Middle and primary school language learning auxiliary method based on interaction of education large model and digital human

Abstract

The invention discloses a middle and primary school language learning auxiliary method based on interaction of an education large model and a digital human, and the method comprises the steps: collecting middle and primary school language learning data, and carrying out the preprocessing of the collected language learning data; generating a corresponding student personalized language learning file; a language knowledge graph is constructed, the language knowledge graph covers multi-dimensional knowledge information of vocabularies, grammar and sentence patterns, and a mapping relation is established between the language knowledge graph and language learning features; forming an adaptive teaching scheme; the digital human virtual teaching assistant performs real-time voice, text and expression interaction with the students by using a multi-mode interaction technology, and adjusts an interaction strategy according to real-time feedback of the students; and the called language knowledge graph dynamically supplements the context of the students, so that situational language teaching is realized, and a real communication environment is simulated in a multi-language application scene. The language migration ability and the communication application level of students are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of language technology in primary and secondary schools, and in particular to a method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction. Background Technology

[0002] With the rapid development of artificial intelligence, large model technology and digital virtual human technology, intelligent assistance and personalized teaching have received widespread attention in the field of language learning in primary and secondary schools. Existing language learning assistance methods in primary and secondary schools mainly include question bank push, fixed process assessment and simple speech recognition scoring methods. They are usually driven by rules or templates and lack the ability to dynamically analyze individual student characteristics, make real-time adaptive adjustments and multimodal interaction.

[0003] Existing systems often only support single-modal data collection and processing, failing to fully integrate the multi-dimensional language data generated by students in different learning activities. This results in an incomplete portrayal of students' true learning abilities, with ability assessments primarily relying on static tests or score-based evaluations. They lack in-depth feature extraction and multi-dimensional ability profile modeling, making it difficult to accurately reflect students' specific weaknesses in pronunciation, grammar, vocabulary, and expression.

[0004] In terms of virtual teaching processes, existing auxiliary systems mainly rely on text-based question-and-answer or voice scoring, which cannot achieve high-quality real-time multimodal interaction of voice, text, and facial expressions. They also lack the ability to dynamically supplement contextualized language application scenarios, making it difficult to effectively simulate real language communication environments and affecting students' actual language expression and comprehensive application abilities. Summary of the Invention

[0005] One objective of this invention is to propose a language learning assistance method for primary and secondary schools based on a large-scale educational model and digital human interaction. This invention effectively improves students' language transfer ability and communicative application level.

[0006] According to an embodiment of the present invention, a method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction includes:

[0007] Collect language learning data from primary and secondary schools, and preprocess the collected language learning data;

[0008] Using a large-scale educational model, we conduct in-depth analysis of preprocessed primary and secondary school language learning data, extract students' language learning characteristics in pronunciation, grammar, vocabulary and expression, and generate corresponding personalized language learning profiles for students.

[0009] Construct a language knowledge graph that covers multi-dimensional knowledge information such as vocabulary, grammar and sentence structure, and establish a mapping relationship with language learning features;

[0010] Personalized learning paths are generated based on mapping relationships. These paths are dynamically planned according to students' weaknesses in various language skills and their level of knowledge mastery, and then form adaptive teaching plans.

[0011] The personalized learning path is implemented through a digital human virtual teaching assistant. The digital human virtual teaching assistant uses multimodal interaction technology to interact with students in real time through voice, text and facial expressions, and adjusts the interaction strategy based on students' real-time feedback.

[0012] During the interaction between the digital human virtual teaching assistant and the student, the language knowledge graph is dynamically supplemented to the student's context, realizing contextualized language teaching and simulating real communication environments in various language application scenarios.

[0013] Optionally, the process of collecting language learning data from primary and secondary schools and preprocessing the collected language learning data includes:

[0014] The original language learning dataset is formed by collecting student voice data, text input data, homework data, and assessment data from language learning activities in primary and secondary schools.

[0015] The original language learning dataset is cleaned to form a cleaned language learning dataset.

[0016] The cleaned language learning dataset is structured by converting speech data into speech signal spectrum vectors, and performing word segmentation, part-of-speech tagging, and syntactic analysis on text input data, assignment data, and assessment data respectively to generate structured language expression vectors and construct a standardized structured dataset.

[0017] The standardized structured dataset is imputed and outliers are detected to form a preprocessed language learning dataset.

[0018] Optionally, generating corresponding personalized language learning profiles for students includes:

[0019] The preprocessed language learning dataset is transformed using a large-scale educational model, converting each piece of language learning data into an embedded representation.

[0020] Based on the embedding representation, the language learning ability features of each student in the pronunciation, grammar, vocabulary and expression dimensions are extracted to obtain the pronunciation ability feature vector, grammar ability feature vector, vocabulary ability feature vector and expression ability feature vector;

[0021] Normalization and cluster analysis were performed on the feature vectors of pronunciation ability, grammar ability, vocabulary ability, and expression ability. Based on the students' performance in the four ability dimensions, a student ability profile was constructed. The feature vectors of pronunciation ability, grammar ability, vocabulary ability, and expression ability of all students were arranged in a way that the students were in rows and the ability dimensions were in columns, forming a multidimensional matrix of students' language ability.

[0022] A personalized language learning portfolio is constructed based on a multidimensional matrix of students' language proficiency, forming a multi-field personalized portfolio structure.

[0023] Optionally, the construction of the language knowledge graph includes:

[0024] Establish a set of vocabulary nodes, unify word form, word meaning and part-of-speech tagging, and configure a unique identifier, difficulty level and use case index for each vocabulary node;

[0025] Establish a set of grammar rule nodes, standardize the descriptions of tense, voice, subject-verb agreement, subordinate structure and dependency relationship rules, and assign a unique identifier and scope of application label to each grammar rule node;

[0026] Establish a set of sentence pattern nodes, define sentence pattern templates for declarative sentences, interrogative sentences, imperative sentences, exclamatory sentences, and compound sentences, and assign a unique identifier, component slots, and collocation constraints to each sentence pattern node;

[0027] Define knowledge relationship types and generate edge sets;

[0028] The knowledge relationship types include at least the lexical-grammatical constraint relationship, the sentence structure-grammatical composition relationship, the lexical-sentence structure filling relationship, and the synonym / antonym / hyper-paragraph relationship, and the relationship edges are configured with directionality, weight and timestamp;

[0029] The vocabulary nodes, grammar rule nodes, and sentence structure nodes are standardized and disambiguated. Duplicate nodes are merged, ambiguous entries are eliminated, and version management is completed to form a language knowledge graph with a consistent structure.

[0030] Based on the multidimensional matrix of students' language proficiency and students' personalized language learning profiles, we extract the feature vectors of each student's pronunciation ability, grammar ability, vocabulary ability, and expression ability, and perform association calculations with the corresponding nodes in the language knowledge graph to generate student-knowledge node mapping pairs and their association weights.

[0031] Establish a bidirectional index mapping table, store student-knowledge node mapping pairs and their associated weights using a double key of student unique identifier and knowledge graph unique identifier, and record update timestamps and version numbers.

[0032] Optionally, the formation of the adaptive teaching plan includes:

[0033] Extract the knowledge mastery vector for each student;

[0034] By multiplying the node's weight parameter by the student's knowledge mastery of the node, a knowledge mastery difference vector for each student is obtained.

[0035] The language knowledge nodes are sorted according to the knowledge mastery difference vector, and the nodes with the largest knowledge mastery difference values ​​are selected to construct the student's personalized target knowledge set.

[0036] In the language knowledge graph, the student's personalized target knowledge set is used as the target node set, and the student's current mastered node set is used as the starting node set. Path search is performed through the directed weighted edges in the language knowledge graph to construct the shortest path set that satisfies the constraints of connectivity, dependency and asymptoticity. The shortest path set is the student's personalized learning path.

[0037] Based on the student's personalized learning path and the student's ability profile, each knowledge node in the learning path is adapted sequentially. According to the ability requirements of each knowledge node, all teaching content in the teaching resource library that matches the knowledge node and can meet the ability requirements corresponding to the student's ability profile is retrieved. The teaching resources most suitable for the student's current knowledge mastery and ability level are selected and arranged into a teaching resource set in the order of the knowledge nodes in the learning path.

[0038] Based on each learning node in the personalized learning path, corresponding teaching resources are allocated to each learning node in sequence, and dynamic adjustment parameters are set for each learning node according to the student's language ability characteristics. Each learning node, the corresponding teaching resources, and the dynamic adjustment parameters set for the learning node are combined in sequence to form multiple sets of triplets. All triplets are arranged in sequence to form the student's adaptive teaching plan.

[0039] Optionally, the formation of the adaptive teaching plan includes:

[0040] Initialize the digital human virtual teaching assistant, load the student's personalized learning path and adaptive teaching plan, collect the voice input, text input and facial expression data generated by the student during the learning process, and form a student interaction feedback dataset according to the time sequence of collection.

[0041] The student interaction feedback dataset is input into the multimodal coding module, and features are extracted from the speech modality, text modality and facial expression modality data respectively. The different modal data are mapped into a set of multimodal feedback vectors of the same dimension.

[0042] The multimodal feedback vector set and the current learning node state are input into the interaction policy discrimination function, which outputs the interaction response category and adjustment coefficient at the current interaction time step.

[0043] Based on the interactive response category and adjustment coefficient, the digital human virtual teaching assistant generates corresponding teaching behaviors;

[0044] The updated dynamic adjustment parameters are applied to the current learning node to adjust the difficulty of the learning content and the task time length of the learning node in real time. The learning node number, corresponding teaching resources and actual application adjustment parameters of all completed learning nodes are arranged in the order of interaction time steps to form a teaching execution record table.

[0045] The teaching execution record form is linked and updated with the student's personalized language learning portfolio to generate a new version of the personalized language learning portfolio that incorporates real-time interactive feedback.

[0046] Optionally, simulating a real communication environment in multiple language application scenarios includes:

[0047] Each time the digital human virtual teaching assistant interacts with a student, the student's current learning context information is collected, and all learning context information is arranged in the order of the interaction time steps to form a learning context set.

[0048] The learning context set and the language knowledge graph are input into the contextualized knowledge supplementation module. Based on the current learning node number and historical feedback in the learning context set, the contextualized knowledge supplementation module retrieves the vocabulary nodes, grammar rule nodes and sentence pattern nodes associated with the current learning node number in the language knowledge graph, and selects knowledge content that is appropriate to the current student's knowledge mastery level and ability profile, and dynamically generates a set of contextualized supplementary content.

[0049] The contextualized supplementary content is integrated with the current interactive content to generate a context-enhanced interactive task package.

[0050] During multiple rounds of interaction, the difficulty level, topic type, and interaction method of the contextualized supplementary content set and interactive task package are dynamically adjusted based on students' performance data in different contexts.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention proposes a data acquisition and feature processing mechanism that integrates modal inputs such as speech, text and facial expressions. A structured language learning dataset with a unified scale is constructed for each student. Through multi-layer embedding expression and cluster analysis algorithms, the ability features of students in the four core dimensions of pronunciation, grammar, vocabulary and expression are automatically extracted and quantified to generate an iterative multi-dimensional student ability profile. This greatly improves the comprehensive characterization granularity and dynamic tracking ability of students' language abilities, and enables the ability profile to have cross-scenario transferability and continuous optimization capabilities.

[0053] (2) This invention constructs a multidimensional language knowledge graph with consistent structure and versionable nodes, introduces a bidirectional dynamic mapping table between students and knowledge nodes, automatically calculates the knowledge mastery difference vector based on the degree of knowledge mastery and weight parameters to accurately locate the weak links of students in each knowledge node, and adopts graph theory path planning and constrained shortest path search algorithm to realize the dynamic determination of personalized target knowledge sets. Combined with the student ability profile, the optimal teaching resources are retrieved and matched to realize the three-dimensional adaptive path generation of ability-knowledge-resource, which significantly improves the personalization, coherence and timeliness of learning content push.

[0054] (3) In the implementation of the digital human virtual teaching assistant, the present invention integrates a multimodal real-time interactive feedback collection, fusion and dynamic strategy adjustment mechanism. It utilizes a contextualized knowledge supplementation module to call up the language knowledge graph in real time to enrich the interactive context. Based on the student's ability performance and learning node status, it simulates the real language communication environment through multimodal output. The system can dynamically recommend and adjust task packages and contextual content based on the student's historical interaction and knowledge structure, effectively solving the problem of the lack of real communication experience and low immersion in the existing system. This makes the language learning process closer to real life and communication practice, and effectively improves the student's language transfer ability and communicative application level. Attached Figure Description

[0055] 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:

[0056] Figure 1 This is a flowchart of a language learning assistance method for primary and secondary schools based on a large-scale educational model and digital human interaction, proposed in this invention. Detailed Implementation

[0057] 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.

[0058] refer to Figure 1 A method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction, comprising:

[0059] Collect language learning data from primary and secondary schools, and preprocess the collected language learning data;

[0060] In this embodiment, language learning data from primary and secondary schools is collected, and the collected language learning data is preprocessed, including:

[0061] The original language learning dataset is formed by collecting student voice data, text input data, homework data, and assessment data from language learning activities in primary and secondary schools.

[0062] Each piece of language learning data in the original language learning dataset includes a student identifier, a data collection timestamp, a data category, and the original content. The total amount of the original language learning dataset is denoted as the total data amount.

[0063] The original language learning dataset is cleaned to form a cleaned language learning dataset.

[0064] By removing duplicate data, null data, and data that failed to be recognized from the original language learning dataset, and retaining the valid samples, a cleaned language learning dataset is formed. The cleaned language learning dataset only contains language learning data whose original content is not empty and which has been determined to be valid.

[0065] The cleaned language learning dataset is structured by converting speech data into speech signal spectrum vectors, and performing word segmentation, part-of-speech tagging, and syntactic analysis on text input data, assignment data, and assessment data respectively to generate structured language expression vectors and construct a standardized structured dataset.

[0066] A standardized structured dataset is composed of a structured speech signal spectrum vector and a structured language expression vector. The number of structured speech data is denoted as the speech data quantity, and the number of structured text data is denoted as the text data quantity.

[0067] The standardized structured dataset is imputed and outliers are detected to form a preprocessed language learning dataset.

[0068] In Example 1, the nearest neighbor interpolation method is used to interpolate and fill in the missing data in the structured speech signal spectrum vector and the structured language expression vector. The three-standard deviation method is used to identify and remove abnormal samples to form a preprocessed language learning dataset. The preprocessed language learning dataset only contains the speech signal spectrum vector and the structured language expression vector that have been interpolated and whose abnormality is identified as non-abnormal data.

[0069] Using a large-scale educational model, we conduct in-depth analysis of preprocessed primary and secondary school language learning data, extract students' language learning characteristics in pronunciation, grammar, vocabulary and expression, and generate corresponding personalized language learning profiles for students.

[0070] In this embodiment, a corresponding personalized language learning profile for the student is generated, including:

[0071] The preprocessed language learning dataset is transformed using a large-scale educational model, converting each piece of language learning data into an embedded representation.

[0072] An embedding is a semantic information of the language learning data in the feature space, and the embedding representation has a unified dimension.

[0073] Based on the embedding representation, the language learning ability features of each student in the pronunciation, grammar, vocabulary and expression dimensions are extracted to obtain the pronunciation ability feature vector, grammar ability feature vector, vocabulary ability feature vector and expression ability feature vector;

[0074] Normalization and cluster analysis were performed on the feature vectors of pronunciation ability, grammar ability, vocabulary ability, and expression ability. Based on the students' performance in the four ability dimensions, a student ability profile was constructed. The feature vectors of pronunciation ability, grammar ability, vocabulary ability, and expression ability of all students were arranged in a way that the students were in rows and the ability dimensions were in columns, forming a multidimensional matrix of students' language ability.

[0075] Each row in the student language proficiency multidimensional matrix corresponds to a student's multidimensional language proficiency information.

[0076] A personalized language learning portfolio is constructed based on a multidimensional matrix of students' language proficiency, forming a multi-field personalized portfolio structure.

[0077] Construct a language knowledge graph that covers multi-dimensional knowledge information such as vocabulary, grammar and sentence structure, and establish a mapping relationship with language learning characteristics;

[0078] In this embodiment, the construction of a language knowledge graph includes:

[0079] Establish a set of vocabulary nodes, unify word form, word meaning and part-of-speech tagging, and configure a unique identifier, difficulty level and use case index for each vocabulary node;

[0080] Establish a set of grammar rule nodes, standardize the descriptions of tense, voice, subject-verb agreement, subordinate structure and dependency relationship rules, and assign a unique identifier and scope of application label to each grammar rule node;

[0081] Establish a set of sentence pattern nodes, define sentence pattern templates for declarative sentences, interrogative sentences, imperative sentences, exclamatory sentences, and compound sentences, and assign a unique identifier, component slots, and collocation constraints to each sentence pattern node;

[0082] Define knowledge relationship types and generate edge sets;

[0083] The knowledge relationship types include at least the lexical-grammatical constraint relationship, the sentence structure-grammatical composition relationship, the lexical-sentence structure filling relationship, and the synonym / antonym / hyper-paragraph relationship, and the relationship edges are configured with directionality, weight and timestamp;

[0084] The vocabulary nodes, grammar rule nodes, and sentence structure nodes are standardized and disambiguated. Duplicate nodes are merged, ambiguous entries are eliminated, and version management is completed to form a language knowledge graph with a consistent structure.

[0085] Based on the multidimensional matrix of students' language proficiency and students' personalized language learning profiles, we extract the feature vectors of each student's pronunciation ability, grammar ability, vocabulary ability, and expression ability, and perform association calculations with the corresponding nodes in the language knowledge graph to generate student-knowledge node mapping pairs and their association weights.

[0086] Establish a bidirectional index mapping table, store student-knowledge node mapping pairs and their associated weights using a double key of student unique identifier and knowledge graph unique identifier, and record update timestamps and version numbers.

[0087] Personalized learning paths are generated based on mapping relationships. These paths are dynamically planned according to students' weaknesses in various language skills and their level of knowledge mastery, and then form adaptive teaching plans.

[0088] In this embodiment, an adaptive teaching plan is formed, including:

[0089] Extract the knowledge mastery vector for each student;

[0090] The knowledge mastery vector is used to describe the student's mastery of each language knowledge node. The knowledge mastery value ranges from zero to one, and the total number of knowledge nodes is the total number of nodes in the language knowledge graph.

[0091] By multiplying the node's weight parameter by the student's knowledge mastery of the node, a knowledge mastery difference vector for each student is obtained.

[0092] The knowledge mastery difference vector is used to measure the learning priority of students at each language knowledge node, the knowledge mastery level is used to measure the students' proficiency in the knowledge of that node, and the weight parameter represents the importance of the knowledge node in the overall learning path.

[0093] The language knowledge nodes are sorted according to the knowledge mastery difference vector, and the nodes with the largest knowledge mastery difference values ​​are selected to construct the student's personalized target knowledge set.

[0094] The personalized target knowledge set is the set of knowledge nodes that the student needs to focus on improving.

[0095] In the language knowledge graph, the student's personalized target knowledge set is used as the target node set, and the student's current mastered node set is used as the starting node set. Path search is performed through the directed weighted edges in the language knowledge graph to construct the shortest path set that satisfies the constraints of connectivity, dependency and asymptoticity. The shortest path set is the student's personalized learning path.

[0096] Based on the student's personalized learning path and the student's ability profile, each knowledge node in the learning path is adapted sequentially. According to the ability requirements of each knowledge node, all teaching content in the teaching resource library that matches the knowledge node and can meet the ability requirements corresponding to the student's ability profile is retrieved. The teaching resources most suitable for the student's current knowledge mastery and ability level are selected and arranged into a teaching resource set in the order of the knowledge nodes in the learning path.

[0097] The length of the teaching resource set is equal to the length of the personalized learning path, and each teaching resource uniquely corresponds to a knowledge node in the learning path.

[0098] Based on each learning node in the personalized learning path, corresponding teaching resources are allocated to each learning node in sequence, and dynamic adjustment parameters are set for each learning node according to the student's language ability characteristics. Each learning node, the corresponding teaching resources, and the dynamic adjustment parameters set for the learning node are combined in sequence to form multiple sets of triplets. All triplets are arranged in sequence to form the student's adaptive teaching plan.

[0099] The dynamic adjustment parameters are used to adjust the teaching difficulty and learning task time of the node according to the student's current learning performance. Each triplet consists of three parts: learning node, teaching resources, and dynamic adjustment parameters.

[0100] Personalized learning paths are implemented through digital human virtual teaching assistants. These assistants use multimodal interaction technology to interact with students in real time via voice, text, and facial expressions, and adjust their interaction strategies based on real-time student feedback.

[0101] In this embodiment, an adaptive teaching plan is formed, including:

[0102] Initialize the digital human virtual teaching assistant, load the student's personalized learning path and adaptive teaching plan, collect the voice input, text input and facial expression data generated by the student during the learning process, and form a student interaction feedback dataset according to the time sequence of collection.

[0103] Each data item in the student interaction feedback dataset records all multimodal feedback information from the student at each interaction time step, and the total number of interaction time steps is the total number of interaction steps.

[0104] The student interaction feedback dataset is input into the multimodal coding module, and features are extracted from the speech modality, text modality and facial expression modality data respectively. The different modal data are mapped into a set of multimodal feedback vectors of the same dimension.

[0105] In Example 1, the student interaction feedback dataset is input into the multimodal encoding module, and each item in the student interaction feedback dataset is processed. Each item includes voice modality data, text modality data, and facial expression modality data.

[0106] For speech modal data, audio features of the speech signal are extracted and converted into speech feature vectors of a unified dimension; for text modal data, semantic features of the text content are extracted and converted into text feature vectors of a unified dimension; for facial expression modal data, emotional features of facial expression images or video frames are extracted and converted into facial expression feature vectors of a unified dimension.

[0107] The speech feature vector, text feature vector, and facial expression feature vector at the same time step are fused according to a predetermined fusion method to obtain the multimodal feedback vector corresponding to the time step. The multimodal feedback vectors obtained from all interactive time steps are arranged in sequence to form a set of multimodal feedback vectors of a unified dimension.

[0108] The multimodal feedback vector set and the current learning node state are input into the interaction policy discrimination function, which outputs the interaction response category and adjustment coefficient at the current interaction time step.

[0109] The interaction strategy discriminant function is a discriminant model trained on a large amount of historical student interaction data and learning outcome data through an educational big data model. The interaction strategy discriminant function outputs the interaction response category and adjustment coefficient based on the multimodal feedback vector of the current interaction time step and the current learning node state. The interaction response category is used to indicate the type of teaching behavior that the digital human virtual teaching assistant needs to take, the adjustment coefficient is used to adjust the dynamic adjustment parameters of the current learning node in the adaptive teaching scheme, and the current learning node state is used to assist the interaction strategy discriminant function in making accurate judgments.

[0110] Based on the interactive response category and adjustment coefficient, the digital human virtual teaching assistant generates corresponding teaching behaviors;

[0111] The teaching behavior is presented through multimodal output methods such as speech synthesis, text push and facial expression display. At the same time, the dynamic adjustment parameters of the current learning node in the adaptive teaching plan are updated according to the adjustment coefficient.

[0112] The updated dynamic adjustment parameters are applied to the current learning node to adjust the difficulty of the learning content and the task time length of the learning node in real time. The learning node number, corresponding teaching resources and actual application adjustment parameters of all completed learning nodes are arranged in the order of interaction time steps to form a teaching execution record table.

[0113] The teaching execution record form is linked and updated with the student's personalized language learning portfolio to generate a new version of the personalized language learning portfolio that incorporates real-time interactive feedback.

[0114] During the interaction between the digital human virtual teaching assistant and the student, the language knowledge graph is dynamically supplemented to the student's context, realizing contextualized language teaching and simulating real communication environments in various language application scenarios.

[0115] In this embodiment, a real communication environment is simulated in multiple language application scenarios, including:

[0116] Each time the digital human virtual teaching assistant interacts with a student, the student's current learning context information is collected, and all learning context information is arranged in the order of the interaction time steps to form a learning context set.

[0117] The learning context information includes the current learning node number, the current interaction content, the student's performance status at the node, and the historical interaction feedback associated with the node.

[0118] The learning context set and the language knowledge graph are input into the contextualized knowledge supplementation module. Based on the current learning node number and historical feedback in the learning context set, the contextualized knowledge supplementation module retrieves the vocabulary nodes, grammar rule nodes and sentence pattern nodes associated with the current learning node number in the language knowledge graph, and selects knowledge content that is appropriate to the current student's knowledge mastery level and ability profile, and dynamically generates a set of contextualized supplementary content.

[0119] The contextualized supplementary content set includes high-frequency vocabulary, practical grammatical structures, and relevant emotional expressions to enrich the current dialogue context.

[0120] The contextualized supplementary content is integrated with the current interactive content to generate a context-enhanced interactive task package.

[0121] The vocabulary nodes, grammar rule nodes, and sentence structure nodes in the contextualized supplementary content set are sorted according to the knowledge point requirements of the current learning node and the student's ability profile, based on the priority and relevance of the knowledge points. Several items with the highest ranking are selected and inserted or replaced in the corresponding knowledge point positions in the current interactive content or introduced as new scene interactive script fragments. This ensures that the integrated interactive task package contains both the student's original learning task and contextual knowledge points closely related to the current context. The interactive task package consists of integrated dialogue text, voice output script, and triggerable facial expressions and actions. All multimodal content to be output corresponds to and is consistent with the knowledge objectives, ability development direction, and contextualized supplementary content of the current learning node, ensuring that the task package has scene integrity, pragmatic relevance, and targeted ability improvement.

[0122] During multiple rounds of interaction, the difficulty level, topic type, and interaction method of the contextualized supplementary content set and interactive task package are dynamically adjusted based on students' performance data in different contexts.

[0123] All performance data and supplementary content during the contextualized interaction process, as well as the version number and output content of the task package, are arranged in the order of the interaction time steps to form a contextualized interaction record table. The contextualized interaction record table is bound and updated with the student's personalized language learning portfolio to form a multi-dimensional personalized portfolio with a contextualized learning history.

[0124] Example 1:

[0125] Using the practical application of this invention in English learning by "Student A," a seventh-grade student at Middle School A in a certain city, as a case study, this paper describes the implementation process and effectiveness comparison of this invention in a real educational scenario, demonstrating the feasibility and effectiveness of this invention.

[0126] Student A is an average first-year junior high school student. He is introverted, has weak communication skills, and his English grades have consistently been in the lower half of his class. Traditional teaching methods rely mainly on standardized textbooks, classroom lectures, and homework correction, lacking in-depth analysis of Student A's individual learning abilities and making it difficult to achieve truly individualized instruction. Teachers can only make rough judgments about students based on test scores and classroom performance, failing to gain a deeper understanding of Student A's strengths and weaknesses in specific areas such as pronunciation, grammar, vocabulary, and expression.

[0127] After this invention was launched on the smart education platform of Middle School A, student A began using the English learning assistance system based on a large-scale educational model and digital human interaction. Throughout the application process, student A's learning activities were fully collected and analyzed from multiple dimensions:

[0128] Student A completes daily English homework and oral practice on the platform. The system automatically collects his audio readings (textbook readings, dialogue imitations, and word spellings in Example 2), text inputs (English essays, vocabulary fill-in-the-blanks, and grammar corrections in Example 2), completion time for each assignment, error analysis reports, and classroom assessment scores, generating a raw dataset. In Example 2, Student A uploaded 82 audio readings, submitted 14 English essays, completed 172 vocabulary and grammar multiple-choice questions, and participated in 3 unit tests, totaling approximately 2100 data entries.

[0129] The system first performs standardized formatting, missing item completion, and anomaly detection on the original dataset, removing 8 incomplete data entries caused by audio loss or network anomalies. It also performs spell correction, sentence segmentation, and annotation on the input essay text, ultimately forming a structured learning dataset that is free of redundancy and anomalies.

[0130] Based on the preprocessed data, the education big data model uses end-to-end acoustic-semantic joint embedding for each audio data point, converting each student's reading audio into a 128-dimensional embedding feature vector. For text essays and homework questions, the system uses a deep Transformer model to extract the text's grammatical structure, lexical richness, and pragmatic expression ability, and finally converts each assignment into numerical features in four ability dimensions.

[0131] After a month of cumulative training and cluster analysis, Student A's ability profile was as follows: pronunciation accuracy 74 points (out of 100, class average 80 points), grammar mastery 72 points (class average 77 points), vocabulary richness 69 points (class average 74 points), and fluency 60 points (class average 71 points). Simultaneously, the model automatically identified significant weaknesses in Student A's verb tenses, long sentence structure, spelling of some high-frequency words, and fluent spoken expression.

[0132] The platform synchronously accesses a language knowledge graph, designating all grammar points, vocabulary, and common sentence patterns currently mastered by student A as "mastered nodes" and weak knowledge points in the ability profile as "nodes to be improved." Based on the logical dependencies between knowledge points and the actual learning difficulty, the platform automatically generates a personalized learning path for student A. The model assigns weights to all knowledge points and calculates priorities accordingly, forming a sequence of target nodes.

[0133] Within a month, Student A's personalized learning path included 12 knowledge nodes, such as "present continuous tense", "past tense usage", "object clause conjunction usage", "high-frequency adjective memorization" and "expression in asking for directions". Each node was accompanied by three task-based exercises of different difficulty levels and 5-7 high-frequency interactive sentence patterns.

[0134] A digital virtual teaching assistant (represented by a 3D virtual teacher "B") interacts with student A throughout the learning process using speech recognition, text understanding, and facial expression analysis modules. During student A's daily self-study, B proactively pushes pronunciation training tasks, situational dialogues, and interactive grammar quizzes based on the learning path. In Example 2, when learning the topic of "asking for directions," B simulates three real-life communication scenarios: asking for directions at a supermarket, asking for directions at an airport, and giving directions on campus, guiding student A to use the learned sentence patterns in practical expressions.

[0135] When student A makes a pronunciation error or shows signs of nervousness, B will automatically adjust the interaction strategy based on real-time feedback. In Example 2, this involves first encouraging the student, then slowing down the speech, or appropriately repeating or simplifying the explanation. All interaction data, adjustment behaviors, and student A's responses are recorded in detail and fed back to the system in real time, automatically updating the ability profile and learning path.

[0136] To verify the effectiveness of this invention, Middle School A conducted a controlled experiment over three months, selecting 74 students from two parallel 7th-grade classes: Class A (the invention group) with 38 students and Class B (the traditional group) with 36 students. Class A used the method of this invention, while Class B only used traditional homework correction and test paper assessment.

[0137] Four months later, the average pronunciation accuracy of students in Class A improved from 76 to 87 points, an increase of 11 points, while that of Class B improved from 75 to 79 points, an increase of 4 points. The average grammar mastery of Class A improved from 73 to 85 points, while that of Class B improved from 74 to 77 points; vocabulary richness of Class A improved from 70 to 82 points, while that of Class B improved from 71 to 74 points; and fluency of expression improved from 61 to 79 points, while that of Class B improved from 60 to 67 points. The number of frequently occurring weaknesses corrected in personalized ability profiles averaged 8.2 for Class A and 3.7 for Class B. The self-reported improvement rate of learning interest among students in Class A was 89%, while that in Class B was 56%. The workload of teachers manually grading in Class A decreased by approximately 47% per week on average, and the feedback time was shortened by an average of 1.5 days.

[0138] Partial sample (data source: actual sampling in a certain month):

[0139] Student A: Pronunciation accuracy improved from 74 to 88 points, fluency improved from 60 to 78 points, and weaknesses decreased from 11 to 2.

[0140] Student B (Class A): The score for vocabulary diversity in the composition improved from 65 points to 80 points.

[0141] Student C (Class B): Pronunciation accuracy improved from 72 to 76 points, and fluency improved from 63 to 68 points.

[0142] The platform's large model has an average monthly training sample count of over 120,000 (including voice, text, and facial expression modalities), and its contextualized task library is dynamically updated more than 2,000 times.

[0143] 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 method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction, characterized in that: include: Collect language learning data from primary and secondary schools, and preprocess the collected language learning data; Using a large-scale educational model, we conduct in-depth analysis of preprocessed primary and secondary school language learning data, extract students' language learning characteristics in pronunciation, grammar, vocabulary and expression, and generate corresponding personalized language learning profiles for students. Construct a language knowledge graph that covers multi-dimensional knowledge information such as vocabulary, grammar and sentence structure, and establish a mapping relationship with language learning features; Personalized learning paths are generated based on mapping relationships. These paths are dynamically planned according to students' weaknesses in various language skills and their level of knowledge mastery, and then form adaptive teaching plans. The personalized learning path is implemented through a digital human virtual teaching assistant. The digital human virtual teaching assistant uses multimodal interaction technology to interact with students in real time through voice, text and facial expressions, and adjusts the interaction strategy based on students' real-time feedback. During the interaction between the digital human virtual teaching assistant and the student, the language knowledge graph is dynamically supplemented to the student's context, realizing contextualized language teaching and simulating real communication environments in various language application scenarios.

2. The method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction as described in claim 1, characterized in that, The process of collecting language learning data from primary and secondary schools and preprocessing the collected language learning data includes: The original language learning dataset is formed by collecting student voice data, text input data, homework data, and assessment data from language learning activities in primary and secondary schools. The original language learning dataset is cleaned to form a cleaned language learning dataset. The cleaned language learning dataset is structured by converting speech data into speech signal spectrum vectors, and performing word segmentation, part-of-speech tagging, and syntactic analysis on text input data, assignment data, and assessment data respectively to generate structured language expression vectors and construct a standardized structured dataset. The standardized structured dataset is imputed and outliers are detected to form a preprocessed language learning dataset.

3. The method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction as described in claim 2, characterized in that, The generation of corresponding personalized language learning profiles for students includes: The preprocessed language learning dataset is transformed using a large-scale educational model, converting each piece of language learning data into an embedded representation. Based on the embedding representation, the language learning ability features of each student in the pronunciation, grammar, vocabulary and expression dimensions are extracted to obtain the pronunciation ability feature vector, grammar ability feature vector, vocabulary ability feature vector and expression ability feature vector; Normalization and cluster analysis were performed on the feature vectors of pronunciation ability, grammar ability, vocabulary ability, and expression ability. Based on the students' performance in the four ability dimensions, a student ability profile was constructed. The feature vectors of pronunciation ability, grammar ability, vocabulary ability, and expression ability of all students were arranged in a way that the students were in rows and the ability dimensions were in columns, forming a multidimensional matrix of students' language ability. A personalized language learning portfolio is constructed based on a multidimensional matrix of students' language proficiency, forming a multi-field personalized portfolio structure.

4. The method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction as described in claim 3, characterized in that, The construction of the language knowledge graph includes: Establish a set of vocabulary nodes, unify word form, word meaning and part-of-speech tagging, and configure a unique identifier, difficulty level and use case index for each vocabulary node; Establish a set of grammar rule nodes, standardize the descriptions of tense, voice, subject-verb agreement, subordinate structure and dependency relationship rules, and assign a unique identifier and scope of application label to each grammar rule node; Establish a set of sentence pattern nodes, define sentence pattern templates for declarative sentences, interrogative sentences, imperative sentences, exclamatory sentences, and compound sentences, and assign a unique identifier, component slots, and collocation constraints to each sentence pattern node; Define knowledge relationship types and generate edge sets; The knowledge relationship types include at least the lexical-grammatical constraint relationship, the sentence structure-grammatical composition relationship, the lexical-sentence structure filling relationship, and the synonym / antonym / hyper-paragraph relationship, and the relationship edges are configured with directionality, weight and timestamp; The vocabulary nodes, grammar rule nodes, and sentence structure nodes are standardized and disambiguated. Duplicate nodes are merged, ambiguous entries are eliminated, and version management is completed to form a language knowledge graph with a consistent structure. Based on the multidimensional matrix of students' language proficiency and students' personalized language learning profiles, we extract the feature vectors of each student's pronunciation ability, grammar ability, vocabulary ability, and expression ability, and perform association calculations with the corresponding nodes in the language knowledge graph to generate student-knowledge node mapping pairs and their association weights. Establish a bidirectional index mapping table, store student-knowledge node mapping pairs and their associated weights using a double key of student unique identifier and knowledge graph unique identifier, and record update timestamps and version numbers.

5. The method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction as described in claim 1, characterized in that, The formation of the adaptive teaching plan includes: Extract the knowledge mastery vector for each student; By multiplying the node's weight parameter by the student's knowledge mastery of the node, a knowledge mastery difference vector for each student is obtained. The language knowledge nodes are sorted according to the knowledge mastery difference vector, and the nodes with the largest knowledge mastery difference values ​​are selected to construct the student's personalized target knowledge set. In the language knowledge graph, the student's personalized target knowledge set is used as the target node set, and the student's current mastered node set is used as the starting node set. Path search is performed through the directed weighted edges in the language knowledge graph to construct the shortest path set that satisfies the constraints of connectivity, dependency and asymptoticity. The shortest path set is the student's personalized learning path. Based on the student's personalized learning path and the student's ability profile, each knowledge node in the learning path is adapted sequentially. According to the ability requirements of each knowledge node, all teaching content in the teaching resource library that matches the knowledge node and can meet the ability requirements corresponding to the student's ability profile is retrieved. The teaching resources most suitable for the student's current knowledge mastery and ability level are selected and arranged into a teaching resource set in the order of the knowledge nodes in the learning path. Based on each learning node in the personalized learning path, corresponding teaching resources are allocated to each learning node in sequence, and dynamic adjustment parameters are set for each learning node according to the student's language ability characteristics. Each learning node, the corresponding teaching resources, and the dynamic adjustment parameters set for the learning node are combined in sequence to form multiple sets of triplets. All triplets are arranged in sequence to form the student's adaptive teaching plan.

6. The method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction as described in claim 5, characterized in that, The formation of the adaptive teaching plan includes: Initialize the digital human virtual teaching assistant, load the student's personalized learning path and adaptive teaching plan, collect the voice input, text input and facial expression data generated by the student during the learning process, and form a student interaction feedback dataset according to the time sequence of collection. The student interaction feedback dataset is input into the multimodal coding module, and features are extracted from the speech modality, text modality and facial expression modality data respectively. The different modal data are mapped into a set of multimodal feedback vectors of the same dimension. The multimodal feedback vector set and the current learning node state are input into the interaction policy discrimination function, which outputs the interaction response category and adjustment coefficient at the current interaction time step. Based on the interactive response category and adjustment coefficient, the digital human virtual teaching assistant generates corresponding teaching behaviors; The updated dynamic adjustment parameters are applied to the current learning node to adjust the difficulty of the learning content and the task time length of the learning node in real time. The learning node number, corresponding teaching resources and actual application adjustment parameters of all completed learning nodes are arranged in the order of interaction time steps to form a teaching execution record table. The teaching execution record form is linked and updated with the student's personalized language learning portfolio to generate a new version of the personalized language learning portfolio that incorporates real-time interactive feedback.

7. The method for assisting primary and secondary school language learning based on a large-scale educational model and digital human interaction as described in claim 6, characterized in that, The simulation of real-world communication environments in multiple language application scenarios includes: Each time the digital human virtual teaching assistant interacts with a student, the student's current learning context information is collected, and all learning context information is arranged in the order of the interaction time steps to form a learning context set. The learning context set and the language knowledge graph are input into the contextualized knowledge supplementation module. Based on the current learning node number and historical feedback in the learning context set, the contextualized knowledge supplementation module retrieves the vocabulary nodes, grammar rule nodes and sentence pattern nodes associated with the current learning node number in the language knowledge graph, and selects knowledge content that is appropriate to the current student's knowledge mastery level and ability profile, and dynamically generates a set of contextualized supplementary content. The contextualized supplementary content is integrated with the current interactive content to generate a context-enhanced interactive task package. During multiple rounds of interaction, the difficulty level, topic type, and interaction method of the contextualized supplementary content set and interactive task package are dynamically adjusted based on students' performance data in different contexts.

Images