Classroom teaching quality evaluation method based on large language model

Abstract

The present invention discloses a classroom teaching quality evaluation method based on a large language model, which mainly comprises: based on classroom monitoring video data and teaching system data, using artificial intelligence technology to extract voice, image, and text multimodal features of teachers, students, and classroom environments; constructing a course knowledge graph according to course lesson plans and course information; enhancing the large language model through the constructed course knowledge graph, and realizing intelligent classroom teaching quality evaluation based on the enhanced large language model; the beneficial effects of the present invention are: enhancing the large language model through classroom teaching quality evaluation knowledge and the course knowledge graph, guiding the large language model to generate an evaluation result description that conforms to the classroom teaching quality evaluation facts based on the semantic association between the classroom teaching quality evaluation knowledge and the course knowledge graph; thereby enabling the model to generate a more reasonable description that conforms to the classroom teaching quality evaluation facts.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for evaluating classroom teaching quality based on a large language model. Background Art

[0002] The development of artificial intelligence (AI) and big data technology has promoted the level of informatization in various fields of education, and research on AI-based classroom teaching quality evaluation has made certain progress.

[0003] Educational researchers mainly focus on theoretical research on teaching quality evaluation methods and programs based on artificial intelligence technology, explore the changes and technical architecture of teaching quality evaluation enabled by artificial intelligence, and conduct research from the aspects of evaluation model, evaluation subject, evaluation content, evaluation results, education data analysis model, theoretical framework, etc.

[0004] With the development of big data and artificial intelligence technology, machine learning and deep learning algorithms have been widely used in classroom teaching quality evaluation, including methods based on neural networks, methods based on convolutional neural networks, methods based on attention mechanisms, etc. These studies mainly include two aspects: (1) Based on the scores of each evaluation indicator in the traditional teaching quality evaluation scale, neural networks are used to model classroom teaching evaluation and give evaluation result grades; (2) Classroom multimodal data analysis based on artificial intelligence algorithms, mainly based on classroom video images, using artificial intelligence algorithms to realize teacher and student behavior analysis, expression analysis, voice analysis, etc.

[0005] However, existing research on classroom teaching quality evaluation based on artificial intelligence mainly tends to focus on teaching and research theories, AI analysis based on the values ​​of teaching quality evaluation scales or classroom multimodal data, and an intelligent, full-process classroom teaching quality evaluation model is still an issue to be studied.

[0006] With the rise of ChatGPT, "AI+Education", driven by large language models, has demonstrated powerful self-supervised learning capabilities, intelligent adaptive teaching capabilities, and multimodal interaction capabilities, bringing profound changes to the educational ecosystem and arousing the interest of many researchers in the application of large language models in the field of education. Currently, good results have been achieved in fields such as mathematics, writing, code writing, and question-answering systems.

[0007] Classroom teaching quality evaluation is mainly based on manual evaluation, which has problems such as low evaluation efficiency, unreasonable evaluation process, strong subjectivity and lack of scientific basis. Therefore, this paper introduces a pre-trained large language model to implement classroom teaching quality evaluation. To address the problems existing in the application of large language models in specific fields, the large language model is enhanced by the course knowledge graph to generate objective, reasonable and effective classroom teaching quality evaluation results. Summary of the Invention

[0008] The purpose of the present invention is to provide a classroom teaching quality evaluation method based on a large language model. The technical problem to be solved is that when a large language model is applied to tasks in specific fields, there is a "hallucination" phenomenon. When generating classroom teaching quality evaluation results, due to the lack of educational knowledge, the generated evaluation result text description does not conform to the facts; when a large language model is applied to tasks in specific fields, there are problems of limited reasoning ability and poor interpretability.

[0009] The technical solution adopted by the present invention is: a classroom teaching quality evaluation method based on a large language model, which mainly includes:

[0010] Step S1: Based on classroom monitoring video data and educational system data, artificial intelligence technology is used to extract multimodal features of voice, image, and text of teachers, students, and classroom environment;

[0011] Step S2: constructing a course knowledge graph based on the course lesson plan and course information;

[0012] Step S3, enhancing the large language model through the course knowledge graph constructed in step S2, and realizing intelligent classroom teaching quality evaluation based on the enhanced large language model;

[0013] The specific steps of step S2 are:

[0014] Step S21: Collect course lesson plans and course information data from the educational system. The data is divided into image data and text data. The data specifically includes course syllabus, course teaching materials, course lesson plan PPT, course resources, and subject information.

[0015] Step S22, preprocessing the data collected in step S21, the preprocessing includes data cleaning, removing redundant information, and ensuring the accuracy and consistency of the data;

[0016] Step S23: perform knowledge extraction, including entity extraction, relationship extraction, and attribute extraction. Based on the data pre-processed in step S22, entities, attributes, and relationships between entities are extracted from the course lesson plan and course information data source to form an ontological knowledge expression.

[0017] Step S24: perform knowledge fusion, integrating the entities, attributes, and relationships between entities obtained from the course teaching plan and course information data source through knowledge extraction in step S23;

[0018] Step S25: perform knowledge construction, and establish knowledge point entities and relationships of course lesson plans and course information based on the data obtained after knowledge fusion in step S24;

[0019] Step S26: perform knowledge integration, and extract a complete course knowledge graph containing a three-level knowledge structure of chapters, sections, and knowledge points based on the knowledge point entities and relationships constructed in step S25;

[0020] Step S27, storing the course knowledge graph of the three-level knowledge structure based on the knowledge integration in step S26.

[0021] Furthermore, the multimodal speech features in step S1 include the teacher's intonation, volume, pitch, tone, rhythm, speaking speed, clarity, fluency, speech noise and teacher-student interaction feature F M ;

[0022] The multimodal features of the image include students’ attendance rate, head-up rate, expression, attention, movement and seat distribution as well as teachers’ expression, behavior and movement trajectory features F I ;

[0023] Text multimodal features include text attributes and speech text content; text attribute multimodal features include extracting teacher, student and course information from the education system data. T The multimodal features of speech text content include keyword distribution, sentiment word distribution, perplexity, and error rate. The speech text content refers to the text obtained by translating the teacher's speech through speech recognition technology. C .

[0024] Furthermore, the data collected in step S22 is preprocessed, specifically in the following steps:

[0025] Step S221: Perform character recognition on the image data using PaddleOCR, an open-source OCR tool library on the Baidu PaddlePaddle platform, and convert it into unstructured text data. The course lesson plan text data is deleted based on regular expression and string matching to form a unit syllabus, unit class teaching materials, and unit class lesson plan PPT.

[0026] Step S222: Obtain the hierarchical relationship of the course lesson plans in the order of teaching arrangement according to the regular expression. The hierarchical relationship includes:

[0027] Course syllabus: Get the chapter titles and subtitles from each unit syllabus;

[0028] Course material content: Obtain the knowledge point descriptions corresponding to the chapter structure from each unit's lesson materials;

[0029] Lesson plan PPT: Get the text information description of the corresponding knowledge points in the course syllabus and course materials from each unit lesson plan PPT;

[0030] Course resources: as supplementary materials for knowledge points, used to enrich the depth and breadth of specific knowledge points;

[0031] Subject information: determine the classification of each course and establish the relationship between subjects and courses;

[0032] Step S223: performing data cleaning based on the hierarchical relationship of the course lesson plan obtained in step S222. The data cleaning includes:

[0033] Remove redundant information: Delete the descriptions of duplicate knowledge points in the unit syllabus, unit class materials, and unit class lesson plan PPTs;

[0034] Data consistency: Ensure that data maintains consistent naming conventions and terminology across different data sources to avoid confusion in subsequent processing;

[0035] Structural verification: Check whether the hierarchical structure of the extracted chapter, section, and knowledge point data is correct, ensure that each "chapter" is appropriately divided into "sections", and each "section" has a clear definition of "knowledge points".

[0036] Furthermore, in step S23, knowledge extraction is performed, and the specific steps are as follows:

[0037] Step S231, entity extraction: Bi-LSTM-CRF model is used to automatically identify the entities of "chapter", "section", and "knowledge point" from the preprocessed text data. After the text data is subjected to entity extraction, a series of entity sets of chapter, section, and knowledge point terms are obtained. ;

[0038] Step S232, relationship extraction: extracting semantic relationships between multiple entities from the text using an open Chinese entity relationship extraction method based on dependency syntax analysis for unstructured text data, clarifying the hierarchical relationships of "chapter", "section", and "knowledge point" and linking the entities;

[0039] Step S233, attribute extraction: collecting attribute information of specific entities of the course chapter theme, section focus, and knowledge point content, and mining the relationship pattern between entity attributes and attribute values ​​from the text data through data mining methods;

[0040] The specific steps of step S231 are:

[0041] Step S2311: Encode the text data preprocessed in step S22 using word vectors and convert it into a word vector sequence. , is the first word vector, is the second word vector, is the nth word vector;

[0042] Each word vector is represented as a vector of fixed dimension. The bidirectional long short-term memory network Bi-LSTM processes the input word vector sequence through bidirectional transmission, captures context information, and generates a hidden state sequence containing context information. , is the first hidden state containing contextual information, is the second hidden state containing context information, is the nth hidden state containing context information;

[0043] Each hidden state containing contextual information is output by a forward long short-term memory network LSTM and reverse LSTM output Composition, to capture the long-term dependencies of word vector sequences; Represents the forward long short-term memory network LSTM at time step The hidden state vector at Represents the forward long short-term memory network LSTM at time step The hidden state vector at Represents the reverse long short-term memory network LSTM at time step The hidden state vector at Represents the reverse long short-term memory network LSTM at time step The hidden state vector at ;

[0044] Step S2312: The conditional random field (CRF) layer generates a hidden state sequence containing context information generated by the bidirectional long short-term memory network Bi-LSTM. Perform global annotation optimization to maximize the conditional probability of the annotation sequence, and input the hidden state sequence containing the context information After the conditional random field CRF layer, the optimal label sequence is output ,in is the entity tag corresponding to the first word, is the entity tag corresponding to the first word, To correspond to The entity tags of words are finally obtained, and the entity sets of chapter, section and knowledge point terms are obtained. ;

[0045] Step S232, an open Chinese entity relationship extraction method based on dependency syntax analysis, specifically the following steps:

[0046] Step S2321, based on the chapter, section, and knowledge point term entity set obtained in step S231 , represents the first term entity, represents the second term entity, represents the nth term entity, Extracting entity pairs using a neural network dependency parser , and Represent the i-th and j-th term entities respectively, where , determine the relationship category between entity pairs, and initially obtain the entity relationship pair set of chapter, section, and knowledge point ; Represents the first entity-relationship pair, Represents the second entity-relationship pair, Represents the nth entity-relationship pair;

[0047] Step S233, mining the relationship pattern between entity attributes and attribute values ​​from the text data by using a data mining method, specifically the following steps:

[0048] Step S2331: The course consists of many chapters and knowledge points, which are divided into vertical courses and horizontal courses;

[0049] Vertical relationship of the course: a chapter consists of multiple sections, a section contains multiple knowledge points, and the relationship between chapters and sections, and between sections and knowledge points is that of parent and child categories;

[0050] Horizontal relationship courses: The learning of later chapters, sections, and knowledge points depends on the content of earlier chapters and knowledge points, and there is a clear order of precedence. Chapters, sections, and knowledge points at the same level have a sequential relationship;

[0051] Step S2332: Based on the data pre-processed in step S22, extract the course chapter, section, knowledge point semantic connection and enrich the chapter, section, knowledge point term entity set by data mining method. Attribute information of

[0052] Step S2333: vertical relationship mining. If a "chapter" contains multiple "sections", the inclusion relationship is defined as: , Indicates the chapter, Indicates the Festival, Indicates chapter inclusion relationship, i.e. chapter Contains Festival If a "section" contains multiple "knowledge points", the inclusion relationship is defined as: ; Indicates the A knowledge point, Indicates that the section contains knowledge point relations, i.e. Section contains Knowledge points ;

[0053] Horizontal relationship mining, if Knowledge points Learning depends on Knowledge points , then the order relationship is defined as: , Indicates the knowledge point dependency; if Festival Learning depends on Festival , then the order relationship is defined as: ; If chapter Learning depends on chapter , then the order relationship is defined as: .

[0054] Furthermore, step S24 performs knowledge fusion, and the specific steps are as follows:

[0055] Step S241, entity linking: identifying entity referents from the text through entity extraction, For step S23, obtain the chapter, section, and knowledge point term entity set, use the clustering method to process multiple referents corresponding to the same entity, and classify multiple referents into the same entity class. ={ | Belong to the same entity class }, while classifying the attributes of multiple referents into the same entity object, for the same entity class , classification attributes , Representing an entity The attribute set of Represents the same entity class The attribute set is used to enhance the accuracy of the course knowledge graph data constructed by chapters, sections, and knowledge points;

[0056] Step S242, knowledge merging: Data fusion merges the unstructured text data with the structured data through knowledge extraction. , It is a set of structured data records in the educational system. Represents the first structured data record, Represents the second structured data record, Represents the mth structured data record, is a set of relations extracted from unstructured text data. represents the relationship between the first pair of term entities, represents the relationship between the second pair of term entities, Represents the relationship between the nth pair of term entities.

[0057] Furthermore, in step S25, knowledge construction is performed, and the specific steps are as follows:

[0058] According to the data of knowledge fusion in step S24, the knowledge point entities and relationships of the course teaching plan and course information are established respectively, and the course knowledge point entities are extracted from the course teaching plan. and relationships and obtain the course information entity from the course information in the educational system and relationships ;

[0059] In step S26, knowledge integration is performed, and the specific steps are as follows:

[0060] According to step S25, entities and relationships of course lesson plans and course information are extracted to construct a complete course knowledge graph containing a three-level knowledge structure of chapters, sections, and knowledge points. The semantic association relationship between course knowledge is explicitly established and represented as a set of entity-relationship triples:

[0061] ;

[0062] Among them, S is the course knowledge graph, represents the first triple, where Represented as the starting entity of the first triple, The target entity represented as the first triple, The type of relationship between the starting entity and the target entity represented as the first triple, represents the second triple, where Represented as the starting entity of the second triple, The target entity represented as the second triple, The type of relationship between the source entity and the target entity represented as the second triple, Indicates the triples, where Expressed as The starting entity of the triple, Expressed as triples of target entities, Expressed as The type of relationship between the starting entity and the target entity of a triple;

[0063] Step S27, storing the course knowledge graph of the three-level knowledge structure based on the knowledge integration in step S26, specifically:

[0064] The course knowledge graph contains a three-level knowledge structure of chapters, sections, and knowledge points, which is expressed as a set of entity-relationship triples stored in the RDF data format.

[0065] Furthermore, in step S3, the course knowledge graph constructed in step S2 is used to enhance the large language model, and intelligent classroom teaching quality evaluation is achieved based on the enhanced large language model; specifically,

[0066] Step S31, extracting classroom teaching knowledge points and classroom teaching knowledge point relationships based on the voice, image, and text multimodal features of the teacher and the classroom environment acquired in step S1;

[0067] Step S32, constructing classroom teaching quality evaluation knowledge constraints C based on classroom teaching evaluation criteria and standards and the classroom teaching knowledge points and classroom teaching knowledge point relationships extracted in step S31;

[0068] Step S33, based on the classroom multimodal features F extracted in step S1 M 、F I 、F T 、F C , the classroom teaching quality evaluation knowledge constraint C of step S32 and the course knowledge graph S of step S2 are used to construct the prompts for classroom teaching quality evaluation as the input of the large language model;

[0069] The prompts for classroom teaching quality evaluation are set as follows:

[0070] ;

[0071] in, Indicates the generated evaluation result, BOS and EOS are the start and end identifiers respectively, and SEP is the separator.

[0072] Furthermore, in step S31, classroom teaching knowledge points and classroom teaching knowledge point relationships are extracted, and the specific steps are:

[0073] Step S311: Use the large model GPT-4 to process and segment the classroom speech text to obtain a sentence token sequence, that is, to subdivide the classroom speech text into multiple independent sentence data units, each sentence data unit represents an element in the original text, and the sentence token sequence formed is represented as ,in represents the first statement token, represents the second statement token, represents the i-th statement token, represents the mth statement token, where m is the total number of statement tokens, A category representing a sequence of sentence tokens;

[0074] Step S312, extract the knowledge points in the classroom voice text as classroom teaching knowledge points based on the sentence token sequence obtained by segmenting the classroom voice text, and then perform joint extraction of entities and relationships on the obtained classroom teaching knowledge points to obtain the logical relationship between the classroom teaching knowledge points, which is expressed in the form of triples as the classroom teaching knowledge point relationship.

[0075] Furthermore, in step S312, the specific steps are:

[0076] Step S3121, using adaptive sampling technology to randomly sample the sentence token sequence, Randomly select candidate sequence fragments from ,in represents the width of the fragment, Represents a sequence of tokens from a statement The selected statement tokens, Represents a sequence of tokens from a statement The selected statement tokens, Represents a sequence of tokens from a statement The selected sentence tokens, for the i-th sentence token in the selected candidate sequence segment Evaluate the relevance of each token not only based on randomness but also based on information gain;

[0077] ;

[0078] in, Represents a statement token The information gain of is the entire candidate sequence fragment The entropy of Indicates that in a given statement token Afterwards, the sequence fragment The conditional entropy of Indicates that in a given statement token Afterwards, the sequence fragment The probability value of

[0079] Step S3122: Use convolutional neural network (CNN) to process the selected candidate sequence segments. , for the input sequence The output is obtained through convolution operation. ,in represents the weight matrix of the convolution kernel, is the bias vector;

[0080] Use global average pooling to extract and simplify the output of the convolution layer to obtain the feature vector , represents the result vector of global average pooling, Represents the output vector of the convolutional layer, and the extracted feature vector is combined with the trained weight vector Merge to generate knowledge point vector , followed by the statement token sequence category With knowledge point vector Splicing to form a comprehensive context knowledge point entity vector , after the logistic regression classifier classifies the knowledge point entity vector Make type predictions ,in is the weight parameter of the logistic regression classifier, which completes the knowledge point recognition and obtains the classroom teaching knowledge points. These are the knowledge points of classroom teaching;

[0081] Step S3123: Use the self-attention mechanism to process two adjacent candidate sequence segments and , the two candidate sequence fragments and The self-attention output of is merged and the information is integrated through the linear layer to obtain the overall context representation:

[0082] ;

[0083] in, and Represents two adjacent candidate sequence fragments and Self-attention output, and are the weights and biases of the linear layer;

[0084] Step S3124, combining the method of extracting knowledge point vectors in step S3122 to obtain the knowledge point vectors of adjacent sequence segments and With context representation Splicing to form a complete knowledge point entity relationship pair , That is the relationship between classroom teaching knowledge points.

[0085] Furthermore, in step S32, the classroom teaching evaluation criteria and standards and the classroom teaching knowledge points extracted in step S31 are used to evaluate the classroom teaching knowledge points. Relationship with classroom teaching knowledge points , construct the knowledge constraint C for classroom teaching quality evaluation; the specific steps are:

[0086] Step S321: According to the classroom teaching evaluation criteria, the teacher's classroom teaching quality evaluation should be evaluated from multiple aspects: teaching attitude, teaching content, teaching methods, teaching effect, classroom atmosphere, learning status, and learning effect, and the evaluation basis of each item is different;

[0087] Step S322: Generate a teaching evaluation description that complies with the classroom teaching evaluation criteria and set the classroom teaching quality evaluation knowledge constraints as follows:

[0088] C={"{Teaching attitude: [Teacher's facial expression characteristics, Teacher's voice characteristics]}"

[0089] "{Teaching content: [Classroom teaching knowledge points, relationships between classroom teaching knowledge points]}"

[0090] "{Teaching Method: [Teacher Behavior Data]}"

[0091] "{Teaching effectiveness: [Teacher's language fluency, Teacher's language clarity, Student's attention data, Student's facial expression data]}"

[0092] "{Learning status: [student behavior data, student expression data]}"

[0093] "{Learning outcomes: [student attendance rate, head-raising rate, distribution of sentiment words]}"

[0094] "{Classroom atmosphere: [Classroom environment characteristics]}"}.

[0095] The beneficial effects of the present invention are:

[0096] (1) Enhance the large language model by combining classroom teaching quality evaluation knowledge and course knowledge graph, and guide the large language model to generate an evaluation result description that conforms to the facts of classroom teaching quality evaluation based on the semantic association between classroom teaching quality evaluation knowledge and course knowledge graph.

[0097] (2) The curriculum knowledge graph-enhanced large language model can solve the problems of "hallucination", limited reasoning ability and poor interpretability when the large language model is applied to tasks in the field of education, thereby enabling the model to generate more reasonable and more consistent descriptions of classroom teaching quality evaluation facts;

[0098] (3) Based on the multimodal features of teachers, students and classroom environment such as speech, images, and text, it is possible to comprehensively characterize teachers' classroom teaching. Using these data to drive the large language model to realize course teaching quality evaluation can truly realize intelligent and scientific classroom teaching quality evaluation and improve the efficiency and intelligence level of teaching management. However, the introduction of large language models for classroom teaching quality evaluation has the problems of "hallucination", limited reasoning ability and poor interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 It is a technical scheme diagram of the method of the present invention. DETAILED DESCRIPTION

[0100] In order to solve the above technical problems, the present invention introduces a large language model to realize intelligent classroom teaching quality evaluation. Through the strong reasoning ability and explanatory characteristics of the classroom knowledge graph, the "hallucination" phenomenon that exists when the large pre-trained language model is applied to specific field tasks is solved. A classroom teaching quality evaluation method based on a large language model enhanced by the course knowledge graph is provided to realize intelligent and scientific classroom teaching quality evaluation, improve the efficiency and intelligence level of teaching management, and can be widely promoted to teaching management platforms and online teaching platform applications to promote the development of smart education.

[0101] like Figure 1 As shown, the classroom teaching quality evaluation method based on the large language model of the present invention mainly includes:

[0102] Step S1: Based on classroom monitoring video data and educational system data, artificial intelligence technology is used to extract multimodal features of voice, image, and text of teachers, students, and classroom environment;

[0103] Step S2: constructing a course knowledge graph based on the course lesson plan and course information;

[0104] Step S3, enhancing the large language model through the course knowledge graph constructed in step S2, and realizing intelligent classroom teaching quality evaluation based on the enhanced large language model;

[0105] The specific steps of step S2 are:

[0106] Step S21: Collect course lesson plans and course information data from the educational system. The data is divided into image data and text data. The data specifically includes course syllabus, course teaching materials, course lesson plan PPT, course resources, and subject information.

[0107] Step S22, preprocessing the data collected in step S21, the preprocessing includes data cleaning, removing redundant information, and ensuring the accuracy and consistency of the data;

[0108] Step S23: perform knowledge extraction, including entity extraction, relationship extraction, and attribute extraction. Based on the data pre-processed in step S22, entities, attributes, and relationships between entities are extracted from the course lesson plan and course information data source to form an ontological knowledge expression.

[0109] Step S24: perform knowledge fusion, integrating the entities, attributes, and relationships between entities obtained from the course teaching plan and course information data source through knowledge extraction in step S23;

[0110] Step S25: perform knowledge construction, and establish knowledge point entities and relationships of course lesson plans and course information based on the data obtained after knowledge fusion in step S24;

[0111] Step S26: perform knowledge integration, and extract a complete course knowledge graph containing a three-level knowledge structure of chapters, sections, and knowledge points based on the knowledge point entities and relationships constructed in step S25;

[0112] Step S27, storing the course knowledge graph of the three-level knowledge structure based on the knowledge integration in step S26.

[0113] Furthermore, the multimodal speech features in step S1 include the teacher's intonation, volume, pitch, tone, rhythm, speaking speed, clarity, fluency, speech noise and teacher-student interaction feature F M ;

[0114] The multimodal features of the image include students’ attendance rate, head-up rate, expression, attention, movement and seat distribution as well as teachers’ expression, behavior and movement trajectory features F I ;

[0115] Text multimodal features include text attributes and speech text content; text attribute multimodal features include extracting teacher, student and course information from the education system data. T The multimodal features of speech text content include keyword distribution, sentiment word distribution, perplexity, and error rate. The speech text content refers to the text obtained by translating the teacher's speech through speech recognition technology. C .

[0116] Furthermore, the data collected in step S22 is preprocessed, specifically in the following steps:

[0117] Step S221: Perform character recognition on the image data using PaddleOCR, an open-source OCR tool library on the Baidu PaddlePaddle platform, and convert it into unstructured text data. The course lesson plan text data is deleted based on regular expression and string matching to form a unit syllabus, unit class teaching materials, and unit class lesson plan PPT.

[0118] Step S222: Obtain the hierarchical relationship of the course lesson plans in the order of teaching arrangement according to the regular expression. The hierarchical relationship includes:

[0119] Course syllabus: Get the chapter titles and subtitles from each unit syllabus;

[0120] Course material content: Obtain the knowledge point descriptions corresponding to the chapter structure from each unit's lesson materials;

[0121] Lesson plan PPT: Get the text information description of the corresponding knowledge points in the course syllabus and course materials from each unit lesson plan PPT;

[0122] Course resources: as supplementary materials for knowledge points, used to enrich the depth and breadth of specific knowledge points;

[0123] Subject information: determine the classification of each course and establish the relationship between subjects and courses;

[0124] Step S223: performing data cleaning based on the hierarchical relationship of the course lesson plan obtained in step S222. The data cleaning includes:

[0125] Remove redundant information: Delete the descriptions of duplicate knowledge points in the unit syllabus, unit class materials, and unit class lesson plan PPTs;

[0126] Data consistency: Ensure that data maintains consistent naming conventions and terminology across different data sources to avoid confusion in subsequent processing;

[0127] Structural verification: Check whether the hierarchical structure of the extracted chapter, section, and knowledge point data is correct, ensure that each "chapter" is appropriately divided into "sections", and each "section" has a clear definition of "knowledge points".

[0128] Furthermore, in step S23, knowledge extraction is performed, and the specific steps are as follows:

[0129] Step S231, entity extraction: Bi-LSTM-CRF model is used to automatically identify the entities of "chapter", "section", and "knowledge point" from the preprocessed text data. After the text data is subjected to entity extraction, a series of entity sets of chapter, section, and knowledge point terms are obtained. ;

[0130] Step S232, relationship extraction: extracting semantic relationships between multiple entities from the text using an open Chinese entity relationship extraction method based on dependency syntax analysis for unstructured text data, clarifying the hierarchical relationships of "chapter", "section", and "knowledge point" and linking the entities;

[0131] Step S233, attribute extraction: collecting attribute information of specific entities of the course chapter theme, section focus, and knowledge point content, and mining the relationship pattern between entity attributes and attribute values ​​from the text data through data mining methods;

[0132] The specific steps of step S231 are:

[0133] Step S2311: Encode the text data preprocessed in step S22 using word vectors and convert it into a word vector sequence. , is the first word vector, is the second word vector, is the nth word vector;

[0134] Each word vector is represented as a vector of fixed dimension. The bidirectional long short-term memory network Bi-LSTM processes the input word vector sequence through bidirectional transmission, captures context information, and generates a hidden state sequence containing context information. , is the first hidden state containing contextual information, is the second hidden state containing context information, is the nth hidden state containing context information;

[0135] Each hidden state containing contextual information is output by a forward long short-term memory network LSTM and reverse LSTM output Composition, to capture the long-term dependencies of word vector sequences; Represents the forward long short-term memory network LSTM at time step The hidden state vector at Represents the forward long short-term memory network LSTM at time step The hidden state vector at Represents the reverse long short-term memory network LSTM at time step The hidden state vector at Represents the reverse long short-term memory network LSTM at time step The hidden state vector at ;

[0136] Step S2312: The conditional random field (CRF) layer generates a hidden state sequence containing context information generated by the bidirectional long short-term memory network Bi-LSTM. Perform global annotation optimization to maximize the conditional probability of the annotation sequence, and input the hidden state sequence containing the context information After the conditional random field CRF layer, the optimal label sequence is output ,in is the entity tag corresponding to the first word, is the entity tag corresponding to the first word, To correspond to The entity tags of words are finally obtained, and the entity sets of chapter, section and knowledge point terms are obtained. ;

[0137] Step S232, an open Chinese entity relationship extraction method based on dependency syntax analysis, specifically the following steps:

[0138] Step S2321, based on the chapter, section, and knowledge point term entity set obtained in step S231 , represents the first term entity, represents the second term entity, represents the nth term entity, Extracting entity pairs using a neural network dependency parser , and Represent the i-th and j-th term entities respectively, where , determine the relationship category between entity pairs, and initially obtain the entity relationship pair set of chapter, section, and knowledge point ; Represents the first entity-relationship pair, Represents the second entity-relationship pair, Represents the nth entity-relationship pair;

[0139] Step S233, mining the relationship pattern between entity attributes and attribute values ​​from the text data by using a data mining method, specifically the following steps:

[0140] Step S2331: The course consists of many chapters and knowledge points, which are divided into vertical courses and horizontal courses;

[0141] Vertical relationship of the course: a chapter consists of multiple sections, a section contains multiple knowledge points, and the relationship between chapters and sections, and between sections and knowledge points is that of parent and child categories;

[0142] Horizontal relationship courses: The learning of later chapters, sections, and knowledge points depends on the content of earlier chapters and knowledge points, and there is a clear order of precedence. Chapters, sections, and knowledge points at the same level have a sequential relationship;

[0143] Step S2332: Based on the data pre-processed in step S22, extract the course chapter, section, knowledge point semantic connection and enrich the chapter, section, knowledge point term entity set by data mining method. Attribute information of

[0144] Step S2333: vertical relationship mining. If a "chapter" contains multiple "sections", the inclusion relationship is defined as: , Indicates the chapter, Indicates the Festival, Indicates chapter inclusion relationship, i.e. chapter Contains Festival If a "section" contains multiple "knowledge points", the inclusion relationship is defined as: ; Indicates the A knowledge point, Indicates that the section contains knowledge point relations, i.e. Section contains Knowledge points ;

[0145] Horizontal relationship mining, if Knowledge points Learning depends on Knowledge points , then the order relationship is defined as: , Indicates the knowledge point dependency; if Festival Learning depends on Festival , then the order relationship is defined as: ; If chapter Learning depends on chapter , then the order relationship is defined as: .

[0146] Furthermore, step S24 performs knowledge fusion, and the specific steps are as follows:

[0147] Step S241, entity linking: identifying entity referents from the text through entity extraction, For step S23, obtain the chapter, section, and knowledge point term entity set, use the clustering method to process multiple referents corresponding to the same entity, and classify multiple referents into the same entity class. ={ | Belong to the same entity class }, while classifying the attributes of multiple referents into the same entity object, for the same entity class , classification attributes , Representing an entity The attribute set of Represents the same entity class The attribute set is used to enhance the accuracy of the course knowledge graph data constructed by chapters, sections, and knowledge points;

[0148] Step S242, knowledge merging: Data fusion merges the unstructured text data with the structured data through knowledge extraction. , It is a set of structured data records in the educational system. Represents the first structured data record, Represents the second structured data record, Represents the mth structured data record, is a set of relations extracted from unstructured text data. represents the relationship between the first pair of term entities, represents the relationship between the second pair of term entities, Represents the relationship between the nth pair of term entities.

[0149] Furthermore, in step S25, knowledge construction is performed, and the specific steps are as follows:

[0150] According to the data of knowledge fusion in step S24, the knowledge point entities and relationships of the course teaching plan and course information are established respectively, and the course knowledge point entities are extracted from the course teaching plan. and relationships and obtain the course information entity from the course information in the educational system and relationships ;

[0151] In step S26, knowledge integration is performed, and the specific steps are as follows:

[0152] According to step S25, entities and relationships of course lesson plans and course information are extracted to construct a complete course knowledge graph containing a three-level knowledge structure of chapters, sections, and knowledge points. The semantic association relationship between course knowledge is explicitly established and represented as a set of entity-relationship triples:

[0153] ;

[0154] Among them, S is the course knowledge graph, represents the first triple, where Represented as the starting entity of the first triple, The target entity represented as the first triple, The type of relationship between the starting entity and the target entity represented as the first triple, represents the second triple, where Represented as the starting entity of the second triple, The target entity represented as the second triple, The type of relationship between the source entity and the target entity represented as the second triple, Indicates the triples, where Expressed as The starting entity of the triple, Expressed as triples of target entities, Expressed as The type of relationship between the starting entity and the target entity of a triple;

[0155] Step S27, storing the course knowledge graph of the three-level knowledge structure based on the knowledge integration in step S26, specifically:

[0156] The course knowledge graph contains a three-level knowledge structure of chapters, sections, and knowledge points, which is expressed as a set of entity-relationship triples stored in the RDF data format.

[0157] Furthermore, in step S3, the course knowledge graph constructed in step S2 is used to enhance the large language model, and intelligent classroom teaching quality evaluation is achieved based on the enhanced large language model; specifically,

[0158] Step S31, extracting classroom teaching knowledge points and classroom teaching knowledge point relationships based on the voice, image, and text multimodal features of the teacher and the classroom environment acquired in step S1;

[0159] Step S32, constructing classroom teaching quality evaluation knowledge constraints C based on classroom teaching evaluation criteria and standards and the classroom teaching knowledge points and classroom teaching knowledge point relationships extracted in step S31;

[0160] Step S33, based on the classroom multimodal features F extracted in step S1 M 、F I 、F T 、F C , the classroom teaching quality evaluation knowledge constraint C of step S32 and the course knowledge graph S of step S2 are used to construct the prompts for classroom teaching quality evaluation as the input of the large language model;

[0161] The prompts for classroom teaching quality evaluation are set as follows:

[0162] ;

[0163] in, Indicates the generated evaluation result, BOS and EOS are the start and end identifiers respectively, and SEP is the separator.

[0164] Furthermore, in step S31, classroom teaching knowledge points and classroom teaching knowledge point relationships are extracted, and the specific steps are:

[0165] Step S311: Use the large model GPT-4 to process and segment the classroom speech text to obtain a sentence token sequence, that is, to subdivide the classroom speech text into multiple independent sentence data units, each sentence data unit represents an element in the original text, and the sentence token sequence formed is represented as ,in represents the first statement token, represents the second statement token, represents the i-th statement token, represents the mth statement token, where m is the total number of statement tokens, A category representing a sequence of sentence tokens;

[0166] Step S312, extract the knowledge points in the classroom voice text as classroom teaching knowledge points based on the sentence token sequence obtained by segmenting the classroom voice text, and then perform joint extraction of entities and relationships on the obtained classroom teaching knowledge points to obtain the logical relationship between the classroom teaching knowledge points, which is expressed in the form of triples as the classroom teaching knowledge point relationship.

[0167] Furthermore, in step S312, the specific steps are:

[0168] Step S3121, using adaptive sampling technology to randomly sample the sentence token sequence, Randomly select candidate sequence fragments from ,in represents the width of the fragment, Represents a sequence of tokens from a statement The selected statement tokens, Represents a sequence of tokens from a statement The selected statement tokens, Represents a sequence of tokens from a statement The selected sentence tokens, for the i-th sentence token in the selected candidate sequence segment Evaluate the relevance of each token not only based on randomness but also based on information gain;

[0169] ;

[0170] in, Represents a statement token The information gain of is the entire candidate sequence fragment The entropy of Indicates that in a given statement token Afterwards, the sequence fragment The conditional entropy of Indicates that in a given statement token Afterwards, the sequence fragment The probability value of

[0171] Step S3122: Use convolutional neural network (CNN) to process the selected candidate sequence segments. , for the input sequence The output is obtained through convolution operation. ,in represents the weight matrix of the convolution kernel, is the bias vector;

[0172] Use global average pooling to extract and simplify the output of the convolution layer to obtain the feature vector , represents the result vector of global average pooling, Represents the output vector of the convolutional layer, and the extracted feature vector is combined with the trained weight vector Merge to generate knowledge point vector , followed by the statement token sequence category With knowledge point vector Splicing to form a comprehensive context knowledge point entity vector , after the logistic regression classifier classifies the knowledge point entity vector Make type predictions ,in is the weight parameter of the logistic regression classifier, which completes the knowledge point recognition and obtains the classroom teaching knowledge points. These are the knowledge points of classroom teaching;

[0173] Step S3123: Use the self-attention mechanism to process two adjacent candidate sequence segments and , the two candidate sequence fragments and The self-attention output of is merged and the information is integrated through the linear layer to obtain the overall context representation:

[0174] ;

[0175] in, and Represents two adjacent candidate sequence fragments and Self-attention output, and are the weights and biases of the linear layer;

[0176] Step S3124, combining the method of extracting knowledge point vectors in step S3122 to obtain the knowledge point vectors of adjacent sequence segments and With context representation Splicing to form a complete knowledge point entity relationship pair , That is the relationship between classroom teaching knowledge points.

[0177] Furthermore, in step S32, the classroom teaching evaluation criteria and standards and the classroom teaching knowledge points extracted in step S31 are used to evaluate the classroom teaching knowledge points. Relationship with classroom teaching knowledge points , construct the knowledge constraint C for classroom teaching quality evaluation; the specific steps are:

[0178] Step S321: According to the classroom teaching evaluation criteria, the teacher's classroom teaching quality evaluation should be evaluated from multiple aspects: teaching attitude, teaching content, teaching methods, teaching effect, classroom atmosphere, learning status, and learning effect, and the evaluation basis of each item is different;

[0179] Step S322: Generate a teaching evaluation description that complies with the classroom teaching evaluation criteria and set the classroom teaching quality evaluation knowledge constraints as follows:

[0180] C={"{Teaching attitude: [Teacher's facial expression characteristics, Teacher's voice characteristics]}"

[0181] "{Teaching content: [Classroom teaching knowledge points, relationships between classroom teaching knowledge points]}"

[0182] "{Teaching Method: [Teacher Behavior Data]}"

[0183] "{Teaching effectiveness: [Teacher's language fluency, Teacher's language clarity, Student's attention data, Student's facial expression data]}"

[0184] "{Learning status: [student behavior data, student expression data]}"

[0185] "{Learning outcomes: [student attendance rate, head-raising rate, distribution of sentiment words]}"

[0186] "{Classroom atmosphere: [Classroom environment characteristics]}"}.

[0187] Among them, in step S33, according to the classroom multimodal features F extracted in step S1 M 、F I 、F T 、F C , the classroom teaching quality evaluation knowledge constraint C of step S32 and the course knowledge graph S of step S2 are used to construct the prompts for classroom teaching quality evaluation as the input of the large language model;

[0188] The prompts for classroom teaching quality evaluation are set as follows:

[0189] #CharacterYou are a professional teacher reviewer with data-driven classroom review capabilities. You can analyze given educational data to conduct in-depth reviews of teachers' teaching quality and provide specific recommendations based on your findings.

[0190] Skills

[0191] ###Skill1: Evaluation based on teaching attitude;

[0192] Analyze the provided data (teacher's facial expression data, teacher's voice characteristics) and make an objective evaluation of the teacher's teaching attitude.

[0193] ###Skill2: Assessment based on teaching content;

[0194] Analyze the provided data (teaching content knowledge points, teaching content knowledge points relationship), and conduct an in-depth review of the teacher's teaching content.

[0195] ###Skill3: Evaluation based on teaching methods;

[0196] Analyze the data provided (teacher behavior data) and conduct a comprehensive evaluation of the teacher's teaching methods.

[0197] ###Skill4: Evaluation based on teaching effectiveness;

[0198] Analyze the provided data (teacher's language fluency, teacher's language clarity, student attention data, student expression data) to conduct a comprehensive evaluation of the teacher's teaching effectiveness.

[0199] ###Skill5: Assessment based on learning status;

[0200] Analyze the provided data (student behavior data, student expression data), evaluate students' learning status, and reflect the teacher's teaching quality.

[0201] ###Skill6: Assessment based on learning outcomes;

[0202] Analyze the provided data (student attendance rate, head-raising rate, and emotional words used in class) to comprehensively evaluate students' learning outcomes.

[0203] ###Skill7: Evaluation based on classroom atmosphere;

[0204] Analyze the characteristics of the classroom environment and accurately evaluate the classroom atmosphere.

[0205] ###Skill8:Evaluation suggestions;

[0206] In response to specific issues observed in each evaluation dimension, clear and feasible suggestions are provided, aiming to solve specific problems and help promote the continuous improvement of teaching quality.

[0207] ###Skill9: Overall teaching quality rating;

[0208] Based on the above evaluation dimensions, give the teacher's overall teaching quality rating (excellent, good, medium), and briefly explain the basis for the rating.

[0209] ###Skill10: Summative evaluation;

[0210] It reflects the overall performance of the teacher by providing an overall evaluation of the teacher and making suggestions for his / her future development.

[0211] ##Constraints; Please strictly follow the following format for your review:

[0212] Teaching attitude

[0213] Description: ...

[0214] Evaluation suggestions: ...

[0215] Teaching content

[0216] Description: ... ......

[0218] Overall teaching quality rating

[0219] Summative evaluation

[0220] During the evaluation process, please maintain objectivity and ensure that the analysis and recommendations are based on concrete data while also encouraging continuous growth and improvement for teachers.

[0221] Please note that all evaluations and recommendations should be based on the educational data provided and should not include any personal subjective opinions.

[0222] By enhancing the large language model with classroom teaching quality evaluation knowledge and course knowledge graph, the large language model is guided to generate an evaluation result description that conforms to the facts of classroom teaching quality evaluation based on the semantic association between classroom teaching quality evaluation knowledge and course knowledge graph.

Claims

1. A classroom teaching quality evaluation method based on a large language model, characterized by: include: Step S1, based on classroom monitoring video data and teaching system data, using artificial intelligence technology to extract voice, image, and text multimodal features of teachers, students, and classroom environment; Step S2, constructing a course knowledge graph based on the course lesson plan and course information; Step S3, enhancing the big language model through the course knowledge graph constructed in step S2, and realizing intelligent classroom teaching quality evaluation based on the enhanced big language model; The multimodal speech features in step S1 include the teacher's intonation, volume, pitch, tone, rhythm, speaking speed, clarity, fluency, speech noise, and teacher-student interaction feature F M ; The multimodal features of the image include students’ attendance rate, head-up rate, expression, attention, movement, and seat distribution, as well as teachers’ expression, behavior, and movement trajectory features. I ; Text multimodal features include text attributes and speech text content; text attribute multimodal features include extracting teacher, student and course information from the education system data. T ; The multimodal features of speech text content include keyword distribution, sentiment word distribution, confusion, and error rate; the speech text content refers to the text obtained by translating the teacher's speech through speech recognition technology F C ; The specific steps of step S2 are: Step S21, collecting course lesson plans and course information data from the education system data, the data is divided into image data and text data, and the data specifically covers course syllabus, course teaching materials, course lesson plan PPT, course resources, and subject information; Step S22, preprocessing the data collected in step S21, the preprocessing includes data cleaning, removing redundant information, and ensuring the accuracy and consistency of the data; Step S23, performing knowledge extraction, including entity extraction, relationship extraction and attribute extraction, extracting entities, attributes and the relationships between entities from the course teaching plan and course information data source according to the data pre-processed in step S22, to form an ontological knowledge expression; Step S24, performing knowledge fusion, integrating the entities, attributes and relationships between entities obtained from the course teaching plan and course information data source through knowledge extraction in step S23; Step S25, performing knowledge construction, respectively establishing the knowledge point entities and relationships of the course lesson plan and course information according to the data after knowledge fusion in step S24; Step S26, performing knowledge integration, extracting a complete course knowledge graph including a three-level knowledge structure of chapters, sections, and knowledge points based on the knowledge point entities and relationships constructed in step S25; Step S27, storing the course knowledge graph of the three-level knowledge structure according to the knowledge integration in step S26; In step S3, the course knowledge graph constructed in step S2 is used to enhance the big language model, and intelligent classroom teaching quality evaluation is realized based on the enhanced big language model; specifically: Step S31, extracting classroom teaching knowledge points and classroom teaching knowledge point relationships based on the voice, image, and text multimodal features of the teacher and the classroom environment acquired in step S1; Step S32, constructing classroom teaching quality evaluation knowledge constraints C according to classroom teaching evaluation criteria and specifications and classroom teaching knowledge points and classroom teaching knowledge point relationships extracted in step S31; Step S33, based on the classroom multimodal features F extracted in step S1 M 、F I 、F T 、F C , the classroom teaching quality evaluation knowledge constraint C of step S32 and the course knowledge graph S of step S2 construct the prompt of classroom teaching quality evaluation as the input of the large language model; The prompts for classroom teaching quality evaluation are set as follows: ； in, Indicates the generated evaluation result, BOS and EOS are the start and end identifiers respectively, and SEP is the separator.

2. The classroom teaching quality evaluation method based on a large language model according to claim 1 is characterized by: The data collected in step S22 is preprocessed, and the specific steps are as follows: Step S221, perform character recognition on the image data through Baidu PaddleOCR, an open source OCR tool library, and convert it into unstructured text data. Delete the course lesson plan text data according to regular expression and string matching to form a unit teaching outline, unit class teaching materials, and unit class lesson plan PPT; Step S222, obtaining the hierarchical relationship of the course lesson plans according to the regular expression and the teaching arrangement sequence, the hierarchical relationship includes: Course Syllabus: Get the chapter titles and subtitles from each unit syllabus content; Course material content: obtain the knowledge point description corresponding to the chapter structure from each unit lesson material; Course lesson plan PPT: Get the text information description of the corresponding knowledge points in the course syllabus and course materials from each unit lesson plan PPT; Course resources: as supplementary materials for knowledge points, used to enrich the depth and breadth of the content of specific knowledge points; Subject information: determine the classification of each course and establish the relationship between subjects and courses; Step S223, performing data cleaning according to the hierarchical relationship of the course lesson plan obtained in step S222, the data cleaning includes: Remove redundant information: Delete the description of repeatedly acquired knowledge points in the unit syllabus, unit class materials and unit class lesson plan PPT; Data consistency: Ensure that data is consistent in naming conventions and terminology across different data sources to avoid confusion in subsequent processing; Structural verification: Check whether the hierarchical structure of the extracted chapter, section, and knowledge point data is correct, ensure that each "chapter" is appropriately divided into "sections", and each "section" has a clear definition of "knowledge points".

3. The classroom teaching quality evaluation method based on a large language model according to claim 2 is characterized by: In step S23, knowledge extraction is performed, and the specific steps are as follows: Step S231, entity extraction: using the Bi-LSTM-CRF model to automatically identify the "chapter", "section", and "knowledge point" entities from the preprocessed text data. After the text data is subjected to entity extraction, a series of chapter, section, and knowledge point term entity sets are obtained. ; Step S232, relationship extraction: extracting semantic relationships between multiple entities from text for unstructured text data through an open Chinese entity relationship extraction method based on dependency syntax analysis, clarifying the hierarchical relationship of "chapter", "section", and "knowledge point" and linking the entities together; Step S233, attribute extraction: collecting the attribute information of the subject of the course chapter, the key points of the section, and the knowledge point content entity, and mining the relationship pattern between the entity attributes and the attribute values ​​from the text data through the data mining method; The specific steps of step S231 are: Step S2311, the text data preprocessed in step S22 is encoded using word vectors and converted into a word vector sequence , is the first word vector, is the second word vector, is the nth word vector; Each word vector is represented as a vector of fixed dimension. The bidirectional long short-term memory network Bi-LSTM processes the input word vector sequence through bidirectional transmission, captures context information, and generates a hidden state sequence containing context information. , is the first hidden state that contains context information, is the second hidden state containing context information, is the nth hidden state containing context information; Each hidden state containing context information is output by a forward long short-term memory network LSTM and the reverse LSTM output Composition, to capture the long-term dependencies of word vector sequences; Represents the forward long short-term memory network LSTM at time step The hidden state vector at Represents the forward long short-term memory network LSTM at time step The hidden state vector at Represents the reverse long short-term memory network LSTM at time step The hidden state vector at Represents the reverse long short-term memory network LSTM at time step The hidden state vector at ; Step S2312: The conditional random field (CRF) layer generates a hidden state sequence containing context information generated by the bidirectional long short-term memory network Bi-LSTM. Perform global annotation optimization to maximize the conditional probability of the annotation sequence, and input the hidden state sequence containing the context information After the conditional random field CRF layer, the optimal label sequence is output ,in is the entity tag corresponding to the first word, is the entity tag corresponding to the first word, To correspond to The entity tags of words are finally obtained, and the entity sets of chapter, section and knowledge point terms are obtained. ; Step S232, an open Chinese entity relationship extraction method based on dependency syntax analysis, the specific steps are: Step S2321, based on the chapter, section, and knowledge point term entity set obtained in step S231 , represents the first term entity, represents the second term entity, represents the nth term entity, Extracting entity pairs using a neural network dependency parser , and Represent the i-th and j-th term entities respectively, where , determine the relationship between entity pairs, and initially obtain the entity relationship pair set of chapters, sections, and knowledge points ; represents the first entity-relationship pair, Represents the second entity-relationship pair, Represents the nth entity-relationship pair; Step S233, mining the relationship pattern between entity attributes and attribute values ​​from the text data by using a data mining method, the specific steps are: Step S2331, the course consists of many chapters and knowledge points, which are divided into vertical courses and horizontal courses; Courses in vertical relationship: a chapter consists of multiple sections, a section contains multiple knowledge points, and the relationship between chapters and sections, and between sections and knowledge points is the parent-child relationship; Courses in horizontal relationship: The learning of later chapters, sections, and knowledge points depends on the content of previous chapters and knowledge points, and there is a clear sequence. Chapters, sections, and knowledge points at the same level are in a sequential relationship; Step S2332, based on the data pre-processed in step S22, extracting the course chapter, section, knowledge point semantically rich chapter, section, knowledge point term entity set by data mining method Attribute information of Step S2333, vertical relationship mining, if a "chapter" contains multiple "sections", the inclusion relationship is defined as: , Indicates chapter, Indicates Festival, Indicates chapter inclusion relationship, that is, chapter Contains Festival , if a "section" contains multiple "knowledge points", the inclusion relationship is defined as: ; Indicates Knowledge points, Indicates that the node contains the knowledge point relationship, that is, Section contains Knowledge Points ; Horizontal relationship mining, if Knowledge Points The learning depends on Knowledge Points , then the order relationship is defined as: , Indicates the knowledge point dependency; if Festival The learning depends on Festival , then the order relationship is defined as: ; If chapter The learning depends on chapter , then the order relationship is defined as: .

4. The classroom teaching quality evaluation method based on a large language model according to claim 3 is characterized by: Step S24 performs knowledge fusion, and the specific steps are as follows: Step S241, entity linking: identifying entity referents from the text through entity extraction, Step S23 obtains the chapter, section, and knowledge point term entity set, uses a clustering method to process multiple referents corresponding to the same entity, and classifies multiple referents into the same entity class. ={ | Belong to the same entity class }, and classify the attributes of multiple referents into the same entity object. , classification attributes , Representing Entities The attribute set of Represents the same entity class A collection of attributes to enhance the accuracy of course knowledge graph data constructed by chapters, sections, and knowledge points; Step S242, knowledge merging: Data fusion merges the unstructured text data with the structured data through knowledge extraction. , It is a set of structured data records in the educational system. Represents the first structured data record, Represents the second structured data record, Represents the mth structured data record, is a set of relations extracted from unstructured text data. represents the relationship between the first pair of term entities, represents the relationship between the second pair of term entities, Represents the relationship between the nth pair of term entities.

5. The classroom teaching quality evaluation method based on a large language model according to claim 4 is characterized by: In step S25, knowledge construction is performed, and the specific steps are as follows: According to the data of knowledge fusion in step S24, knowledge point entities and relationships of course lesson plan and course information are established respectively, and course knowledge point entities are extracted from course lesson plan. and relationship , and obtain the course information entity from the course information in the academic affairs system and relationship ; In step S26, knowledge integration is performed, and the specific steps are as follows: According to step S25, entities and relationships of course lesson plans and course information are extracted to construct a complete course knowledge graph containing a three-level knowledge structure of chapters, sections, and knowledge points, and semantic association relationships between course knowledge are explicitly established, which are represented as a set of entity-relationship triples: ； Among them, S is the course knowledge graph, represents the first triple, where Represents the starting entity of the first triple, is represented as the target entity of the first triple, The type of relationship between the starting entity and the target entity represented by the first triple, represents the second triple, where Represented as the starting entity of the second triple, The target entity represented as the second triple, The type of relationship between the source entity and the target entity represented as the second triple, Indicates triples, where Expressed as The starting entity of the triple, Expressed as triples of target entities, Expressed as The type of relationship between the starting entity and the target entity of a triple; Step S27, storing the course knowledge graph of the three-level knowledge structure integrated according to step S26, specifically: The course knowledge graph contains a three-level knowledge structure of chapters, sections, and knowledge points, which is expressed as a set of entity-relationship triples stored in the RDF data format.

6. The classroom teaching quality evaluation method based on a large language model according to claim 5 is characterized by: In step S31, classroom teaching knowledge points and classroom teaching knowledge point relationships are extracted, and the specific steps are: Step S311, use the large model GPT-4 to process and segment the classroom speech text to obtain a sentence token sequence, that is, to subdivide the classroom speech text into multiple independent sentence data units, each sentence data unit represents an element in the original text, and the formed sentence token sequence is expressed as ,in represents the first statement token, represents the second statement token, represents the i-th statement token, represents the mth statement token, where m is the total number of statement tokens, Represents the category of a sequence of sentence tokens; Step S312, extract the knowledge points in the classroom voice text as classroom teaching knowledge points according to the sentence token sequence obtained by segmenting the classroom voice text, and then perform joint extraction of entities and relationships on the acquired classroom teaching knowledge points to obtain the logical relationship between the classroom teaching knowledge points, which are expressed in the form of triples as the classroom teaching knowledge point relationship.

7. The classroom teaching quality evaluation method based on a large language model according to claim 6 is characterized by: Step S312, the specific steps are: Step S3121, using adaptive sampling technology to randomly sample the sentence token sequence, Randomly select candidate sequence fragments from ,in represents the fragment width, Represents a sequence of tokens from a statement The selected statement tokens, Represents a sequence of tokens from a statement The selected statement tokens, Represents a sequence of tokens from a statement The selected sentence tokens, for the i-th sentence token in the selected candidate sequence fragment The relevance of each token is evaluated not only based on randomness but also on information gain; ； in, Represents a statement token The information gain of is the entire candidate sequence fragment The entropy of Indicates that in a given statement token After that, the sequence fragment The conditional entropy of Indicates that in a given statement token After that, the sequence fragment The probability value of Step S3122, using a convolutional neural network (CNN) to process the selected candidate sequence fragments , for the input sequence The output is obtained through the convolution operation. ,in represents the weight matrix of the convolution kernel, is the bias vector; Use global average pooling to extract and simplify the output of the convolutional layer to obtain the feature vector , represents the result vector of global average pooling, Represents the output vector of the convolutional layer, which is the extracted feature vector and the trained weight vector Merge to generate knowledge point vector , followed by the statement token sequence category With knowledge point vector Splicing to form a knowledge point entity vector with comprehensive context , after the logistic regression classifier is used to classify the knowledge point entity vector Making type predictions ,in is the weight parameter of the logistic regression classifier, which completes the knowledge point recognition and obtains the classroom teaching knowledge points. That is, the knowledge points of classroom teaching; Step S3123: Use the self-attention mechanism to process two adjacent candidate sequence fragments and , the two candidate sequence fragments and The self-attention outputs are merged and the information is integrated through a linear layer to obtain the overall context representation: ； in, and Represents two adjacent candidate sequence fragments and Self-attention output, and are the weights and biases of the linear layer; Step S3124, combining the method of extracting knowledge point vectors in step S3122 to obtain knowledge point vectors of adjacent sequence segments and With context representation Splicing to form a complete knowledge point entity relationship pair , That is the relationship between classroom teaching knowledge points.

8. The classroom teaching quality evaluation method based on a large language model according to claim 7 is characterized by: Step S32 is based on the classroom teaching evaluation criteria and specifications and the classroom teaching knowledge points extracted in step S31. Relationship with classroom teaching knowledge points , construct the knowledge constraint C of classroom teaching quality evaluation; the specific steps are: Step S321, according to the classroom teaching evaluation criteria, the teacher's classroom teaching quality evaluation should be evaluated from multiple aspects such as teaching attitude, teaching content, teaching methods, teaching effect, classroom atmosphere, learning status, and learning effect, and the evaluation basis of each item is different; Step S322, generating a teaching evaluation description that complies with the classroom teaching evaluation criteria and setting classroom teaching quality evaluation knowledge constraints as follows: C={"{Teaching attitude: [Teacher's facial expression characteristics, teacher's voice characteristics]}" "{Teaching content: [classroom teaching knowledge points, relationship between classroom teaching knowledge points]}" "{teaching method: [teacher behavior data]}" "{Teaching effect: [Teacher's language fluency, teacher's language clarity, student's attention data, student's expression data]}" "{Learning status: [student behavior data, student expression data]}" "{Learning effect: [student attendance rate, head raising rate, distribution of sentiment words]}" "{Classroom atmosphere: [characteristics of the classroom environment]}"}.

Images