AI intelligent live broadcast system

Abstract

The invention discloses an AI intelligent live broadcast system, and the system specifically comprises a data collection module which is used for obtaining original text data generated by a plurality of students during live broadcast interaction; the text processing module is used for performing word segmentation and part-of-speech tagging on the original text data, and identifying terminologies and abbreviations in the original text data through a preset terminology library and an abbreviation library to form first text data; the emotion recognition module is used for matching all emotion words in the first text data according to a preset emotion dictionary and determining the emotion state of each student; and the state judgment module is used for associating and combining the emotional state of each student with the corresponding student eye movement data and learning behavior data to determine the current learning state of each student. According to the invention, accurate evaluation of the learning state of the student and dynamic adjustment of the personalized teaching content are realized, the adaptability of the teaching scene is optimized, and the sharing efficiency of teaching resources is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI intelligent live broadcast system. Background Art

[0002] With the rapid development of the online education industry, AI intelligent live broadcast systems have gradually become an important tool in the field of education. These systems integrate various advanced technologies such as natural language processing, emotion recognition, eye tracking, etc. to provide a more efficient and personalized online learning experience. However, existing AI intelligent live broadcast systems still face many challenges in dealing with students' real learning status and emotional tendencies in specific teaching scenarios.

[0003] First, during the training of deep learning models, a significant problem is that the model has a low accuracy rate in recognizing subtle emotions. Students often have very different words and sentences to express the same emotion in different courses. For example, in math class, students may express confusion by saying "this question is too difficult", while in Chinese class, they may express the same emotion by saying "this article is really difficult to understand". This difference makes it difficult for existing emotion recognition models to accurately capture students' true emotions when faced with changing expressions.

[0004] Secondly, in live courses with a strong professional focus, the questions asked by students often contain a lot of professional terms and abbreviations. These terms and abbreviations have their own unique meanings and contexts in specific fields, but existing natural language processing models often have difficulty accurately understanding the deep meanings of these terms. This results in the intelligent question-answering system being unable to give effective answers when dealing with such questions, thus affecting students' learning outcomes and experience.

[0005] In addition, students may use some obscure or ironic expressions to express confusion or dissatisfaction during the live broadcast. These expressions often deviate greatly from the conventional emotion recognition model training data, so it is difficult for the system to accurately capture these subtle emotional changes. For example, when a student says "This question is so 'simple' that I can't understand it", his true emotion may be confusion or dissatisfaction, but the system may not be able to recognize this ironic expression.

[0006] The existence of these challenges makes it very difficult for existing AI intelligent live broadcast systems to judge whether students truly understand the teaching content and whether teaching strategies need to be adjusted according to students' personalized needs. Summary of the invention

[0007] The purpose of the present invention is to provide an AI intelligent live broadcast system, which realizes accurate evaluation of students' learning status and dynamic adjustment of personalized teaching content, optimizes the adaptability of teaching scenarios, and improves the sharing efficiency of teaching resources, so as to solve at least one of the above-mentioned prior art problems.

[0008] The present invention discloses an AI intelligent live broadcast system, which specifically includes: The data collection module is used to obtain the original text data generated by multiple students during live interaction; A text processing module, used to perform word segmentation and part-of-speech tagging on the original text data, and identify professional terms and abbreviations in the original text data through a preset professional term library and abbreviation library to form first text data; An emotion recognition module, used to match each emotion word in the first text data according to a preset emotion dictionary to determine the emotional state of each student; The state judgment module is used to associate each student's emotional state with the corresponding student's eye movement data and learning behavior data to determine each student's current learning state; A comprehension judgment module is used to obtain the online answering data of each student and determine the comprehension data of each student by integrating the current learning state and the online answering data; The strategy adjustment module is used to dynamically adjust the personalized teaching strategy for each student based on the comprehension data and a preset teaching strategy matching library.

[0009] Compared with the prior art, the present invention has at least one of the following technical effects: 1. It realizes accurate assessment of students' learning status and dynamic adjustment of personalized teaching content, optimizes the adaptability of teaching scenarios, and improves the sharing efficiency of teaching resources.

[0010] 2. Through the preset professional terminology library and abbreviation library, professional terms and abbreviations in the text can be effectively identified and processed, avoiding analysis errors caused by improper understanding of terminology.

[0011] 3. Use the preset sentiment dictionary to accurately match sentiment words in the text, preliminarily determine the students’ emotional state, and combine the deep learning model’s judgment of word types to further correct the emotional state, thereby improving the accuracy and robustness of emotion recognition.

[0012] 4. Comprehensively evaluate students' current learning status by integrating their emotional state, eye movement data and learning behavior data. And integrate multiple data through the decision tree model to improve the accuracy and reliability of learning status judgment.

[0013] 5. Combined with students’ current learning status and online answering data, the system comprehensively evaluates students’ understanding, and uses a recurrent neural network model for integrated evaluation, which improves the accuracy and adaptability of understanding judgment.

[0014] 6. By periodically collecting students’ learning behavior logs, we can form student portraits and provide a basis for the formulation of personalized teaching strategies.

[0015] 7. Based on the student portrait and teaching strategy matching library, dynamically adjust the personalized teaching strategy for each student to improve the pertinence and effectiveness of teaching.

[0016] 8. Automatically adjust scene parameter configurations, such as background parameters and sound effect parameters, according to the type of live broadcast room, optimize the live broadcast environment, and improve students' viewing experience and learning effects.

[0017] 9. The association rule mining algorithm is used to analyze the relationship between teaching content, interaction methods and scene parameters, revealing the key factors affecting teaching effectiveness.

[0018] 10. By comparing the student participation corresponding to the association rules with the preset participation threshold, a teaching effect analysis report is formed, providing strong support for teaching improvement.

[0019] 11. According to the format requirements of different educational platforms, the live broadcast source data is transcoded to achieve multi-platform simultaneous live broadcast and sharing of teaching resources, improve the compatibility and scalability of the live broadcast system, and facilitate students to watch live broadcasts and learn on different platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 It is a structural diagram of an AI intelligent live broadcast system provided by the first embodiment of the present invention; Figure 2 It is a structural diagram of an AI intelligent live broadcast system provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0022] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0023] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0024] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0025] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0026] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0027] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0028] Figure 1 The structure diagram of the AI ​​intelligent live broadcast system disclosed in one embodiment of the present invention is shown, and the details are as follows: The data collection module 101 is used to obtain the original text data generated by multiple students during live interaction; The text processing module 102 is used to perform word segmentation and part-of-speech tagging on the original text data, and identify professional terms and abbreviations in the original text data through a preset professional term library and abbreviation library to form first text data; The emotion recognition module 103 is used to match each emotion word in the first text data according to a preset emotion dictionary to determine the emotion state of each student; The state determination module 104 is used to associate and combine the emotional state of each student with the corresponding student eye movement data and learning behavior data to determine the current learning state of each student; The comprehension judgment module 105 is used to obtain the online answering data of each student, and determine the comprehension data of each student by integrating the current learning state and the online answering data; The strategy adjustment module 106 is used to dynamically adjust the personalized teaching strategy for each student according to the comprehension data and a preset teaching strategy matching library.

[0029] In this embodiment, the data collection module is used to collect the original text data posted by students in the discussion area in real time through the chat room, bullet screen or voice-to-text function during the live interaction. For example, student A asked, "How do you calculate the time complexity of this algorithm?" Student B replied, "Use O(n), which means that as the input size n increases, the execution time increases linearly." The text processing module is used to use natural language processing technology (NLP) to segment and tag the original text data. For example, "How to calculate the time complexity of this algorithm?" is segmented into "How to calculate the time complexity of this algorithm? ", and then use the preset professional terminology library (such as "time complexity", "algorithm") and abbreviation library (such as "O(n)" is interpreted as "linear time complexity") to identify and mark the professional terms and abbreviations to form the first text data.

[0030] The emotion recognition module is used to match the emotion words in the first text data according to the preset emotion dictionary (including positive, negative, neutral emotion words and their intensities). For example, the words "difficult" and "cannot learn" in student C's statement "This is too difficult, I can't learn it!" are identified as negative emotions, and the emotional state of each student is determined, such as student A showing curiosity, student B showing confidence, and student C showing frustration.

[0031] The state judgment module is used to combine students' eye movement data (such as fixation time and pupil diameter changes, obtained through cameras and eye tracking technology) and learning behavior data (such as speaking frequency and online time) to further refine the learning state. For example, although student A is curious, his wandering eyes may indicate distraction; although student C is frustrated, he continues to try to answer questions, indicating that he is still working hard.

[0032] The comprehension judgment module is used to collect students' online answering data during the live broadcast, including accuracy, answering speed, etc., and comprehensively evaluate students' comprehension by combining their current learning status (such as attention and emotions) and online answering data. For example, although student B is confident, his high error rate in answering questions may mean that his understanding is biased; although student C is frustrated, he answers some questions correctly, indicating that he has mastered certain knowledge points.

[0033] The strategy adjustment module is used to dynamically adjust the teaching strategy based on the comprehension data and the preset teaching strategy matching library. For example, for student C who has low comprehension and shows frustration, the system can recommend more basic exercises and explanation videos; for student A who has high comprehension but is distracted, interactive sessions or questions can be added to guide his attention.

[0034] Furthermore, the identifying of professional terms and abbreviations in the original text data by using a preset professional term library and abbreviation library specifically includes: Based on a preset professional terminology library, a similarity algorithm is used to calculate the similarity between each word in the original text data and each professional term in the professional terminology library to obtain a professional terminology recognition result; Based on a preset abbreviation library, a similarity algorithm is used to calculate the similarity between each word in the original text data and each abbreviation in the abbreviation library to obtain an abbreviation recognition result; Determining a number of professional terms in the original text data according to the professional term recognition result, and classifying each professional term into different professional categories using an information classification method; Determine a number of abbreviations in the original text data according to the abbreviation recognition result, and determine complete semantic information of each abbreviation using an understanding method; According to the context information of each professional term or each abbreviation, it is judged whether each professional term or each abbreviation belongs to the current course content, and the professional terms or abbreviations that do not belong to the current course content are marked.

[0035] In this embodiment, a preset professional term library and abbreviation library are used to obtain word units in the original text data. For each word unit, the similarity value between it and each term or abbreviation in the term library and the abbreviation library is calculated by a similarity algorithm. According to the similarity value, it is determined whether the word unit matches the content in the term library or the abbreviation library. If it matches, it is determined that the word unit is a professional term or abbreviation, and the term recognition result and the abbreviation recognition result are obtained.

[0036] Using the professional terminology recognition algorithm, the original text data is deeply analyzed to extract the professional terminology list in the text, and the frequency and location information of each term is recorded. According to the pre-established professional terminology classification library, the information classification method is used to match each extracted professional term to the corresponding professional category one by one to generate the professional terminology classification result. Through the abbreviation recognition technology, the cleaned text data is further analyzed to identify the abbreviation list in the text, and the frequency and location information of each abbreviation is recorded. Using the semantic understanding technology, each identified abbreviation is semantically parsed to obtain its corresponding complete semantic information and generate the abbreviation semantic parsing result. The professional terminology classification results and the abbreviation semantic parsing results are integrated to build a comprehensive semantic information library for subsequent information retrieval and application. If new professional terms or abbreviations are found, the professional terminology classification library and the abbreviation semantic library are updated to ensure continuous learning and optimization of the system.

[0037] For the extracted terms and abbreviations, a preset library containing all professional terms and abbreviations of the current course is constructed to ensure the integrity and accuracy of the data in the library. The extracted terms are compared one by one with the terms in the preset library, and a string matching algorithm is used to ensure the accuracy of the comparison. If the term does not exist in the preset library, it is determined that the term does not belong to the current course content and is marked as a "non-course term". The marked terms are returned to the original text and highlighted or other visual methods are used to allow users to quickly identify terms that do not belong to the course content. Through statistical analysis, the frequency and distribution of non-course terms are determined, and a term distribution report is generated to provide data support for the subsequent optimization of course content. Combined with the term distribution report, the course content is revised, and non-course terms are deleted or replaced to ensure the accuracy and pertinence of the course content.

[0038] Exemplarily, the preset professional terminology library is a data set containing professional vocabulary in various fields, such as a medical terminology library that may contain vocabulary such as "myocardial infarction" and "thrombocytopenia". The similarity algorithm can use methods such as cosine similarity or edit distance to calculate the similarity between each word in the original text and the words in the terminology library. For example, "myocardial infarction" appears in the text, which is very similar to "myocardial infarction" in the terminology library and may be identified as a professional term. The abbreviation library is also a pre-established data set that contains various common abbreviations and their full names, such as "DNA" corresponding to "deoxyribonucleic acid" and "CPU" corresponding to "central processing unit". When using a similarity algorithm to identify abbreviations, you may encounter a word with multiple meanings, such as "IC" may refer to "integrated circuit" or "identity card", which needs to be judged in combination with the context. Professional term classification is to classify the identified terms into different professional fields. Machine learning methods such as naive Bayes or support vector machines can be used for classification. For example, "carbon dioxide" may be classified as the chemical field, and "stock options" may be classified as the financial field. This classification helps to understand the subject and professional background of the text. Semantic understanding of abbreviations involves restoring abbreviations to complete phrases and understanding their meanings. For example, after "WHO" is identified as "World Health Organization", it is also necessary to understand that it is a specialized agency of the United Nations dedicated to international public health affairs. Judging whether professional terms or abbreviations belong to the current course content based on the context can improve the pertinence of learning materials. For example, in a computer network course, "IP" (Internet Protocol) belongs to the course content, while "IP" (intellectual property) may be marked as irrelevant. This screening can help learners focus on the core content of the course and avoid unnecessary interference. Through this series of steps, professional terms and abbreviations in the text can be effectively identified, understood and screened, providing a basis for subsequent text analysis, knowledge extraction and personalized learning. This method can not only improve the accuracy of text understanding, but also provide a more targeted reading experience for readers from different backgrounds.

[0039] Furthermore, matching each emotion word in the first text data according to a preset emotion dictionary to determine the emotion state of each student specifically includes: Based on a preset sentiment dictionary, text matching is performed on each word in the first text data to determine a number of sentiment words in the first text data; Accumulate the emotion value corresponding to each emotion word to obtain the first emotion state of each student; Using a pre-trained deep learning model to determine the word types in the first text data to obtain word classification results, wherein the word types include commendatory words, derogatory words, and ironic words; The first emotional state is modified according to the word classification result to obtain a second emotional state.

[0040] In this embodiment, the first text data is obtained, segmented, and a segmentation list is generated. The sentiment words and their corresponding sentiment values ​​are read from a preset sentiment dictionary to construct a sentiment word library. The segmentation list is traversed, each word is matched with the sentiment word library, and the sentiment words are identified. For the identified sentiment words, the corresponding sentiment values ​​are obtained, and an accumulation operation is performed to obtain a preliminary sentiment sum. According to the accumulated sentiment sum, the first emotional state of each student is calculated to generate the emotional state data.

[0041] A pre-trained deep learning model, such as BERT or GPT, is used to extract word-level features from the input first text data. The emotional tendency of each word is identified and classified through the attention mechanism and multi-layer neural network inside the model, and the classification labels of commendatory words, derogatory words and ironic words are output. The classification results are represented in the form of vectors, and each word corresponds to a vector containing emotional category information. According to the word classification results, an emotional state correction matrix is ​​constructed. The matrix is ​​pre-defined and contains the correction weights of different emotional categories on the emotional state. The classification vector of each word in the first text data is multiplied by the correction matrix to obtain an emotional correction vector. The emotional correction vectors of all words are weighted and summed to generate a total emotional correction value. The first emotional state is represented as a multi-dimensional emotional vector, which contains multiple dimensions of emotions, such as pleasure, excitement, etc. The obtained emotional correction value is added to the first emotional state vector to obtain a corrected second emotional state vector. The second emotional state vector is normalized to ensure that the emotional state is within the preset emotional space. If the difference between the second emotional state vector and the first emotional state vector exceeds a preset threshold, the emotional state update mechanism is triggered. The update mechanism adjusts the second emotional state vector according to the difference value and the preset emotion transfer rules to make it more consistent with the overall emotional tendency of the text data. Through the emotional state update mechanism, the second emotional state is finally determined and used as the output. The second emotional state contains the corrected emotional dimension information, which can more accurately reflect the true emotional tendency of the text data. In the model training stage, a large-scale emotional annotation corpus is used to fine-tune the pre-trained model to improve the accuracy and robustness of the model in the emotion classification task. Through cross-validation and model evaluation indicators such as accuracy, recall rate and F1 score, ensure that the classification performance of the model meets the preset standards. In practical applications, user feedback and text data are continuously collected, and the pre-trained model and emotional state correction matrix are regularly updated and optimized to adapt to the ever-changing text emotional expression and emotional state change rules.

[0042] For example, a sentiment dictionary is a preset data set containing words and their corresponding sentiment values, for example, "happy" may be assigned +2 points, and "sad" may be assigned -2 ​​points. Through text matching technology, these sentiment words can be identified in the student's text data. Assuming that a student's text contains "Today's exam went smoothly, but the weather was not very good", the system may identify two sentiment words, "smooth" (+1 point) and "not very good" (-1 point). Accumulating sentiment values ​​is the process of adding the corresponding scores of the identified sentiment words. In the above example, the student's initial emotional state score is 0 (+1-1=0). This method can quickly obtain the overall emotional tendency of the text, but may ignore the subtle differences in context and tone. Pre-trained deep learning models, such as BERT or XLNet, can more accurately determine the type of words. These models are trained on large-scale corpora and can understand the meaning of words in different contexts. For example, "awesome" is a commendatory word in most cases, but may be ironic in a specific context. The application of word classification results is crucial to the correction of emotional states. Suppose the system detects a sentence "This question is so easy", the initial sentiment value may be positive, but if the deep learning model determines that "easy" is used ironically here, the system will correct the emotional state to negative. This correction can capture subtle emotional changes in students' expressions. For example, if the system finds that multiple students frequently use negative words when discussing a topic, teachers may need to adjust their teaching strategies or provide additional support.

[0043] Furthermore, the emotional state of each student is associated with the corresponding student eye movement data and learning behavior data to determine the current learning state of each student, specifically including: Acquire students' teaching video viewing data, and extract eye movement data of each student based on the students' teaching video viewing data; Determine the gaze point position and the corresponding dwell time of each student according to the student eye movement data, and obtain the concentration of each student according to the gaze point position and the corresponding dwell time of each student; Extracting learning behavior data of each student in a preset time period based on the teaching video data watched by the students, obtaining the number of interactions of each student in the preset time period based on the learning behavior data, and determining the participation of each student according to the number of interactions of each student in the preset time period; Based on a pre-built decision tree model, the emotional state of each student, the concentration and the participation are associated and fused to obtain the current learning state of each student.

[0044] In this embodiment, the original behavior data of students watching teaching videos is obtained, and basic information such as the viewing time and viewing frequency of each student is recorded. The eye movement data of students watching the video is collected through eye tracking equipment, including detailed information such as pupil position and gaze point coordinates. The collected eye movement data is preprocessed to remove noise and invalid data to ensure the accuracy and reliability of the data. A data mining algorithm is used to extract the gaze point position of each student and its corresponding dwell time from the preprocessed eye movement data to form a gaze point distribution map and time series data. According to the gaze point position and dwell time, the concentration calculation model is used to calculate the concentration score of each student. The concentration score reflects the student's attention to the video content.

[0045] Obtain each student's viewing record from the teaching video system, including student ID, viewing time, viewing duration, interactive behavior and other learning data. According to the preset time period, perform time slicing on each student's viewing record to obtain each student's learning behavior data within the preset time period to form a learning record. Analyze each student's learning record, extract interactive behavior, count the number of interactions of each student within the preset time period, and record the interaction type. Based on the number of interactions, use the preset participation calculation model to calculate the degree of participation of each student and form a participation score.

[0046] The emotional state, concentration and engagement data are integrated, and feature fusion technology is used to combine the leaf value and feature mapping relationship of the decision tree model to generate a comprehensive learning state feature vector. The comprehensive learning state feature vector is scored and classified, and the current learning state of each student is judged through the classification rules and preset thresholds of the decision tree model, and the learning state label is output. Based on the learning state label, personalized learning suggestions are generated, and a recommendation algorithm is used to push targeted learning resources and improvement strategies in combination with students' historical learning data and course content.

[0047] Exemplarily, first, the eye movement data of students is obtained through eye tracking technology, which can accurately capture the movement trajectory and the position of students' sight. For example, when students watch a video explaining a mathematical formula, the system can record whether their eyes are focused on the key steps and how long they stay on each step. The calculation of concentration is based on the position of the gaze point and the dwell time. Assuming that when a student is watching the key content in the video, his gaze point frequently appears in the center of the screen and stays for a long time each time, which indicates that the student has a high degree of concentration. On the contrary, if the student's sight frequently wanders to the edge of the screen or changes frequently, it may indicate a low degree of concentration. The extraction of learning behavior data involves multiple aspects, such as pause, fast forward, rewind and other operations. The measurement of participation is mainly determined by the number of interactions of students within a preset time. For example, an actively participating student may frequently use the functions of the video player, such as repeatedly watching difficult passages or using the note function. These behaviors reflect the student's active learning attitude. The fusion of emotional state, concentration and participation uses a pre-built decision tree model. This model can comprehensively consider multiple factors and give a more comprehensive assessment of learning status. For example, a student may show high concentration, but have a negative emotional state, which may indicate that the student faces learning difficulties. The decision tree model can weigh these factors and give a more accurate judgment of learning status. The advantage of this comprehensive analysis method is that it can comprehensively evaluate the learning status of students. It not only considers students' objective behaviors, such as eye movement and interactive operations, but also combines the subjective factor of emotional state. This multi-dimensional analysis helps educators better understand the learning needs of each student and provide personalized teaching support.

[0048] Furthermore, the comprehension data of each student is determined by integrating the current learning status and the online answering data, specifically including: Determine the correct answer rate of each student based on the online answer data; A pre-built recurrent neural network model is used to integrate and evaluate the correct answer rate and the current learning status to determine the comprehension data of each student.

[0049] In this embodiment, the online answer record of each student is obtained, the total number of answers and the number of correct answers are counted, and the answer accuracy is calculated. The answer accuracy data is input into the pre-constructed recurrent neural network model, combined with the current learning status data, a fusion evaluation is performed, and the understanding data of each student is output. According to the understanding data, it is determined whether the student needs additional tutoring. If the understanding is lower than the preset threshold, a personalized learning suggestion is generated and relevant learning resources are pushed. The changes in the student's learning status are continuously monitored, the understanding data is updated regularly, and the learning suggestions and resource push strategies are adjusted. The online answer record of each student is obtained, the total number of answers and the number of correct answers are counted, and the answer accuracy is calculated. The answer accuracy data is input into the pre-constructed recurrent neural network model, combined with the current learning status data, a fusion evaluation is performed, and the understanding data of each student is output. According to the understanding data, it is determined whether the student needs additional tutoring. If the understanding is lower than the preset threshold, a personalized learning suggestion is generated and relevant learning resources are pushed. The changes in the student's learning status are continuously monitored, the understanding data is updated regularly, and the learning suggestions and resource push strategies are adjusted.

[0050] For example, first, it is necessary to determine the correct rate of each student's answer. This step involves collecting and processing the raw data of students' online answers, including the answer status and correctness of each question. For example, if a student completes a set of 50 online tests and answers 40 questions correctly, then the student's correct rate is 80%. The calculation of the correct rate seems simple, but it may face some challenges in practical applications. For example, the difficulty of different types of questions may vary, and directly using a simple correct rate may not accurately reflect the student's true level. To solve this problem, the question difficulty coefficient can be introduced to give different weights to questions of different difficulty levels. For example, a question with a difficulty coefficient of 1.5 may be more reflective of the student's actual ability than a question with a difficulty coefficient of 0.8. Next, the correct rate of the answer needs to be integrated with the current learning status to determine the understanding data of each student. Here, a pre-built recurrent neural network model is used to complete this task. Recurrent neural network (RNN) is a special neural network structure that is very suitable for processing sequence data and can capture dependencies in time series. In the field of education, the learning process of students is essentially a time series, so the RNN model has a unique advantage here. Specifically, the student's correct answer rate can be used as an input sequence, combined with other indicators that reflect the current learning status, such as learning time, video viewing completion, and exercise completion. These data together constitute a multidimensional vector that describes the student's learning status. By processing these sequence data, the RNN model can learn the dynamic changes in the student's learning status. For example, suppose a student has three consecutive weeks of online learning, and his weekly answer rate is 75%, 80%, and 85%, respectively, and his weekly study time is 10 hours, 12 hours, and 15 hours, respectively. The RNN model can capture this trend over time, not only considering the data of the current week, but also taking into account the performance of the previous weeks. The advantage of this method is that it can identify the improvement or regression trend of the student's learning status, rather than just a static evaluation. Through the processing of the RNN model, the final result is the student's understanding data. This data may be a comprehensive score or an assessment result of multiple dimensions. For example, a student's understanding data may be displayed as "85% mastery of basic knowledge, 70% application ability, and 60% comprehensive analysis ability." This multi-dimensional assessment result can more comprehensively reflect the student's learning situation and help teachers formulate targeted teaching strategies. The advantage of this integrated assessment method is that it not only takes into account the student's immediate performance (answer accuracy), but also incorporates other important factors in the learning process (current learning status). This comprehensive consideration can more accurately reflect the student's actual level of understanding and avoid the deviation that may be caused by relying on a single indicator.For example, some students may perform well in answering questions, but spend less time on actual learning, which may mean that their deep understanding of the knowledge needs to be improved. On the contrary, some students may not answer questions correctly, but have a high degree of learning investment, which may indicate that they are struggling to overcome learning difficulties and need more support and encouragement.

[0051] Furthermore, the personalized teaching strategy for each student is dynamically adjusted according to the comprehension data and the preset teaching strategy matching library, specifically including: Periodically collect students’ interaction records and homework information on the online learning platform to form a learning behavior log; According to the learning behavior log, each student's mastery of different knowledge points is clustered by a clustering algorithm to form a student portrait; Associating the student portraits with different teaching strategies, determining personalized teaching strategy combinations for different student groups, and forming a teaching strategy matching library; The comprehension data is matched through the teaching strategy matching library to determine the personalized teaching strategy for each student.

[0052] In this embodiment, the interaction records and homework information of the student online learning platform are obtained, including login time, browsing pages, answer records, homework submission status, etc., to form an initial learning behavior data set. The initial learning behavior data set is preprocessed, invalid data and noise are cleaned, the data format is standardized, and key features are extracted, such as knowledge point access frequency, homework completion time, answer accuracy, etc., to generate cleaned learning behavior feature data. The K-means clustering algorithm is used to perform cluster analysis on the cleaned learning behavior feature data. According to the students' mastery of different knowledge points, the students are divided into several groups, each group has similar learning behavior characteristics, and the student group clustering results are formed. Based on the student group clustering results, a student portrait is constructed, which includes information such as students' learning habits, knowledge point mastery, and learning difficulties, and detailed student portrait data is generated. A variety of teaching strategies are established in advance, including explanation depth, exercise difficulty, tutoring frequency, etc., and student portraits are associated with teaching strategies. The personalized teaching strategy combination applicable to different student groups is determined by the decision tree algorithm to form a teaching strategy matching library. The comprehension data is matched through the teaching strategy matching library to determine the optimal personalized teaching strategy for each student's current learning state, and generate a personalized teaching strategy recommendation list. According to the personalized teaching strategy recommendation list, the teaching content and method are dynamically adjusted, personalized learning resources are pushed, precise teaching is achieved, students' learning effects are continuously tracked, learning behavior logs and student portraits are updated, and a closed-loop optimization mechanism is formed.

[0053] For example, first, the platform will periodically collect students' learning behavior logs, including login time, learning time, video viewing progress, homework completion, etc. For example, a student logs in to the platform an average of 3 times a day in a week, with a cumulative learning time of 15 hours, completes 80% of video courses, and submits 90% of homework. These data constitute the learning behavior characteristics of the student. Next, the clustering algorithm is used to analyze the students' mastery of knowledge points. Assume that the K-means clustering algorithm is used to divide students' performance on different knowledge points into several categories. For example, in the course of "data structure", there may be several types of students such as "solid basic knowledge type", "strong algorithm application ability type", and "lack of comprehensive analysis type". This clustering result forms a student's knowledge mastery portrait, which helps teachers quickly identify students' strengths and weaknesses. Based on student portraits, a teaching strategy matching library can be established. For example, for students with "solid basic knowledge type", more advanced application questions and actual case analysis can be provided; while for students with "lack of comprehensive analysis type", more concept analysis and thinking training may need to be added. The establishment of this matching relationship is not achieved overnight, and requires continuous practice and adjustment. Teachers can continuously optimize the matching rules based on students' feedback and learning effects. Finally, the students' understanding data is connected to the teaching strategy matching library to generate personalized teaching strategies for each student. For example, for a student who is identified as "strong in algorithm application ability" but "weak in theoretical foundation" in the "Data Structure" course, the system may recommend review materials for basic concepts and provide more difficult algorithm implementation questions. This personalized strategy can not only make up for the students' shortcomings, but also further play their advantages. The advantage of this personalized teaching method based on data analysis is that it can capture changes in students' learning status in a timely manner and make corresponding adjustments. For example, if the system detects that a student has always performed poorly on a specific type of question in a test for several consecutive weeks, it will automatically adjust the student's study plan and increase the amount of practice on the relevant knowledge points. This dynamic adjustment mechanism ensures that the teaching strategy is always consistent with the students' actual needs, thereby maximizing the learning effect. With this method, educators can break through the traditional "one-size-fits-all" teaching model and provide each student with a tailored learning experience. This not only improves learning efficiency, but also enhances students' learning interest and self-confidence, thereby achieving better teaching results.

[0054] In some embodiments, Figure 2 As shown, the system also includes: The live broadcast parameter adjustment module 107 is used to obtain the scene parameter configuration data stream corresponding to the current live broadcast room type according to the pre-established live broadcast room type and scene parameter mapping database, and adjust the background parameters and sound effect parameters of the live broadcast room through the scene parameter configuration data stream.

[0055] In this embodiment, when the system starts, the live broadcast room type and scene parameter mapping database are preloaded to ensure that the database contains all known live broadcast room types and their corresponding scene parameter configurations. The type information of the current live broadcast room is monitored in real time, and the scene parameter configuration data stream corresponding to the type is queried from the mapping database through the API interface. The acquired scene parameter configuration data stream is parsed to extract the specific values ​​and configuration instructions of the background parameters and sound effect parameters. According to the parsed background parameters, the image processing module is called to adjust the live broadcast room background in real time, including dynamic modification of parameters such as color, brightness, and contrast. According to the parsed sound effect parameters, the audio processing module is called to adjust the live broadcast room sound effects in real time, including dynamic modification of parameters such as volume, pitch, and reverberation. If the type of the live broadcast room changes, the above steps are re-executed to ensure that the live broadcast room parameters always match the current type. The operation log of each parameter adjustment is recorded, including the adjustment time, parameter values ​​before and after adjustment, and other information, for subsequent analysis and optimization.

[0056] Exemplarily, the live broadcast parameter adjustment module is an intelligent live broadcast scene optimization tool designed to enhance user experience and interactive effects. The core function of this module is to obtain the scene parameter configuration data stream corresponding to the current live broadcast room type according to the pre-established live broadcast room type and scene parameter mapping database, and adjust the background parameters and sound effect parameters of the live broadcast room in real time through these data streams. The live broadcast room type and scene parameter mapping database is the basis of this module. This database may contain a variety of live broadcast types, such as educational lectures, online classes, seminars, etc. Each type corresponds to a series of optimized scene parameters. For example, for live broadcasts of educational lecture types, the database may recommend the use of soft background tones, clear subtitle displays, and moderate ambient sound effects to create an atmosphere of focused learning. The scene parameter configuration data stream is a set of specific parameters extracted from the database, which may include multiple dimensions such as background color, brightness, contrast, volume, reverberation, etc. The parameter adjustment process is dynamic and real-time. When the live broadcast type changes, the system will immediately extract the new parameter configuration from the database and smoothly transition to the new settings. This smooth transition can be achieved through a gradient effect to avoid abrupt changes affecting the audience experience. In addition, the module can also self-optimize based on audience feedback and interaction data. By analyzing data such as audience dwell time and interaction frequency, the system can fine-tune parameter configurations and continuously improve user experience. For example, if it is found that the average viewing time of viewers is longer under a certain background tone, the system will tend to use this tone more often.

[0057] In some embodiments, Figure 2 As shown, the system also includes: The teaching effect analysis module 108 is used to analyze the association relationship between the teaching content, interaction mode and scene parameters of multiple students in the live interaction process by using an association rule mining algorithm, and obtain multiple association rules between the teaching content, interaction mode and scene parameters; By comparing the student participation corresponding to each association rule with the preset participation threshold, the effective association rule set and the invalid association rule set are determined to form a teaching effect analysis report.

[0058] In this embodiment, an association rule mining algorithm is used to extract the characteristic vectors of teaching content, interaction mode and scene parameters from live interactive data, and construct a characteristic matrix; frequent item sets are generated by the Apriori algorithm, support and confidence thresholds are set, and qualified association rules are screened out; the student participation corresponding to each association rule is calculated, compared with the preset participation threshold, and a valid association rule set and an invalid association rule set are divided; based on the valid association rule set, the association relationship between teaching content, interaction mode and scene parameters is analyzed, and a teaching effect analysis report is generated. If the support of the association rule is higher than the preset support threshold, and the confidence is higher than the preset confidence threshold, the rule is included in the frequent item set; otherwise, the rule is eliminated. Association rules are generated through frequent item sets, and the antecedents and consequents of the rules are recorded to form an association rule library. Student participation data, including indicators such as viewing time, number of interactions, and correct answer rate, are obtained to construct a student participation vector; the student participation vector corresponding to each association rule is compared with the preset participation threshold. If it is higher than the threshold, the rule is marked as a valid association rule; otherwise, it is marked as an invalid association rule. According to the effective association rule set, the association relationship between teaching content and interaction mode and scene parameters is analyzed to identify the key factors affecting teaching effect; combined with the confidence and support of the association rules, the contribution of each factor to the teaching effect is evaluated to form a multi-dimensional teaching effect analysis report. Through data visualization technology, the key indicators and association relationships in the teaching effect analysis report are displayed in the form of charts, which is convenient for intuitive understanding and decision support. Regularly update the live interaction data, re-execute the association rule mining and analysis process, dynamically adjust the teaching strategy, and optimize the teaching effect. Using the decision tree model in the machine learning algorithm, the effective association rules are further analyzed to identify the key features that affect student participation, generate feature importance ranking, and provide data support for teaching optimization. Combined with the clustering algorithm, cluster analysis is performed on the student participation data to identify student groups with different levels of participation and formulate differentiated teaching strategies. Through natural language processing technology, sentiment analysis is performed on the teaching content text to identify positive and negative sentiment words. Combined with association rule analysis, the impact of teaching content on students' emotions is evaluated to provide a reference for teaching content optimization.

[0059] For example, the teaching effect analysis module is an innovative educational technology tool designed to improve the quality of live teaching and student participation. This module uses an association rule mining algorithm to deeply analyze the relationship between various factors in the live interaction process of students, thereby optimizing teaching strategies. The association rule mining algorithm is a data mining technology used to discover hidden patterns in large amounts of data. In teaching scenarios, this algorithm can reveal the potential connections between teaching content, interaction methods, and scene parameters. For example, the algorithm may find a rule such as "when the teacher uses animation demonstrations (teaching content) and conducts real-time questions and answers (interaction methods), student participation is significantly improved." In actual applications, the system may collect and analyze the following data: teaching content (such as video explanations, text handouts, chart displays, etc.), interaction methods (such as barrage interaction, voice questions and answers, voting, etc.), and scene parameters (such as background music, screen layout, subtitle size, etc.). By analyzing these data, the system may derive multiple association rules, such as "using vivid animation demonstrations + real-time barrage interaction + soft background music" and "higher student participation" have a strong correlation. In order to determine which association rules are effective, the system compares the student engagement corresponding to each rule with the preset engagement threshold. Assuming the preset threshold is 70%, any rule that leads to a student engagement of more than 70% will be classified as a valid association rule set. Conversely, those rules with an engagement below the threshold are classified as an invalid association rule set. This classification method can help educators quickly identify which teaching strategies are most effective and which need to be improved. For example, if it is found that the rule "using professional terms to explain + no interaction + monotonous background" leads to a student engagement of only 30%, which is far below the threshold, then this rule will be classified as an invalid set, prompting teachers to adjust their teaching methods. Based on these analysis results, the system will generate a detailed teaching effectiveness analysis report. This report not only lists effective and invalid association rules, but may also contain specific improvement suggestions. For example, the report may suggest "adding chart presentations and real-time question-and-answer sessions when explaining complex concepts, while keeping light music in the background, which can increase student engagement from the current 50% to about 85%." In this way, the teaching effectiveness analysis module can help educators continuously optimize their teaching strategies and improve students' learning interest and effectiveness. It not only provides teachers with data-supported decision-making basis, but also promotes the continuous improvement and innovation of teaching methods. This data-based teaching optimization method has the potential to completely change the traditional education model and make education more personalized and efficient.

[0060] In some embodiments, Figure 2 As shown, the system also includes: The cross-platform publishing module 109 is used to perform transcoding operations on the live source data according to pre-established encoding rules to obtain live streams adapted to different educational platform formats, wherein the transcoding operations include video encoding format conversion, audio encoding format conversion, and resolution adjustment; According to the platform interface protocol, live streams in different formats are distributed to the corresponding education platform interface to achieve multi-platform simultaneous live broadcast and sharing of teaching resources.

[0061] In this embodiment, the live source data is obtained, the original encoding formats of the video and audio are parsed, and the current resolution parameters are recorded. According to the preset format requirements of the education platform, the target video encoding format, audio encoding format and resolution standard are determined. The source data is converted into a video encoding format, and the H.264 or H.265 encoding standard is adopted to generate a video stream that meets the requirements of the target platform. The source data is converted into an audio encoding format, and the AAC or MP3 encoding standard is adopted to generate an audio stream that meets the requirements of the target platform. The video stream resolution is adjusted, and a scaling algorithm is adopted to ensure that the output resolution is consistent with the target platform standard. According to the interface protocols of each education platform, the converted video stream and audio stream are encapsulated to generate a live stream data packet adapted to each platform. Through a multi-threaded distribution mechanism, the live stream data packet is transmitted to each education platform interface in real time to realize multi-platform synchronous live broadcast and teaching resource sharing.

[0062] Exemplarily, the cross-platform publishing module is a key component for realizing multi-platform synchronous live broadcast and teaching resource sharing. The module first transcodes the live source data according to the pre-established encoding rules to adapt to the format requirements of different education platforms. The transcoding operation includes video encoding format conversion, audio encoding format conversion and resolution adjustment. In terms of video encoding format conversion, common formats include H.264, H.265 and VP9, ​​for example, an online education platform requires the use of H.264 encoding, while another platform may require H.265 encoding. The transcoding module converts the source video to the corresponding encoding format according to the requirements of the target platform. This ensures that the video can be played smoothly on different platforms while ensuring the picture quality. Audio encoding format conversion is equally important. Common audio encoding formats include AAC, MP3 and Opus. Assuming that the source audio is in AAC format, but an education platform only supports MP3 format, the transcoding module will convert the audio from AAC to MP3. This conversion ensures the compatibility of audio on various platforms, allowing students to clearly hear the teacher's explanation. Resolution adjustment is a necessary step to adapt to different devices and network conditions. For example, the source video may be 1080p resolution, but when viewed on a mobile device, 720p or 480p may be more appropriate. The transcoding module generates multiple versions with different resolutions, such as 1080p, 720p, and 480p, according to the platform characteristics and user devices to adapt to different viewing needs. After completing the transcoding, the module distributes the live streams in different formats to the corresponding platforms according to the interface protocols of each educational platform. This process involves data transmission and protocol adaptation. For example, one platform may use the RTMP protocol, while another platform may use the HLS protocol. The module packages the transcoded live stream into an appropriate format according to the requirements of each platform and sends it through the specified interface. This cross-platform publishing method brings significant technical effects. First, it realizes the wide sharing of educational resources, allowing high-quality courses to be broadcast live on multiple platforms at the same time, expanding the audience range. Second, it improves the compatibility and accessibility of live broadcasts, so that students can have a good learning experience regardless of the device or platform they use. Finally, through intelligent transcoding and distribution, the module optimizes the utilization of network resources and ensures stable live broadcast services under different network conditions. In practical applications, this module can greatly improve the efficiency and coverage of online education. For example, an important math class can be broadcast live on multiple educational platforms at the same time, and students can choose the platform to watch according to their habits. Whether using a computer, tablet or mobile phone, you can get the best viewing experience suitable for your device. This not only facilitates students, but also provides greater flexibility for educational institutions, enabling them to use teaching resources more effectively and reach more learners.

[0063] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0064] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0065] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0066] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Claims

1. An AI intelligent live broadcast system, characterized in that: The system specifically comprises: The data collection module is used to obtain the original text data generated by multiple students during live interaction; A text processing module, used to perform word segmentation and part-of-speech tagging on the original text data, and identify professional terms and abbreviations in the original text data through a preset professional term library and abbreviation library to form first text data; An emotion recognition module, used to match each emotion word in the first text data according to a preset emotion dictionary to determine the emotional state of each student; The state judgment module is used to associate each student's emotional state with the corresponding student's eye movement data and learning behavior data to determine each student's current learning state; A comprehension judgment module is used to obtain the online answering data of each student, and determine the comprehension data of each student by integrating the current learning state and the online answering data; The strategy adjustment module is used to dynamically adjust the personalized teaching strategy for each student based on the comprehension data and a preset teaching strategy matching library.

2. The system according to claim 1, characterized in that The identifying of professional terms and abbreviations in the original text data by using a preset professional term library and abbreviation library specifically includes: Based on a preset professional terminology library, a similarity algorithm is used to calculate the similarity between each word in the original text data and each professional term in the professional terminology library to obtain a professional terminology recognition result; Based on a preset abbreviation library, a similarity algorithm is used to calculate the similarity between each word in the original text data and each abbreviation in the abbreviation library to obtain an abbreviation recognition result; Determining a number of professional terms in the original text data according to the professional term recognition result, and classifying each professional term into different professional categories using an information classification method; Determine a number of abbreviations in the original text data according to the abbreviation recognition result, and determine complete semantic information of each abbreviation using an understanding method; According to the context information of each professional term or each abbreviation, it is judged whether each professional term or each abbreviation belongs to the current course content, and the professional terms or abbreviations that do not belong to the current course content are marked.

3. The system according to claim 1, characterized in that The step of matching each emotion word in the first text data according to a preset emotion dictionary to determine the emotion state of each student specifically includes: Based on a preset sentiment dictionary, text matching is performed on each word in the first text data to determine a number of sentiment words in the first text data; Accumulate the emotion value corresponding to each emotion word to obtain the first emotion state of each student; Using a pre-trained deep learning model to determine the word types in the first text data to obtain word classification results, wherein the word types include commendatory words, derogatory words, and ironic words; The first emotional state is modified according to the word classification result to obtain a second emotional state.

4. The system according to claim 1, characterized in that The step of associating each student's emotional state with the corresponding student eye movement data and learning behavior data to determine each student's current learning state specifically includes: Acquire students' teaching video viewing data, and extract eye movement data of each student based on the students' teaching video viewing data; Determine the gaze point position and the corresponding dwell time of each student according to the student eye movement data, and obtain the concentration of each student according to the gaze point position and the corresponding dwell time of each student; Extracting learning behavior data of each student in a preset time period based on the teaching video data watched by the students, obtaining the number of interactions of each student in the preset time period based on the learning behavior data, and determining the participation of each student according to the number of interactions of each student in the preset time period; Based on a pre-built decision tree model, the emotional state of each student, the concentration and the participation are associated and fused to obtain the current learning state of each student.

5. The system according to claim 1, characterized in that The step of determining the understanding data of each student by integrating the current learning status and the online answering data specifically includes: Determine the correct answer rate of each student based on the online answer data; A pre-built recurrent neural network model is used to integrate and evaluate the correct answer rate and the current learning status to determine the comprehension data of each student.

6. The system according to claim 1, characterized in that The method of dynamically adjusting the personalized teaching strategy for each student based on the comprehension data and the preset teaching strategy matching library specifically includes: Periodically collect students’ interaction records and homework information on the online learning platform to form a learning behavior log; According to the learning behavior log, each student's mastery of different knowledge points is clustered by a clustering algorithm to form a student portrait; Associating the student portraits with different teaching strategies, determining personalized teaching strategy combinations for different student groups, and forming a teaching strategy matching library; The comprehension data is matched through the teaching strategy matching library to determine the personalized teaching strategy for each student.

7. The system according to any one of claims 1 to 6, characterized in that The system further comprises: The live broadcast parameter adjustment module is used to obtain the scene parameter configuration data stream corresponding to the current live broadcast room type according to the pre-established live broadcast room type and scene parameter mapping database, and adjust the background parameters and sound effect parameters of the live broadcast room through the scene parameter configuration data stream.

8. The system according to any one of claims 1 to 6, characterized in that The system further comprises: The teaching effect analysis module is used to analyze the association relationship between the teaching content, interaction mode and scene parameters of multiple students in the live interaction process by using the association rule mining algorithm, and obtain multiple association rules between the teaching content, interaction mode and scene parameters; By comparing the student participation corresponding to each association rule with the preset participation threshold, the effective association rule set and the invalid association rule set are determined to form a teaching effect analysis report.

9. The system according to any one of claims 1 to 6, characterized in that The system further comprises: A cross-platform publishing module is used to perform transcoding operations on live source data according to pre-established encoding rules to obtain live streams adapted to different educational platform formats. The transcoding operations include video encoding format conversion, audio encoding format conversion, and resolution adjustment. According to the platform interface protocol, live streams in different formats are distributed to the corresponding education platform interface to achieve multi-platform simultaneous live broadcast and sharing of teaching resources.

Images