A pre-school education teaching content generation system based on a large language model

Abstract

The application discloses a pre-school education teaching content generation system based on a large language model, comprising: a teaching demand analysis module for extracting demand fields and forming structured teaching demand data; a teaching link arrangement module for dividing teaching targets and marking target types to form a teaching link sequence; a RetNet target retention coding module for writing target sub-items and teaching links to obtain a retention vector; a target coverage degree marking module for forming a target coverage degree table; a target coverage residual retention module for marking un-covered targets and forming a residual retention state; a link adaptation retention module for determining target supplementary links to form a target closed link state sequence; and a large language model constraint generation module for constraining the large language model to generate teaching content. The application improves the target coverage rate and link consistency of pre-school teaching content generation, and reduces the cost of teacher modification.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education technology, and in particular to a preschool education teaching content generation system based on a large language model. Background Technology

[0002] With the development of AI-powered educational technology, pre-trained large language models have been used for lesson plan writing, classroom activity design, teaching question and answer generation, and educational text organization. Current methods for generating preschool education content largely rely on teachers inputting the theme, age group, and teaching objectives, after which the pre-trained large language model directly outputs activity plans, questions, summaries, evaluations, and parent-teacher collaboration content. While this approach improves text generation efficiency, the generation process primarily depends on input prompts and lacks continuous and structured control over the implementation of teaching objectives across different teaching stages.

[0003] In existing technologies, some solutions check whether teaching content meets the age group and teaching objectives through rule templates, keyword verification, or post-generation review. This type of processing often occurs after text generation, making it difficult to record the coverage relationship between target sub-items and activity introductions, exploration tasks, game interactions, teacher questioning, children's responses, summary evaluations, and home-school collaboration within the generation process. When a preceding teaching stage has already covered some objectives, but subsequent teaching stages still need to supplement the uncovered objectives, existing methods lack a structure to continuously maintain and purposefully allocate the uncovered objectives to appropriate teaching stages.

[0004] When generating preschool education content using existing large language models, problems easily arise such as a disconnect between teaching objectives and teaching steps, some objectives not being addressed by activities or questions, and summaries, evaluations, and parent-teacher collaboration content deviating from the objectives. Current technologies lack the technical means to incorporate objective sub-items into the retention network state, generate residual retention states based on objective coverage, and use these residual retention states to constrain the output content of the large language model. Therefore, how to provide a preschool education content generation system based on a large language model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a preschool education teaching content generation system based on a large language model. This invention utilizes the RetNet persistence network and a pre-trained large language model to construct teaching objective retention, objective coverage residual retention, and process adaptation retention mechanisms, thereby achieving closed-loop generation of preschool education teaching content with the advantages of clear objective continuity, stable content structure, and controllable generation process.

[0006] A preschool education teaching content generation system based on a large language model according to an embodiment of the present invention includes:

[0007] The teaching needs parsing module is used to receive preschool teaching needs text, extract fields, and form structured teaching needs data.

[0008] The teaching segment arrangement module is used to read teaching objectives from structured teaching needs data, divide them into target sub-items, label the target types, and form a teaching segment sequence according to the preset teaching segment order;

[0009] The RetNet target preservation encoding module is used to compose the target sub-items, target types and teaching sequence into a target preservation input sequence, write it into the key vector and value vector of the RetNet preservation network, and output the target sub-item preservation vector and the sequence preservation vector.

[0010] The target coverage annotation module is used to calculate the similarity value between the target sub-item preservation vector and the link preservation vector to form a target coverage table;

[0011] The target coverage residual preservation module is used to accumulate the similarity value of the same target sub-item, mark the target sub-item below the preset coverage threshold as the uncovered target sub-item, and write it into the target coverage residual preservation status;

[0012] The segment adaptation and preservation module is used to select the mapping teaching segment with the lowest similarity value as the target supplement segment based on the target type and teaching segment mapping table, and write the target coverage residual preservation state into the preservation state of the target supplement segment to form a target closed teaching segment state sequence.

[0013] The large language model constraint generation module is used to write the state sequence of the target closed teaching link into the decoding control data, and control the pre-trained large language model to output preschool education teaching content.

[0014] Optionally, the operation of forming structured teaching requirements data includes:

[0015] Fields were extracted from the preschool teaching needs text to obtain structured teaching needs data containing teaching themes, age groups, teaching areas, teaching objectives, and class duration.

[0016] Optionally, the operation of breaking down the teaching objective into sub-objectives includes:

[0017] Read the teaching objective field from the structured teaching needs data, segment the teaching objective field into sentences, and obtain a set of target sentences;

[0018] The target statement set is deduplicated and numbered according to the order in which the target statements appear in the teaching objective field, resulting in target sub-items consisting of target sub-item numbers and target sub-item texts;

[0019] Label the target sub-items as one of the following target types: cognitive target sub-items, ability target sub-items, expression target sub-items, and affective target sub-items;

[0020] Write the target sub-items and target type into the target sub-item set, and send the target sub-item set to the RetNet target preservation encoding module and link adaptation preservation module.

[0021] Optionally, the operation of forming the teaching sequence includes:

[0022] The teaching steps are arranged in the following order: activity introduction, exploration tasks, interactive games, teacher questioning, children's responses, summary and evaluation, and home-school collaboration.

[0023] The cognitive objectives are written into the teaching segments corresponding to the activity introduction, exploration tasks, and teacher questions; the ability objectives are written into the teaching segments corresponding to the exploration tasks and game interactions; the expression objectives are written into the teaching segments corresponding to the teacher questions and children's responses; and the affective objectives are written into the teaching segments corresponding to the summary evaluation and home-school collaboration.

[0024] The teaching steps are arranged in the order of their positions to form a sequence.

[0025] Optionally, the operation of maintaining the target encoding includes:

[0026] Read the target sub-item number, target sub-item text, and target type from the target sub-item set; read the teaching segment position from the teaching segment sequence.

[0027] The target sub-item number, target sub-item text, target type, and teaching link position are matched one by one to form the target retention input unit;

[0028] Arrange the target retention input units according to the sequence of teaching steps to form a target retention input sequence;

[0029] Perform text embedding encoding on the target sub-item text in the input sequence to obtain the target sub-item encoding;

[0030] Perform category embedding encoding on the target type to obtain the target type encoding;

[0031] Position embedding encoding is performed on the location of each teaching segment to obtain the teaching segment code;

[0032] Write the target sub-item encoding into the key vector of the RetNet preserving network, and write the target type encoding and teaching segment encoding into the value vector of the RetNet preserving network;

[0033] The RetNet network reads the instruction code sequentially according to the instruction sequence, reads the key vector corresponding to the target sub-item code at each instruction position, and writes the value vector corresponding to the key vector into the hold state of that instruction position.

[0034] After writing the state of all teaching steps according to the teaching step sequence, output the target sub-item holding vector and the step holding vector.

[0035] Optionally, the operation of establishing the target coverage table includes:

[0036] Read target sub-items from the target sub-item set, read the teaching segment positions from the teaching segment sequence, and build a blank coverage table with target sub-items as rows and teaching segment positions as columns;

[0037] Read the target sub-item preservation vector and link preservation vector output by the RetNet target preservation encoding module;

[0038] Perform vector normalization on the target sub-item preservation vector and the link preservation vector to obtain the target sub-item normalized vector and the link normalized vector;

[0039] The similarity value between the target sub-item and the teaching segment is obtained by multiplying the normalized vector of the target sub-item and the normalized vector of the segment dimension by dimension and then summing them.

[0040] Write the similarity values ​​into the cells corresponding to the target sub-item and the corresponding teaching segment in the blank coverage table;

[0041] Fill all cells according to the correspondence between the target sub-items and the teaching links to form a target coverage table.

[0042] Optionally, the operation of preserving the target coverage residual includes:

[0043] Read the similarity values ​​row by row from the target coverage table according to the target sub-items, and add the similarity values ​​corresponding to the same target sub-item in the order of the teaching link sequence to obtain the target sub-item coverage cumulative value;

[0044] The cumulative coverage value of the target sub-item is compared with the preset coverage threshold. When the cumulative coverage value of the target sub-item is lower than the preset coverage threshold, the corresponding target sub-item is marked as an uncovered target sub-item.

[0045] Read the target item number, target item text, and target type corresponding to the uncovered target item from the target item set;

[0046] Read all similarity values ​​of the rows containing uncovered target sub-items from the target coverage table, and establish a correspondence between each similarity value and the corresponding teaching segment position in the teaching segment sequence;

[0047] Write the target sub-item number, target sub-item text, target type, target sub-item coverage cumulative value, teaching link location and corresponding similarity value of the uncovered target sub-item into the residual preservation record;

[0048] Write the target sub-item number and target sub-item text in the residual preservation record into the residual key vector of the RetNet preservation network, and write the target type, target sub-item coverage accumulation value, teaching link position and corresponding similarity value in the residual preservation record into the residual value vector of the RetNet preservation network.

[0049] RetNet keeps the network reading the residual key vector and residual value vector, transfers the uncovered target sub-items to the target coverage residual retention state, and sends the target coverage residual retention state to the link adaptation retention module.

[0050] Optionally, the operation of adapting and maintaining the execution phase includes:

[0051] Read the uncovered target sub-items, target types, cumulative coverage values ​​of target sub-items, teaching link locations, and corresponding similarity values ​​from the target coverage residual preservation status;

[0052] Read the mapping table between target type and teaching link. The mapping table records the corresponding activities, exploration tasks and teacher questions for the cognitive target sub-items, the exploration tasks and game interactions for the ability target sub-items, the teacher questions and children's responses for the expression target sub-items, and the summary evaluation and home-school cooperation for the affective target sub-items.

[0053] Based on the target type of the uncovered target sub-item, read the corresponding mapped teaching segment from the target type and teaching segment mapping table;

[0054] In the target coverage residual preservation state, read the similarity value between the uncovered target sub-item and each mapping teaching link, and arrange the mapping teaching links in order of similarity value from low to high;

[0055] The mapping teaching segment, which is ranked first, is identified as a supplementary objective segment.

[0056] When two mapping teaching segments have the same similarity value, the mapping teaching segment that is arranged later in the preset teaching segment order is selected as the target supplementary segment.

[0057] Write the uncovered target sub-items, target types, and target sub-item coverage accumulation values ​​from the target coverage residual retention state into the retention state of the target supplementation stage;

[0058] Write the target replenishment marker into the target replenishment state;

[0059] The state of each teaching step is maintained after being written according to the teaching step sequence, forming a target-closed teaching step state sequence, which is then sent to the large language model constraint generation module.

[0060] Optionally, the operation of generating execution constraints includes:

[0061] Read the state sequence of the closed teaching links of the target, and extract the teaching link name, teaching link maintenance state, target sub-item, target type and target supplementary mark item by item according to the order of the teaching link sequence;

[0062] The teaching segment name, target sub-item, target type, teaching segment hold status, and target supplementary mark are assembled into the corresponding teaching segment decoding control data in the order of segment name, target sub-item, target type, hold status, and supplementary mark;

[0063] The decoding control data of each teaching segment is concatenated into the model input text for the corresponding teaching segment;

[0064] The input text to the model is used as the current input to the pre-trained large language model to generate the corresponding teaching text.

[0065] When there are target supplementary markers in the input text of the model, supplementary content containing the target sub-items corresponding to the target supplementary markers will be generated in the corresponding teaching segment text;

[0066] The texts of each teaching segment are merged according to the order of the teaching segment sequence to form the preschool education teaching content.

[0067] Optionally, the preschool education teaching content includes activity introduction text segments, exploration task text segments, game interaction text segments, teacher question text segments, children's response text segments, summary and evaluation text segments, and home-school co-education text segments; the large language model constraint generation module writes the text of each teaching segment into the corresponding text segment according to the arrangement order of the teaching segment sequence, establishes the correspondence between text segment identifiers, teaching segment positions, target sub-items, and target supplementary markers, and forms preschool education teaching content with target correspondence.

[0068] The beneficial effects of this invention are:

[0069] (1) The present invention divides the teaching objectives into target sub-items and uses RetNet to preserve the encoding of target sub-items, target types and teaching sequence, so that the preschool education teaching content generation process can continuously retain the correspondence between teaching objectives and teaching links, and reduce the problem of teaching objectives being weakened in subsequent generation links.

[0070] (2) The present invention calculates the similarity value between the target sub-item and the teaching link through the target coverage table, and writes the target sub-item below the preset coverage threshold into the target coverage residual retention state, so that the uncovered target sub-item can be clearly recorded and continuously maintained, reducing the problems of unimplemented goals and disconnect between goals and activities in the teaching content.

[0071] (3) This invention determines the target supplementary link through the target type and teaching link mapping table, and writes the target coverage residual retention state into the retention state of the target supplementary link, so that the pre-trained large language model generates preschool education teaching content under the constraints of decoding control data, and improves the consistency of target acceptance between activity introduction, exploration task, teacher questioning, summary evaluation and home-school co-education. Attached Figure Description

[0072] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0073] Figure 1 This is a schematic diagram of the overall structure of a preschool education teaching content generation system based on a large language model proposed in this invention.

[0074] Figure 2 This is a schematic diagram of RetNet target-preserving encoding for a preschool education teaching content generation system based on a large language model proposed in this invention;

[0075] Figure 3 This is a schematic diagram illustrating the target residual adaptation and preservation of a preschool education teaching content generation system based on a large language model proposed in this invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0077] refer to Figures 1-3 A preschool education teaching content generation system based on a large language model includes:

[0078] The teaching needs parsing module is used to receive preschool teaching needs text, identify and separate the teaching theme field, age group field, teaching domain field, teaching objective field and class duration field from the preschool teaching needs text, and form structured teaching needs data;

[0079] The teaching process arrangement module is used to read the teaching objective field from the structured teaching needs data, divide the teaching objective field into target sub-items, label the target type, and form a teaching process sequence according to the preset teaching process order. The target sub-item is a single teaching objective unit obtained by dividing the teaching objective field, and the target type is used to characterize the target attribute of the target sub-item in the preschool education teaching content.

[0080] The RetNet target preservation encoding module is used to form a target preservation input sequence from the target sub-items, target types, and teaching sequence, and write it into the key vector and value vector of the RetNet preservation network. It outputs the target sub-item preservation vector and the sequence preservation vector. The RetNet preservation network is a sequence modeling network built based on the Retentive Network preservation mechanism. The RetNet preservation network preserves the correspondence between target sub-items and teaching sequences through key vectors, value vectors, and preservation states.

[0081] The target coverage annotation module is used to calculate the similarity value between the target sub-item preservation vector and the stage preservation vector to form a target coverage table. The target coverage table is used to record the coverage degree of each target sub-item in each teaching stage position.

[0082] The target coverage residual preservation module is used to accumulate the similarity value of the same target sub-item, mark the target sub-item below the preset coverage threshold as the uncovered target sub-item, and write it into the target coverage residual preservation state. The target coverage residual preservation state is the preservation state of the uncovered target sub-item, target type, coverage accumulation value and teaching link similarity value stored separately in the RetNet network.

[0083] The segment adaptation and preservation module is used to select the teaching segment with the lowest similarity value as the target supplement segment based on the target type and teaching segment mapping table. The target coverage residual preservation state is written into the preservation state of the target supplement segment to form a target closed teaching segment state sequence. The target supplement segment is the teaching segment position that matches the target type of the uncovered target sub-item and has the lowest coverage.

[0084] The large language model constraint generation module is used to write the state sequence of the closed teaching link into the decoding control data, and control the output of pre-trained large language model preschool education teaching content. The decoding control data is used to limit the teaching link name, target sub-item, target type, maintenance state and target supplementary label used by the pre-trained large language model when generating text for each teaching link.

[0085] In this embodiment, the operations for generating structured teaching requirement data include:

[0086] Fields were extracted from the preschool teaching needs text to obtain the teaching theme field, age group field, teaching domain field, teaching objective field, and class duration field. The teaching theme field consists of noun phrases representing the course theme, the age group field consists of words representing the kindergarten class or age range, the teaching domain field consists of the name of one of the domains: health, language, social, science, and art, the teaching objective field consists of statements representing the cognitive, operational, expressive, and emotional requirements of the children, and the class duration field consists of the class duration value and time unit.

[0087] Write the teaching topic field, age group field, teaching domain field, teaching objective field, and class duration field into the same structured teaching requirement record to form structured teaching requirement data. The structured teaching requirement record stores the teaching topic field, age group field, teaching domain field, teaching objective field, and class duration field according to the field name.

[0088] Send the teaching objective field to the teaching process arrangement module, and send the teaching topic field, age group field, teaching domain field, and class duration field to the large language model constraint generation module for splicing the model input text.

[0089] In this embodiment, the operation of dividing the teaching objective field into objective sub-items includes:

[0090] The teaching objective field is read from the structured teaching requirements data. The teaching objective field is segmented into sentences to obtain a set of target sentences. The sentences are segmented according to the period, semicolon, comma and newline character in the teaching objective field.

[0091] The target statement set is deduplicated and numbered according to the order in which the target statements appear in the teaching objective field, resulting in target sub-items consisting of target sub-item numbers and target sub-item text. Deduplication involves deleting duplicate statements with the same target sub-item text from the target statement set and retaining the first target statement that appears. Numbering involves assigning a unique number to each target sub-item according to the order in which the target statements appear.

[0092] The target sub-items are labeled as one of the following target types: cognitive target sub-items, ability target sub-items, expression target sub-items, and affective target sub-items. The target type labeling is based on the classification operation performed on the core verb phrases in the target sub-item text. Target sub-items containing "recognize," "understand," and "know" are labeled as cognitive target sub-items; target sub-items containing "observe," "compare," and "operate" are labeled as ability target sub-items; target sub-items containing "tell," "answer," and "express" are labeled as expression target sub-items; and target sub-items containing "willing to participate," "cooperate," and "form a habit" are labeled as affective target sub-items.

[0093] Write the target sub-items and target type into the target sub-item set, and send the target sub-item set to the RetNet target preservation encoding module and link adaptation preservation module. The target sub-item set includes the target sub-item number, target sub-item text and target type.

[0094] In this embodiment, the operations for forming a sequence of teaching steps include:

[0095] The teaching steps are arranged in the following order: activity introduction, exploration task, game interaction, teacher questioning, children's response, summary evaluation, and home-school cooperation. The teaching step position consists of a step number and a step name. The step number is used to indicate the position of the teaching step in the teaching step sequence.

[0096] The cognitive objectives are written into the teaching segments corresponding to the activity introduction, exploration tasks, and teacher questions; the ability objectives are written into the teaching segments corresponding to the exploration tasks and game interactions; the expression objectives are written into the teaching segments corresponding to the teacher questions and children's responses; and the affective objectives are written into the teaching segments corresponding to the summary evaluation and home-school collaboration. The writing operation is to create a storage record of the objective sub-item number, objective sub-item text, and objective type under the corresponding teaching segment position.

[0097] The teaching steps are arranged in the order of their positions to form a teaching step sequence. The teaching step sequence includes the activity introduction position, the exploration task position, the game interaction position, the teacher questioning position, the children's response position, the summary and evaluation position, and the home-school cooperation position. Each teaching step position stores at least one target sub-item number and the corresponding target type.

[0098] In this embodiment, the operation of performing target preservation encoding includes:

[0099] Read the target sub-item number, target sub-item text, and target type from the target sub-item set; read the teaching segment position from the teaching segment sequence.

[0100] The target sub-item number, target sub-item text, target type, and teaching link position are matched item by item to form a target retention input unit. The item-by-item matching means that each target sub-item is associated with each teaching link position in the teaching link sequence. Each associated record includes the target sub-item number, target sub-item text, target type, link number, and link name.

[0101] The target retention input units are arranged according to the sequence of teaching steps to form a target retention input sequence. The target retention input sequence is arranged with the position of the teaching step as the main order and the target sub-item number as the subordinate order.

[0102] Text embedding encoding is performed on the target sub-item text in the input sequence to obtain the target sub-item encoding. Text embedding encoding is an encoding operation that converts the target sub-item text into a fixed-dimensional numerical vector.

[0103] Perform category embedding encoding on the target type to obtain the target type encoding. Category embedding encoding is an encoding operation that converts cognitive target sub-items, ability target sub-items, expression target sub-items, and emotion target sub-items into fixed-dimensional numerical vectors.

[0104] The location embedding code is performed on the teaching segment location to obtain the teaching segment code. The location embedding code is an encoding operation that converts the segment number and segment name corresponding to activity introduction, exploration task, game interaction, teacher questioning, children's response, summary evaluation and home-school co-education into a fixed-dimensional numerical vector.

[0105] The target sub-item encoding is written into the key vector of the RetNet persistence network, and the target type encoding and teaching segment encoding are written into the value vector of the RetNet persistence network. The key vector is used to index the target sub-item, and the value vector is used to store the persistence content corresponding to the target type and the teaching segment position.

[0106] The RetNet persistence network reads the instruction code sequentially according to the instruction sequence. At each instruction position, it reads the key vector corresponding to the target sub-item code and writes the value vector corresponding to the key vector into the persistence state at that instruction position. The persistence state is the state storage unit where the RetNet persistence network saves the target type code and instruction code at the corresponding instruction position.

[0107] After writing the state of all teaching segments according to the teaching segment sequence, output the target sub-item holding vector and the segment holding vector. The target sub-item holding vector is obtained by aggregating the holding states of the same target sub-item at all teaching segment positions, and the segment holding vector is obtained by aggregating the holding states of all target sub-items at the same teaching segment position. The aggregation method is to calculate the average of each dimension.

[0108] In this embodiment, the operation of establishing the target coverage table includes:

[0109] Read target sub-items from the target sub-item set, read teaching segment positions from the teaching segment sequence, and build a blank coverage table with target sub-items as rows and teaching segment positions as columns. The row identifiers of the blank coverage table are target sub-item numbers, and the column identifiers are segment names of teaching segment positions.

[0110] Read the target sub-item preservation vector and link preservation vector output by the RetNet target preservation encoding module;

[0111] Perform vector normalization on the target sub-item preservation vector and the link preservation vector to obtain the target sub-item normalized vector and the link normalized vector. The vector normalization process is to divide the value of each dimension of the vector by the vector magnitude, and the vector magnitude is the square root of the sum of the squares of each dimension of the vector.

[0112] The similarity value between the target sub-item normalized vector and the link normalized vector is obtained by multiplying them dimension by dimension and summing them. The similarity value is used to represent the degree of coverage between the target sub-item and the corresponding teaching link position.

[0113] Write the similarity values ​​into the cells corresponding to the target sub-item and the corresponding teaching segment in the blank coverage table;

[0114] Fill all cells according to the correspondence between target sub-items and teaching segment locations to form a target coverage table. Each row in the target coverage table corresponds to a target sub-item, each column corresponds to a teaching segment location, and each cell stores a similarity value.

[0115] In this embodiment, the operation of preserving the target coverage residual includes:

[0116] Read the similarity values ​​row by row from the target coverage table according to the target sub-items, and add the similarity values ​​corresponding to the same target sub-item in the order of the teaching link sequence to obtain the target sub-item coverage cumulative value. The target sub-item coverage cumulative value is the sum of the coverage degree of the same target sub-item in all teaching link positions.

[0117] The cumulative coverage value of the target sub-item is compared with the preset coverage threshold. When the cumulative coverage value of the target sub-item is lower than the preset coverage threshold, the corresponding target sub-item is marked as an uncovered target sub-item. The preset coverage threshold is a fixed threshold parameter stored in the target coverage residual preservation module, which is used to determine whether the target sub-item has been covered by the teaching sequence. An uncovered target sub-item means that the cumulative coverage value of the target sub-item has not reached the preset coverage threshold.

[0118] Read the target item number, target item text, and target type corresponding to the uncovered target item from the target item set;

[0119] Read all similarity values ​​of the rows containing uncovered target sub-items from the target coverage table, and establish a correspondence between each similarity value and the corresponding teaching segment position in the teaching segment sequence. The correspondence includes the target sub-item number, the teaching segment position, and the similarity value.

[0120] Write the target sub-item number, target sub-item text, target type, target sub-item coverage cumulative value, teaching link location and corresponding similarity value of the uncovered target sub-items into the residual preservation record. The residual preservation record is a data record that saves the uncovered target sub-items and corresponding insufficient coverage information.

[0121] Write the target item number and target item text in the residual preservation record into the residual key vector of the RetNet preservation network. Write the target type, target item coverage accumulation value, teaching link position and corresponding similarity value in the residual preservation record into the residual value vector of the RetNet preservation network. The residual key vector is used to index the uncovered target items, and the residual value vector is used to store the target type, coverage accumulation value and teaching link similarity distribution of the uncovered target items.

[0122] The RetNet maintains the network by reading the residual key vector and residual value vector, transferring the uncovered target sub-items into the target coverage residual maintain state, and sending the target coverage residual maintain state to the link adaptation maintain module. The target coverage residual maintain state is stored separately from the maintain state in the RetNet maintain network. The target coverage residual maintain state is used to continuously maintain the uncovered target sub-items and the corresponding insufficient coverage information.

[0123] In this embodiment, the operations for maintaining adaptation during execution include:

[0124] Read the uncovered target sub-items, target types, cumulative coverage values ​​of target sub-items, teaching link locations, and corresponding similarity values ​​from the target coverage residual preservation status;

[0125] Read the target type and teaching segment mapping table. The target type and teaching segment mapping table records the corresponding activity introduction, exploration task and teacher questioning for cognitive target sub-items, the corresponding exploration task and game interaction for ability target sub-items, the corresponding teacher questioning and children's response for expression target sub-items, and the corresponding summary evaluation and home-school cooperation for affective target sub-items. The target type and teaching segment mapping table is used to limit the position of the teaching segment to which different target types can be supplemented.

[0126] According to the target type of the uncovered target sub-item, the corresponding mapped teaching segment is read from the target type and teaching segment mapping table. The mapped teaching segment is the position of the teaching segment in the target type and teaching segment mapping table that corresponds to the target type of the uncovered target sub-item.

[0127] In the target coverage residual preservation state, read the similarity value between the uncovered target sub-item and each mapping teaching link, and arrange the mapping teaching links in order of similarity value from low to high;

[0128] The mapping teaching segment ranked first is determined as the target supplementary segment. The target supplementary segment is the teaching segment position with the lowest similarity value that the target type can inherit.

[0129] When the similarity values ​​of two mapping teaching links are the same, the mapping teaching link that is arranged later in the preset teaching link order is selected as the target supplementary link. The teaching link that is arranged later in the preset teaching link order is used to carry over the target content that was not completed in the previous teaching link.

[0130] Write the uncovered target sub-items, target types, and target sub-item coverage accumulation values ​​from the target coverage residual retention status to the retention status of the target supplementation stage. The write operation is to add the target sub-item number, target sub-item text, target type, and target sub-item coverage accumulation value corresponding to the status record in the retention status of the target supplementation stage.

[0131] Write a target supplementation marker in the hold state of the target supplementation stage. The target supplementation marker is used to instruct the large language model constraint generation module to generate teaching content corresponding to the uncovered target sub-items in the target supplementation stage.

[0132] The state of each teaching step after it is written is maintained according to the teaching step sequence to form a target closed teaching step state sequence. The target closed teaching step state sequence is sent to the large language model constraint generation module. The target closed teaching step state sequence includes the maintained state of each teaching step position, the maintained state of the target supplementary steps, and the target supplementary marker.

[0133] In this embodiment, the operation of generating constraints includes:

[0134] Read the state sequence of the closed teaching links of the target, and extract the teaching link name, teaching link maintenance state, target sub-item, target type and target supplementary mark item by item according to the order of the teaching link sequence;

[0135] The teaching segment name, target sub-item, target type, teaching segment hold state, and target supplementary marker are assembled into the corresponding teaching segment decoding control data in the order of segment name, target sub-item, target type, hold state, and supplementary marker. The decoding control data is the structured input data used by the pre-trained large language model to generate the corresponding teaching segment text.

[0136] The decoding control data of each teaching segment is concatenated into the model input text of the corresponding teaching segment. The model input text consists of segment name text, target sub-item text, target type text, state preservation summary text, and target supplementary label text.

[0137] The input text to the model is used as the current input to the pre-trained large language model to generate the corresponding teaching text. The pre-trained large language model is a language generation model that has been trained on a large corpus and can generate natural language text based on the input text.

[0138] When there are target supplementary markers in the input text of the model, supplementary content containing the target sub-items corresponding to the target supplementary markers is generated in the corresponding teaching segment text. The supplementary content is consistent with the uncovered target sub-items corresponding to the target supplementary markers.

[0139] The texts of each teaching segment are merged according to the sequence of teaching segments to form preschool education teaching content. The merging order of the texts of each teaching segment is consistent with the order of activity introduction, exploration task, game interaction, teacher questioning, children's response, summary evaluation and home-school cooperation.

[0140] In this implementation, the preschool education teaching content includes activity introduction text segments, exploration task text segments, game interaction text segments, teacher question text segments, children's response text segments, summary and evaluation text segments, and home-school collaboration text segments. The large language model constraint generation module writes the text of each teaching segment into the corresponding text segment according to the arrangement order of the teaching segment sequence, establishes the correspondence between text segment identifiers, teaching segment positions, target sub-items, and target supplementary markers, and forms preschool education teaching content with target correspondence. The text segment identifiers are used to distinguish between activity introduction text segments, exploration task text segments, game interaction text segments, teacher question text segments, children's response text segments, summary and evaluation text segments, and home-school collaboration text segments. The target correspondence is used to record the target sub-items and target supplementary markers corresponding to each text segment.

[0141] Example 1: To verify the feasibility of this invention in practice, it was applied to a preschool education teaching content generation scenario. This scenario is designed for teachers' daily lesson preparation. After teachers input the course theme, age group, teaching domain, teaching objectives, and class duration, the system generates activity introductions, exploration tasks, interactive games, teacher questions, children's responses, summary evaluations, and home-school collaboration texts. Comparison methods included direct generation using a standard pre-trained large language model and generation using rule template validation. All three methods used the same teaching requirement texts, the same number of objectives, and the same evaluation criteria.

[0142] During implementation, the preschool teaching needs text first enters the teaching needs analysis module, which extracts the teaching theme field, age group field, teaching domain field, teaching objective field, and class duration field to form structured teaching needs data. The teaching process arrangement module segments the teaching objective field into sub-items, numbers and labels each sub-item as one of the following objective types: cognitive objective sub-item, ability objective sub-item, expression objective sub-item, and affective objective sub-item. Then, the sub-items are assigned to the corresponding teaching process positions in activity introduction, exploration task, game interaction, teacher questioning, child response, summary evaluation, and home-school collaboration.

[0143] The RetNet target retention encoding module constructs a target retention input sequence from the target sub-item number, target sub-item text, target type, and teaching stage position. The target sub-item text, target type, and teaching stage position are each embedded and encoded. The target sub-item encoding is fed into the key vector of the RetNet retention network, while the target type encoding and teaching stage encoding are fed into the value vector. The RetNet retention network updates the retention state according to the teaching stage sequence, outputting the target sub-item retention vector and the stage retention vector.

[0144] The target coverage annotation module normalizes the target sub-item preservation vector and the stage preservation vector, then multiplies them dimension by dimension and sums them to obtain the similarity value between the target sub-item and the teaching stage location, and writes it into the target coverage table. The target coverage residual preservation module accumulates the similarity values ​​for each target sub-item to obtain the target sub-item coverage accumulation value. When the target sub-item coverage accumulation value is lower than the preset coverage threshold, the corresponding target sub-item is marked as an uncovered target sub-item and written into the residual key vector and residual value vector, forming the target coverage residual preservation state.

[0145] The module for adapting and maintaining the learning process reads the target coverage residual retention status and the mapping table between target type and teaching process. Cognitive target sub-items correspond to activity introductions, exploration tasks, and teacher questions; ability target sub-items correspond to exploration tasks and game interactions; expression target sub-items correspond to teacher questions and children's responses; and affective target sub-items correspond to summary evaluation and home-school collaboration. The system selects the teaching process with the lowest similarity value from the mapped teaching processes as the target supplementary process, writes the target coverage residual retention status into the retention status of the target supplementary process, and adds a target supplementary marker.

[0146] The large language model constraint generation module extracts the teaching step name, teaching step hold state, target sub-item, target type, and target supplementary marker from the target closed teaching step state sequence, assembles them into decoding control data, and then concatenates them into the model input text. The pre-trained large language model generates the corresponding teaching step text based on the model input text. When the model input text contains target supplementary markers, the pre-trained large language model generates supplementary content containing uncovered target sub-items in the corresponding teaching step text. The teaching step texts are merged according to the teaching step sequence to obtain the preschool education teaching content.

[0147] This embodiment sets up 60 course themes, covering five teaching areas: health, language, social studies, science, and art. Each course theme contains 3-5 teaching objectives, with a total of 212 sub-objectives. Teacher evaluators review the generated content based on whether the sub-objectives are integrated into activities, questions, summaries, and home-school collaboration. The direct generation method using a conventional pre-trained large language model sometimes results in complete objectives but insufficient activity integration. While rule-template-based generation can reduce some omissions, there are still insufficient integrations between affective objectives and home-school collaboration. This invention maintains the target coverage residual state and the target supplementation stage state, ensuring that uncovered target sub-objectives are retained in the generation chain and enter the adaptation teaching stage. The comparison data table is shown in the figure below:

[0148] Table 1: Comparison of Teaching Objective Coverage Effectiveness

[0149] Number of course topics 60 60 60 Number of sub-items of teaching objectives 212 212 212 Average coverage of target sub-items 76.8% 84.5% 95.6% Average of uncovered target sub-items 1.84 1.12 0.32 Accuracy of target correspondence in each stage 72.5% 80.9% 94.1% Target supplementary phase hit rate none 68.3% 91.7% Average number of revisions per teacher 2.4 wheels 1.7 wheels 0.6 wheels Teachers adjust time manually 18.6 minutes 13.2 minutes 6.1 minutes Activity missing rate 13.8% 9.6% 2.9% Age group expression mismatch rate 11.2% 8.5% 3.1% Home-school co-education target implementation rate 61.4% 70.8% 90.5% Content adoption rate 63.0% 71.5% 88.6% Content revision rate 42.0% 28.5% 9.0% Number of times safety reminders were missed 17 times 12 times 4 times Target correspondence traceability rate 0% 48.3% 96.2%

[0150] The data in the table shows that, under the same 60 course themes, 212 teaching objective sub-items, and the same evaluation criteria, this invention outperforms both the conventional pre-trained large language model direct generation method and the rule template verification generation method in multiple evaluation indicators. The average coverage rate of objective sub-items generated by the conventional pre-trained large language model direct generation method is 76.8%, the rule template verification generation method is 84.5%, while this invention's system reaches 95.6%. This indicates that this invention can effectively integrate more teaching objective sub-items into teaching stages such as activity introduction, exploration tasks, game interaction, teacher questioning, children's responses, summary evaluation, and home-school collaboration. Regarding the average number of uncovered objective sub-items, the conventional pre-trained large language model direct generation method has 1.84, the rule template verification generation method has 1.12, while this invention's system reduces it to 0.32. This shows that the objective coverage residual maintenance state can continuously maintain uncovered objective sub-items, and the stage adaptation and maintenance module can supplement uncovered objective sub-items into the objective supplementation stage.

[0151] Regarding the correspondence between teaching segments and teaching objectives, the accuracy rate of segment-objective correspondence in the direct generation method of ordinary pre-trained large language models is 72.5%, and in the rule template verification generation method it is 80.9%, while the system of this invention improves it to 94.1%. In terms of the hit rate of objective supplementation segments, the direct generation method of ordinary pre-trained large language models lacks a target supplementation segment selection mechanism, and the rule template verification generation method has a hit rate of 68.3%, while the system of this invention reaches 91.7%, indicating that this invention, through the mapping table between target types and teaching segments and the sorting by similarity values, can more accurately select target supplementation segments. Regarding the home-school co-education objective acceptance rate, the direct generation method of ordinary pre-trained large language models has a hit rate of 61.4%, and in the rule template verification generation method it is 70.8%, while the system of this invention reaches 90.5%, indicating that emotional target sub-items and extended target content can more stably enter the summary evaluation and home-school co-education segments.

[0152] Regarding the cost of subsequent adjustments by teachers, the average number of revision rounds for the conventional pre-trained large language model direct generation method is 2.4 rounds, and for the rule template validation generation method it is 1.7 rounds, while the system of this invention reduces this to 0.6 rounds; the time spent on manual adjustments by teachers is reduced from 18.6 minutes and 13.2 minutes to 6.1 minutes. In terms of the revision rate of generated content, the conventional pre-trained large language model direct generation method has a revision rate of 42.0%, and the rule template validation generation method has a revision rate of 28.5%, while the system of this invention reduces this to 9.0%. Regarding the content adoption rate, the conventional pre-trained large language model direct generation method has a adoption rate of 63.0%, and the rule template validation generation method has a adoption rate of 71.5%, while the system of this invention increases this to 88.6%. These data demonstrate that this invention can reduce the number of secondary revisions required by teachers for the generated content and increase the proportion of generated content directly used for lesson preparation.

[0153] Regarding content completeness and applicability, the missing activity rate of the direct generation method using ordinary pre-trained large language models was 13.8%, the rate of the rule template verification generation method was 9.6%, and the system of this invention reduced it to 2.9%; the age group expression mismatch rate decreased from 11.2% and 8.5% to 3.1%; and the number of missed safety prompts decreased from 17 and 12 times to 4 times. In terms of the traceability rate of target correspondence, the direct generation method using ordinary pre-trained large language models had a rate of 0%, the rate of the rule template verification generation method was 48.3%, and the system of this invention reached 96.2%. These data indicate that this invention not only improves the coverage of teaching objectives but also makes the target continuation process in the generated content traceable through the correspondence between target sub-items, teaching link positions, target supplementary markers, and text segments, thereby improving the structural stability, target consistency, and ease of use for teachers in the generation of preschool education teaching content.

[0154] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A preschool education teaching content generation system based on a large language model, characterized in that, include: The teaching needs parsing module is used to receive preschool teaching needs text, extract fields, and form structured teaching needs data. The teaching segment arrangement module is used to read teaching objectives from structured teaching needs data, divide them into target sub-items, label the target types, and form a teaching segment sequence according to the preset teaching segment order; The RetNet target preservation encoding module is used to compose the target sub-items, target types and teaching sequence into a target preservation input sequence, write it into the key vector and value vector of the RetNet preservation network, and output the target sub-item preservation vector and the sequence preservation vector. The target coverage annotation module is used to calculate the similarity value between the target sub-item preservation vector and the link preservation vector to form a target coverage table; The target coverage residual preservation module is used to accumulate the similarity value of the same target sub-item, mark the target sub-item below the preset coverage threshold as the uncovered target sub-item, and write it into the target coverage residual preservation status; The segment adaptation and preservation module is used to select the mapping teaching segment with the lowest similarity value as the target supplement segment based on the target type and teaching segment mapping table, and write the target coverage residual preservation state into the preservation state of the target supplement segment to form a target closed teaching segment state sequence. The large language model constraint generation module is used to write the state sequence of the target closed teaching link into the decoding control data, and control the pre-trained large language model to output preschool education teaching content.

2. The preschool education teaching content generation system based on a large language model according to claim 1, characterized in that, The operations for generating structured teaching needs data include: Fields were extracted from the preschool teaching needs text to obtain structured teaching needs data containing teaching themes, age groups, teaching areas, teaching objectives, and class duration.

3. The preschool education teaching content generation system based on a large language model according to claim 2, characterized in that, The operation of breaking down teaching objectives into sub-objectives includes: Read the teaching objective field from the structured teaching needs data, segment the teaching objective field into sentences, and obtain a set of target sentences; The target statement set is deduplicated and numbered according to the order in which the target statements appear in the teaching objective field, resulting in target sub-items consisting of target sub-item numbers and target sub-item texts; Label the target sub-items as one of the following target types: cognitive target sub-items, ability target sub-items, expression target sub-items, and affective target sub-items; Write the target sub-items and target type into the target sub-item set, and send the target sub-item set to the RetNet target preservation encoding module and link adaptation preservation module.

4. The preschool education teaching content generation system based on a large language model according to claim 3, characterized in that, The operations for forming the sequence of teaching steps include: The teaching steps are arranged in the following order: activity introduction, exploration tasks, interactive games, teacher questioning, children's responses, summary and evaluation, and home-school collaboration. The cognitive objectives are written into the teaching segments corresponding to the activity introduction, exploration tasks, and teacher questions; the ability objectives are written into the teaching segments corresponding to the exploration tasks and game interactions; the expression objectives are written into the teaching segments corresponding to the teacher questions and children's responses; and the affective objectives are written into the teaching segments corresponding to the summary evaluation and home-school collaboration. The teaching steps are arranged in the order of their positions to form a sequence.

5. The preschool education teaching content generation system based on a large language model according to claim 4, characterized in that, The operation of maintaining the encoding of the execution target includes: Read the target sub-item number, target sub-item text, and target type from the target sub-item set; read the teaching segment position from the teaching segment sequence. The target sub-item number, target sub-item text, target type, and teaching link position are matched one by one to form the target retention input unit; Arrange the target retention input units according to the sequence of teaching steps to form a target retention input sequence; Perform text embedding encoding on the target sub-item text in the input sequence to obtain the target sub-item encoding; Perform category embedding encoding on the target type to obtain the target type encoding; Position embedding encoding is performed on the location of each teaching segment to obtain the teaching segment code; Write the target sub-item encoding into the key vector of the RetNet preserving network, and write the target type encoding and teaching segment encoding into the value vector of the RetNet preserving network; The RetNet network reads the instruction code sequentially according to the instruction sequence, reads the key vector corresponding to the target sub-item code at each instruction position, and writes the value vector corresponding to the key vector into the hold state of that instruction position. After writing the state of all teaching steps according to the teaching step sequence, output the target sub-item holding vector and the step holding vector.

6. The preschool education teaching content generation system based on a large language model according to claim 5, characterized in that, The operation of establishing the target coverage table includes: Read target sub-items from the target sub-item set, read the teaching segment positions from the teaching segment sequence, and build a blank coverage table with target sub-items as rows and teaching segment positions as columns; Read the target sub-item preservation vector and link preservation vector output by the RetNet target preservation encoding module; Perform vector normalization on the target sub-item preservation vector and the link preservation vector to obtain the target sub-item normalized vector and the link normalized vector; The similarity value between the target sub-item and the teaching segment is obtained by multiplying the normalized vector of the target sub-item and the normalized vector of the segment dimension by dimension and then summing them. Write the similarity values ​​into the cells corresponding to the target sub-item and the corresponding teaching segment in the blank coverage table; Fill all cells according to the correspondence between the target sub-items and the teaching links to form a target coverage table.

7. The preschool education teaching content generation system based on a large language model according to claim 6, characterized in that, The operation of maintaining residual coverage for the execution target includes: Read the similarity values ​​row by row from the target coverage table according to the target sub-items, and add the similarity values ​​corresponding to the same target sub-item in the order of the teaching link sequence to obtain the target sub-item coverage cumulative value; The cumulative coverage value of the target sub-item is compared with the preset coverage threshold. When the cumulative coverage value of the target sub-item is lower than the preset coverage threshold, the corresponding target sub-item is marked as an uncovered target sub-item. Read the target item number, target item text, and target type corresponding to the uncovered target item from the target item set; Read all similarity values ​​of the rows containing uncovered target sub-items from the target coverage table, and establish a correspondence between each similarity value and the corresponding teaching segment position in the teaching segment sequence; Write the target sub-item number, target sub-item text, target type, target sub-item coverage cumulative value, teaching link location and corresponding similarity value of the uncovered target sub-item into the residual preservation record; Write the target sub-item number and target sub-item text in the residual preservation record into the residual key vector of the RetNet preservation network, and write the target type, target sub-item coverage accumulation value, teaching link position and corresponding similarity value in the residual preservation record into the residual value vector of the RetNet preservation network. RetNet keeps the network reading the residual key vector and residual value vector, transfers the uncovered target sub-items to the target coverage residual retention state, and sends the target coverage residual retention state to the link adaptation retention module.

8. The preschool education teaching content generation system based on a large language model according to claim 7, characterized in that, The adaptation and maintenance operations in the execution phase include: Read the uncovered target sub-items, target types, cumulative coverage values ​​of target sub-items, teaching link locations, and corresponding similarity values ​​from the target coverage residual preservation status; Read the mapping table between target type and teaching link. The mapping table records the corresponding activities, exploration tasks and teacher questions for the cognitive target sub-items, the exploration tasks and game interactions for the ability target sub-items, the teacher questions and children's responses for the expression target sub-items, and the summary evaluation and home-school cooperation for the affective target sub-items. Based on the target type of the uncovered target sub-item, read the corresponding mapped teaching segment from the target type and teaching segment mapping table; In the target coverage residual preservation state, read the similarity value between the uncovered target sub-item and each mapping teaching link, and arrange the mapping teaching links in order of similarity value from low to high; The mapping teaching segment, which is ranked first, is identified as a supplementary objective segment. When two mapping teaching segments have the same similarity value, the mapping teaching segment that is arranged later in the preset teaching segment order is selected as the target supplementary segment. Write the uncovered target sub-items, target types, and target sub-item coverage accumulation values ​​from the target coverage residual retention state into the retention state of the target supplementation stage; Write the target replenishment marker into the target replenishment state; The state of each teaching step is maintained after being written according to the teaching step sequence, forming a target-closed teaching step state sequence, which is then sent to the large language model constraint generation module.

9. A preschool education teaching content generation system based on a large language model according to claim 8, characterized in that, The operations for generating execution constraints include: Read the state sequence of the closed teaching links of the target, and extract the teaching link name, teaching link maintenance state, target sub-item, target type and target supplementary mark item by item according to the order of the teaching link sequence; The teaching segment name, target sub-item, target type, teaching segment hold status, and target supplementary mark are assembled into the corresponding teaching segment decoding control data in the order of segment name, target sub-item, target type, hold status, and supplementary mark; The decoding control data of each teaching segment is concatenated into the model input text for the corresponding teaching segment; The input text to the model is used as the current input to the pre-trained large language model to generate the corresponding teaching text. When there are target supplementary markers in the input text of the model, supplementary content containing the target sub-items corresponding to the target supplementary markers will be generated in the corresponding teaching segment text; The texts of each teaching segment are merged according to the order of the teaching segment sequence to form the preschool education teaching content.

10. A preschool education teaching content generation system based on a large language model according to claim 9, characterized in that, The preschool education teaching content includes activity introduction text segments, exploration task text segments, game interaction text segments, teacher question text segments, children's response text segments, summary and evaluation text segments, and home-school co-education text segments. The large language model constraint generation module writes the text of each teaching segment into the corresponding text segment according to the arrangement order of the teaching segment sequence, establishes the correspondence between text segment identifiers, teaching segment positions, target sub-items, and target supplementary markers, and forms preschool education teaching content with target correspondence.

Images