the same time, analyze the trainees' learning time and frequency to observe whether the trainees have invested more time and energy in the adjusted path. If the trainees' learning experience becomes more positive, it is confirmed that the learning experience goal has been achieved. According to the trainees' test scores or task completion in different learning areas, it is evaluated whether the trainees have fully mastered the predetermined domain knowledge. For example, if the trainees performed poorly in a certain knowledge area in the learning path before the adjustment, after the adjustment, the learning tasks in this area are added to supplement it. The evaluation results show that the trainees' performance in this area has improved significantly, and the learning area goal will be confirmed to have been achieved. By analyzing the trainees' mastery of knowledge points before and after the adjustment, it is judged whether the trainees have mastered more knowledge. For example, compare the knowledge point coverage or learning task completion of the trainees in the test before and after the adjustment. If the evaluation results show that the amount of knowledge mastered by the trainees has increased from 60% before the adjustment to more than 80% after the adjustment, it will be confirmed that the learning knowledge target has been successfully achieved.

[0057] By comprehensively analyzing the achievement of the above three goals, the dynamic adjustment effect evaluation of the target dynamic learning path is completed.

[0058] Furthermore, the method provided in the application embodiment also includes: The predetermined learning goals include learning experience goals, learning field goals, and learning knowledge amount goals.

[0059] In the embodiment of the present application, the predetermined learning goals include learning experience goals, learning domain goals and learning knowledge amount goals. Among them, the learning experience goal refers to the student's subjective feelings about the learning process, and evaluates the student's emotional experience such as interest and pressure in learning. The learning domain goal refers to the knowledge field involved in the student's learning process, ensuring that the student can cover the predetermined subject or knowledge category. The learning knowledge amount goal refers to the total amount of knowledge that the student needs to master in the learning process, which is used to measure the number of knowledge points acquired by the student in the entire learning path.

[0060] In the embodiments of the present application, in summary, the embodiments of the present application have at least the following technical effects: This application obtains a predetermined learning path in a predetermined learning outline; obtains a RAG model, the RAG model includes a multimodal retriever and an integrated fusion device; performs multimodal information retrieval analysis on the target dynamic learning portrait of the target student through the multimodal retriever to obtain the target real-time learning status; performs a correlation comparison analysis on the target real-time learning status and the predetermined knowledge base to obtain a candidate knowledge sequence; performs a knowledge fusion analysis on the target predetermined knowledge of the target period in the predetermined learning path and the candidate knowledge sequence through the integrated fusion device to obtain a target knowledge sequence; dynamically adjusts the predetermined learning path in combination with the target knowledge sequence and the target period to generate a target dynamic learning path for the target student. The present invention solves the technical problem that the prior art cannot perform intelligent dynamic adjustment of subsequent learning paths according to the real-time learning situation of students, resulting in the lack of pertinence in the students' learning process, unsatisfactory learning quality and effect, and low efficiency. The present invention obtains a predetermined learning path and a RAG model, uses a multimodal retriever to analyze the dynamic learning portrait of the students, obtains the real-time learning status, and then compares the learning status with the predetermined knowledge base to generate a candidate knowledge sequence. The predetermined learning path and the candidate knowledge are fused through an integrated fuser to generate a target knowledge sequence, and the learning path is dynamically adjusted to finally generate a personalized dynamic learning path, thereby achieving the technical effect of improving the personalization and flexibility of the learning path and improving the overall learning effect of the students.

[0061] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. The processes depicted in the accompanying drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0062] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0063] This specification and drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A RAG-based large language model learning path dynamic generation optimization method, characterized in that: include: Get the scheduled learning path in the scheduled learning outline; Acquire a RAG model, wherein the RAG model includes a multimodal retriever and an integrated fusion device; The multimodal retriever performs multimodal information retrieval analysis on the target dynamic learning profile of the target student to obtain the target real-time learning status; Performing correlation comparison analysis on the target real-time learning state and the predetermined knowledge base to obtain a candidate knowledge sequence; Performing knowledge fusion analysis on the target predetermined knowledge of the target cycle in the predetermined learning path and the candidate knowledge sequence by the integrated fusion device to obtain a target knowledge sequence; The predetermined learning path is dynamically adjusted in combination with the target knowledge sequence and the target period to generate a target dynamic learning path for the target student.

2. According to the RAG-based large language model learning path dynamic generation optimization method according to claim 1, it is characterized in that: include: The multimodal retriever includes a text retrieval channel and a non-text retrieval channel; Analyzing the target dynamic learning portrait to obtain the target learned knowledge of the target student, wherein the target learned knowledge includes a target learned text sequence and a target learned video sequence; Performing content retrieval on multiple text segments corresponding to multiple lessons in the target learned text sequence through the text retrieval channel to obtain a first target learned content; Performing content retrieval on a plurality of videos corresponding to a plurality of courses in the target learned video sequence through the non-text retrieval channel to obtain a second target learned content; The first target learned content and the second target learned content are used as the target real-time learning status.

3. According to the RAG-based large language model learning path dynamic generation optimization method according to claim 2, it is characterized in that: include: The non-text retrieval channel stores a format processing strategy; Matching a first video corresponding to a first course among the multiple videos, where the first course is any one of the multiple courses; Parsing the first video according to the format processing strategy to obtain a first video parsing result, wherein the first video parsing result includes a first image timing and a first audio; Convert the first audio into a first text, and form a formatted text with the first included text of the first image in the first image sequence; Performing content retrieval on the format text to obtain the second target learned content.

4. According to claim 1, the RAG-based large language model learning path dynamic generation optimization method is characterized in that: include: Extracting a first predetermined text of the first predetermined knowledge in the predetermined knowledge base, and obtaining a first predetermined word vector of the first predetermined text; Obtain a target real-time word vector of the target real-time learning state; When the first cosine similarity between the first predetermined word vector and the target real-time word vector reaches a first similarity threshold, adding the first predetermined knowledge to the candidate knowledge sequence; The calculation expression of the first cosine similarity between the first predetermined word vector and the target real-time word vector is as follows: ； refers to the first predetermined word vector With the target real-time word vector The first cosine similarity, refers to the first predetermined word vector With the target real-time word vector The dot product of and They refer to the first predetermined word vectors The Euclidean norm of the target real-time word vector The Euclidean norm of , and They refer to the first predetermined word vectors With the target real-time word vector The elements, refers to the first predetermined word vector With the target real-time word vector Dimension.

5. According to claim 4, the RAG-based large language model learning path dynamic generation optimization method is characterized in that: include: Dividing the target real-time learning state to obtain N learned contents of N periods, wherein the N periods correspond to N weight coefficients, and 2≤N<4, where N is an integer; Extracting a first cycle from the N cycles, and matching first learned content of the first cycle in the N learned content, where the first learned content includes multiple contents of multiple nodes; Randomly collect the multiple nodes to obtain a first node set, and match a first content set corresponding to the first node set in the multiple contents; Performing fusion analysis on the first content set to obtain a first content vector, and weighting it with a first weight coefficient to obtain a first target content vector; The first target content vector is analyzed to obtain the target real-time word vector.

6. According to claim 1, the RAG-based large language model learning path dynamic generation optimization method is characterized in that: include: Step a: Arrange the candidate knowledge sequence in descending order of cosine similarity to obtain a candidate knowledge descending list; Step b: reading the first adjustment ratio; Step c: deleting and adjusting the target predetermined knowledge based on the first adjustment ratio to obtain a first adjustment result; Step d: extracting the first candidate knowledge in the candidate knowledge descending list based on the first adjustment ratio; Step e: adding the first candidate knowledge to the first adjustment result to obtain a first knowledge fusion result; Step f: obtaining a first fused word vector of the first knowledge fusion result; Step g: Calculate the first fused cosine similarity between the first fused word vector and the target real-time word vector; Step h: determining whether the first fusion cosine similarity reaches a second similarity threshold; Step i: If not reached, repeat steps b to i until the predetermined number of iterations is reached, and output the knowledge fusion result at that time as the target knowledge sequence.

7. According to claim 1, the RAG-based large language model learning path dynamic generation optimization method is characterized in that: Also includes: The predetermined learning outline also includes predetermined learning acceptance and predetermined learning objectives; Obtaining a target learning result of the target learner based on the target dynamic learning path; Determining whether the target learning result meets the predetermined learning acceptance; If it meets the requirements, a quantitative assessment instruction will be issued; Based on the quantitative evaluation instruction and in combination with the predetermined learning goal, the target dynamic learning path is dynamically adjusted and evaluated for its effect.

8. According to claim 7, the RAG-based large language model learning path dynamic generation optimization method is characterized in that: The predetermined learning goals include learning experience goals, learning field goals, and learning knowledge amount goals.

Images