Classroom diagnosis analysis method based on deep learning

Abstract

The invention discloses a classroom diagnosis analysis method based on deep learning, and belongs to the technical field of data processing, and the method comprises the steps: obtaining a plurality of classroom voices, and cutting the classroom voices into a plurality of voice segments; converting the voice segment into a language text by using a first deep learning model, cutting the language text into a plurality of theme texts, and labeling a first label in the theme texts; extracting a teaching summary in each subject text based on a second deep learning model, matching a corresponding first knowledge point for the teaching summary, and labeling the first knowledge point in the subject text as a second label; counting second error frequencies of the first knowledge points, and acquiring first N first knowledge points with higher second error frequencies as second knowledge points; and matching a theme text corresponding to the second knowledge point based on the second label, calculating a score of each evaluation index in the theme text, and generating an improvement suggestion according to the scores. According to the invention, the classroom teaching condition of the teacher can be comprehensively and deeply analyzed and evaluated.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a classroom diagnosis and analysis method based on deep learning. Background Art

[0002] Traditional classroom teaching evaluation mostly adopts methods such as asking questions, feedback on homework, and distributing questionnaires. However, these methods have disadvantages such as subjectivity, lag, and incompleteness. With the emergence of deep learning technology, by collecting image data and voice data of each classroom teaching and analyzing them using deep learning technology, it is convenient to evaluate the teaching quality.

[0003] For example, a Chinese patent document with a publication number of CN116976317A discloses a classroom discourse analysis method, which uses the FILD framework for observation and evaluation. By counting the number of teachers' and students' discourses and the total number of classroom discourses, the frequency of discourse in the classroom can be quantified. These indicators can be used to analyze teaching models and student participation, helping teachers understand the distribution and balance of classroom discourse. By calculating the amount of information and information contribution ratio of teachers' and students' discourses, the richness of information in the classroom and the effect of information transmission can be evaluated. For another example, a Chinese patent document with a publication number of CN117195892A discloses a classroom teaching evaluation method and system based on big data. The method first obtains the teacher's teaching video and student interaction information, and then converts it into visual data, and then analyzes the connection between the teacher's teaching content and the student's interactive content based on the visual data, thereby obtaining evaluation indicators of the interaction frequency and activity level of the class, and then evaluates the quality of classroom teaching based on the teaching quality evaluation indicators.

[0004] The above-mentioned existing technologies all evaluate the classroom teaching quality based on the classroom interaction situation, and the evaluation dimension is relatively single, so there is a problem that the teaching quality cannot be evaluated in an all-round way. Summary of the invention

[0005] To solve the above problems, the present invention provides a classroom diagnosis and analysis method based on deep learning to achieve a comprehensive evaluation of teaching quality.

[0006] In order to achieve the above-mentioned purpose of the invention, the present invention proposes a classroom diagnosis and analysis method based on deep learning, comprising: Acquire multiple classroom voices, locate turning points in each classroom voice, and cut the classroom voice into multiple voice segments based on the turning points, each of the voice segments having a first label; Using a first deep learning model to convert the speech segment into a language text, cutting the language text into a plurality of subject texts, and marking the first label in the subject text according to the speech segment to which the subject text belongs; Extracting the teaching summary in each of the subject texts based on the second deep learning model, generating a plurality of first knowledge points based on the teaching materials, matching the teaching summary with the corresponding first knowledge points, and marking them as second labels in the subject texts; Analyze the first knowledge point tested by each test question in the test paper, calculate the second error frequency of the first knowledge point based on the first error frequency of the test question, and obtain the top N first knowledge points with higher second error frequencies as second knowledge points; Based on the second tag matching the subject text corresponding to the second knowledge point, multiple evaluation indicators are preset, the score of each evaluation indicator in the subject text is calculated, and improvement suggestions are generated according to the score.

[0007] Further, locating the turning point in the classroom speech comprises the following steps: Splitting the classroom speech into a plurality of speech intervals, calculating a first score for each of the speech intervals, generating a fluctuation curve of the first score changing over time, selecting a plurality of reference points at intervals on the fluctuation curve, and obtaining the slopes of the reference points; Set multiple numerical intervals, divide the slope into corresponding numerical intervals according to the value of the slope, count the number of coverages including the slope in the numerical intervals, obtain M numerical intervals with smaller coverage numbers, define them as target intervals, select X slopes from the target intervals under a constraint condition, the constraint condition is that the horizontal coordinate interval of the corresponding coordinate points of the selected slopes is greater than a first threshold, determine the corresponding reference point based on the selected slope, obtain the first time point when the reference point appears in the classroom speech, and define the first time point as the turning point.

[0008] Furthermore, drawing the fluctuation curve comprises the following steps: Counting a first number of voiceprints appearing in each of the voice intervals, generating a corresponding speech text based on each voiceprint, counting a second number of characters included in the speech text, setting a character limit for the voice interval, and calculating the first score corresponding to the voice interval based on the first number, the second number, and the character limit; A coordinate system is established with the speech duration of the classroom speech as the horizontal axis and the first score as the vertical axis to locate the existence interval of each speech interval in the speech duration, and coordinate points corresponding to each speech interval are drawn in the coordinate system. The vertical coordinate of the coordinate point is the first score corresponding to the speech interval, and the horizontal coordinate is within the existence interval of the speech interval. The coordinate points are fitted to generate the fluctuation curve.

[0009] Furthermore, cutting the language text into the subject text comprises the following steps: A first dictionary library is set, wherein the first dictionary library includes a plurality of keywords, and the language text is divided into a plurality of initial language blocks by taking the keywords appearing in the language text as segmentation points, and the initial language blocks adjacent to each other are defined as a first language block and a second language block, respectively. If a first word and a second word exist in the first language block and the second language block, and a distribution rule of the first word and the second word in the first language block and the second language block meets a preset condition, the first language block is retained as the subject text, and the second language block and the language block located thereafter are continuously compared, otherwise the first language block and the second language block are re-cut based on the distribution rule until the distribution rule meets the preset condition.

[0010] Furthermore, the distribution rule includes a first frequency and a second frequency of the first word appearing in the first language block and the second language block, and a third frequency and a fourth frequency of the second word appearing in the first language block and the second language block. If the absolute value of the difference between the first frequency and the second frequency, and the absolute value of the difference between the third frequency and the fourth frequency are both greater than a second threshold, it is determined that the distribution rule meets the preset condition.

[0011] Furthermore, during re-segmentation, the first language block and the second language block are merged into a combined language block. If there is a segmentation point in the combined language block, after the combined language block is segmented based on the segmentation point to obtain a third language block and a fourth language block, if the first words and the second words satisfying the distribution rule exist in the third language block and the fourth language block, the third language block is retained as the subject text. If there is no segmentation point, the combined language block and the language blocks following it continue to be compared.

[0012] Further, extracting the teaching summary includes the following steps: The second deep learning model is a neural network based on the Transformer mechanism, which splits the subject text into multiple first independent sentences, generates an initial summary in the subject text based on the second deep learning model, obtains the first independent sentence corresponding to each second independent sentence in the initial summary, calculates the second score of the first independent sentence according to the second independent sentence, filters the second score greater than a third threshold, extracts the corresponding first independent sentences and combines them into the teaching summary.

[0013] Further, obtaining the first independent statement corresponding to the second independent statement includes the following steps: Obtain the weights of each of the first independent statements used by the second deep learning model to generate the second independent statement, and define the first independent statement corresponding to the largest weight as corresponding to the second independent statement.

[0014] Furthermore, matching the first knowledge point corresponding to the teaching summary includes the following steps: A second dictionary library is established to extract multiple focus words in the teaching summary, and the superordinate words of the focus words are obtained based on the second dictionary library. The subject text is grouped based on the superordinate words to obtain multiple first categories, and in each of the first categories, the subject text is grouped again into second categories based on the focus words. When matching the first knowledge point, the superordinate words corresponding to the first knowledge point are first matched to locate the corresponding first category, and then the second category is matched according to the focus words.

[0015] Furthermore, the first deep learning model is an LSTM model.

[0016] Compared with the prior art, the beneficial effects of the present invention are at least as follows: The present invention locates the turning points in classroom speech, thereby accurately dividing classroom speech into two types of speech segments: explanation and discussion, which helps to analyze classroom interaction and teaching content in more detail later. The summary of classroom content is automatically extracted by the second deep learning model, reducing the workload of manual processing. By analyzing the error frequency in the test paper and identifying the knowledge points where students are prone to make mistakes, only the subject text that explains the knowledge point can be located, and only the located text can be analyzed, thereby further improving the pertinence of the analysis process and reducing the workload of analysis. By presetting multiple evaluation indicators to comprehensively evaluate the classroom content and provide teachers with multi-dimensional teaching feedback, it is helpful for teachers to optimize in terms of explanation time, interaction frequency, etc., thereby improving students' participation and knowledge reception. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1This is a flowchart of the steps of classroom diagnostic analysis based on deep learning in the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0020] like Figure 1 As shown, a classroom diagnostic analysis method based on deep learning includes: S1: Acquire multiple classroom voices, locate turning points in each classroom voice, and cut the classroom voice into multiple voice segments based on the turning points, each voice segment having a first label.

[0021] During offline teaching, a microphone can be installed in the classroom to record classroom voice. During online teaching, classroom voice can be recorded through a mobile terminal used for teaching. This embodiment takes a complete teaching process as a classroom voice, which includes the voice of the teacher and the voice of the students. For each classroom voice, a turning point is located in it. The turning point in this embodiment is used to separate the part where the teacher teaches individually and the part where the group discusses. After segmentation, the first label is marked in the voice segment. The first label includes an explanation label and a discussion label. The voice segment with the explanation label is the part where the teacher explains the class individually, and the voice segment with the discussion label is the voice segment where the teacher interacts with the students, or where the students interact and discuss with each other.

[0022] S2: Use the first deep learning model to convert the speech segment into language text, cut the language text into multiple topic texts, and mark the first label in the topic text according to the speech segment to which the topic text belongs.

[0023] In this embodiment, the first deep learning model is an LSTM model.

[0024] The existing trained LSTM model can be used as the first deep learning model. The full name of the LSTM model is long short-term memory network, which is particularly suitable for processing natural text sequence data; the segmented voice segment is input into the first deep learning model, and the first deep learning model converts it into a language text. Here, Chinese and physics knowledge content are used as examples for explanation. In other embodiments, the language text can also be English, Japanese, etc. Since the content of a class may include multiple knowledge points, such as "understanding movement" and "time displacement", or the teacher has multiple knowledge points because of explaining the test paper, the language text is cut into multiple subject texts here to facilitate more detailed analysis and diagnosis later; afterwards, according to the first label of the voice segment from which the subject text comes, the same first label is marked in the subject text.

[0025] S3: Extract the teaching summary in each subject text based on the second deep learning model, generate multiple first knowledge points based on the teaching materials, match the corresponding first knowledge points for the teaching summary, and annotate them in the subject text as the second label.

[0026] The second deep learning model is constructed based on the neural model of the Transformer mechanism. The Transformer model is a deep learning model for processing sequence data. The trained Transformer model is used as the second deep learning model, and the subject text is input into the second deep learning model. The second deep learning model extracts and generates a relatively brief teaching summary, which includes more important keywords. In addition, the first knowledge point is generated according to the teaching materials. Specifically, it can be generated according to the title of the textbook catalog, and matched according to the keywords included in the catalog title and the teaching summary. For example, if the teaching summary includes the decomposition of force, the corresponding knowledge point can be matched to "decomposition and synthesis of force", and "decomposition and synthesis of force" is marked as the second label in the subject text.

[0027] S4: Analyze the first knowledge point tested by each test question in the test paper, calculate the second error frequency of the first knowledge point based on the first error frequency of the test question, and obtain the top N first knowledge points with higher second error frequencies as the second knowledge points.

[0028] S5: Matching the subject text corresponding to the second knowledge point based on the second label, presetting multiple evaluation indicators, calculating the score of each evaluation indicator in the subject text, and generating improvement suggestions based on the scores.

[0029] The knowledge points tested by each test question in the test paper can be manually marked, and then the error frequency of the first knowledge point can be determined by counting the wrong questions of the students in the class of the instructor. For example, there are 5 errors in test A and 6 errors in test B, that is, the corresponding first error frequencies are 5 and 6 respectively. Test A and test B both test the first knowledge point Q, and the second error frequency of the first knowledge point Q is 11 times. Set N to 3. After the statistics are completed, select the first 3 first knowledge points with higher second error frequencies as the second knowledge points. The second knowledge points are error-prone knowledge points.

[0030] Find the second label containing the second knowledge point, and then obtain the corresponding subject text. In particular, a second knowledge point may correspond to multiple subject texts. For example, new knowledge explanation, example explanation, mutual discussion, etc. will be marked with the same second label. The subject text is analyzed by setting a variety of evaluation indicators. During the analysis, the comprehensive evaluation of the explanation content can be performed through the pre-set evaluation indicators. The evaluation indicators of this embodiment include the length of the explanation and the number of interactions. For the scoring of the length of the explanation, a standard number of characters can be set. When the number of characters in the subject text is greater than the standard number of characters, it is full marks. When it is less than the standard number of characters, the ratio of the number of characters in the subject text to the standard number of characters is used as the score; the number of interactions can be determined according to the number of subject texts that match the first label as the discussion label; in addition, indicators such as the number of occurrences of related knowledge points and the number of example explanations can also be included. The specific scoring method can be set according to the actual situation.

[0031] The present invention locates the turning points in classroom speech, thereby accurately dividing classroom speech into two types of speech segments: explanation and discussion, which helps to analyze classroom interaction and teaching content in more detail later. The summary of classroom content is automatically extracted by the second deep learning model, reducing the workload of manual processing. By analyzing the error frequency in the test paper and identifying the knowledge points where students are prone to make mistakes, only the subject text that explains the knowledge point can be located, and only the located text can be analyzed, thereby further improving the pertinence of the analysis process and reducing the workload of analysis. By presetting multiple evaluation indicators to comprehensively evaluate the classroom content and provide teachers with multi-dimensional teaching feedback, it is helpful for teachers to optimize in terms of explanation time, interaction frequency, etc., thereby improving students' participation and knowledge reception.

[0032] It is particularly noteworthy that the present invention can conduct in-depth analysis and evaluation of teachers' classroom teaching situations in all aspects, making teaching activities more scientific and precise.

[0033] In this embodiment, locating the turning point in classroom speech includes the following steps: Split the classroom voice into multiple voice intervals, calculate the first score for each voice interval, generate a fluctuation curve showing how the first score changes over time, select multiple reference points at intervals on the fluctuation curve, and obtain the slopes of the reference points.

[0034] Specifically, the steps for drawing the fluctuation curve in this embodiment are as follows: Count the first number of voiceprints that appear within each voice interval, generate corresponding speech texts for each voiceprint as a unit, count the second number of characters included in the speech texts, set a character limit for the voice interval, and calculate the first score for the corresponding voice interval based on the first number, the second number, and the character limit.

[0035] Establish a coordinate system with the voice duration of the classroom voice as the horizontal axis and the first score as the vertical axis, locate the existence interval of each voice interval within the voice duration, draw coordinate points corresponding to each voice interval in the coordinate system, where the vertical coordinate of the coordinate point is the first score of the corresponding voice interval, and the horizontal coordinate is within the existence interval of the voice interval. Fit the coordinate points to generate a fluctuation curve.

[0036] In this embodiment, the length of each voice interval is 2 minutes. When the voice duration of the classroom voice is 40 minutes, the classroom voice is split into 20 voice intervals. Then, calculate the first score for each voice interval. After the calculation is completed, draw a fluctuation curve showing how the voice intervals change over time. When calculating the first score of a voice interval, first obtain the number of voiceprints included in the voice interval, and voiceprint recognition can be performed based on the first deep learning model. Since each person has a unique voiceprint, the number of voiceprints represents the number of speakers within the voice interval. Then, count the second number of characters included in the speech text, and calculate the first score of the voice interval through the first formula. The first formula is: , where P represents the first score, K represents the first number corresponding to the voice interval, Ei represents the second number of the speech text corresponding to the i-th voiceprint, and L represents the character limit, which is set to 300 in this embodiment.

[0037] After that, establish a coordinate system with the voice duration of the classroom voice as the horizontal coordinate axis and the first score as the vertical coordinate axis. When drawing the coordinate points, since the length of each voice interval is 2 minutes, the middle of each voice interval is selected as the horizontal coordinate of the coordinate point, that is, the horizontal coordinate is within the existence interval of the voice interval. For example, if the voice interval corresponds to the 21st - 22nd minutes in the classroom voice, 21 - 22 minutes is the existence duration of this voice interval, and the midpoint 21.5 minutes is selected as the horizontal coordinate of the corresponding coordinate point, and the first score of this voice interval is used as the vertical coordinate of the coordinate point. Then, fit each coordinate point through the least squares method to generate the corresponding fluctuation curve. Finally, select multiple reference points on the fluctuation curve, the selection interval between the reference points is 1 minute, and calculate the slope of each reference point on the fluctuation curve.

[0038] Set multiple numerical intervals, divide the slopes into corresponding numerical intervals according to their values, count the number of coverages including the slopes in the numerical intervals, obtain M numerical intervals with smaller coverages, define them as target intervals, select X slopes from the target intervals under constraints, the constraint condition is that the horizontal coordinate intervals of the coordinate points corresponding to the selected slopes are greater than a first threshold, determine the corresponding reference points based on the selected slopes, obtain the first time point when the reference points appear in the classroom speech, and define the first time point as a turning point.

[0039] In a classroom scenario, when a discussion is needed, it is usually accompanied by the teacher calling the roll or waiting for the hand to be raised, that is, the idle period. According to the first formula, the first score calculated for the corresponding voice interval will become smaller at this time. For example, voice intervals 1-5 are for individual teaching by the teacher, voice interval 6 is for the waiting stage for hand raising, and voice intervals 7-9 are for the discussion and answering stage. The first score is larger in voice intervals 1-5, and the values ​​in each voice interval are relatively close. It is smaller in voice interval 6 and becomes larger in voice intervals 7-9, and a trough will appear in the fluctuation curve. Or, voice intervals 6-9 are all waiting for hand raising stages, and the fluctuation curve will tilt downward and then stabilize in a certain value range.

[0040] To this end, this embodiment proposes the following method: first, set multiple numerical intervals, such as [-5,5), [-15,-5), and divide them into corresponding numerical intervals according to the value of the slope. Then, count the number of coverages including the slope in the numerical interval. According to the previous analysis, discussions will not appear frequently in the classroom. Therefore, the intervals with fewer slopes are selected as target intervals. For example, the above numerical interval [-5,5) includes 4 slopes, namely -1, -2, -2, and -3. The first threshold is set to 3min. -1 and -2 are adjacent reference points with an interval of 1min, which is equal to the first threshold and does not meet the constraint condition. Therefore, only one slope is selected from them, and the first time point at which the slope corresponds to the reference point is the turning point. The horizontal coordinate between the two reference points corresponding to the slopes of -2 and -3 is greater than 1min, which meets the constraint condition, so both slopes will be extracted. The purpose of this step is to find the trough or the position where the slope changes greatly in the fluctuation curve, and select a reference point from it as the turning point.

[0041] In this embodiment, cutting the language text into subject texts includes the following steps: A first dictionary library is set, and the first dictionary library includes multiple keywords. The keywords appearing in the language text are used as segmentation points to segment the language text into multiple initial language blocks. The initial language blocks adjacent to each other are defined as the first language block and the second language block. If the first word and the second word exist in the first language block and the second language block, and the distribution pattern of the first word and the second word in the first language block and the second language block meets the preset conditions, the first language block is retained as the subject text, and the second language block and the language block located thereafter are continuously compared. Otherwise, the first language block and the second language block are re-segmented based on the distribution pattern until the distribution pattern meets the preset conditions.

[0042] Among them, the distribution pattern includes the first frequency and the second frequency of the first word appearing in the first language block and the second language block, and the third frequency and the fourth frequency of the second word appearing in the first language block and the second language block. If the absolute value of the difference between the first frequency and the second frequency, and the absolute value of the difference between the third frequency and the fourth frequency are both greater than the second threshold, it is judged that the distribution pattern meets the preset conditions.

[0043] Specifically, the first dictionary includes a plurality of pre-set keywords, such as "next", "next knowledge point", etc. Such keywords can effectively segment the explanation process of the language text. Assume that the language text is currently segmented into three language blocks based on the keyword detection method, wherein the first two language blocks are defined as the first language block and the second language block, respectively. First, the first language block and the second language block are analyzed. The first language block and the second language block are segmented to obtain a plurality of words, and the frequency of each word appearing in the first language block and the second language block is counted. If there is such a group of words, which are defined as the first word and the second word, for example, the first word is friction, and the second word is gravity, the first word appears more frequently in the first language block, but less frequently in the second language block, and the second word appears less frequently in the first language block, but more frequently in the second language block, indicating that there is a significant difference in the distribution of the first word and the second word in the two language blocks, that is, the first language block and the second language block correspond to different knowledge points, respectively. It is judged that the first word and the second word meet the preset conditions, so the division of the first language block can be retained. Then, by analogy, continue to analyze the second block and its subsequent blocks to ensure that the word distribution patterns of each pair of adjacent blocks meet the preset conditions.

[0044] If the first word and the second word do not exist in the first block and the second block, that is, the words in the first block and the second block are evenly distributed, the first block and the second block are split again.

[0045] When re-segmenting, the first block and the second block are merged into a combined block. If there is a segmentation point in the combined block, the combined block is segmented based on the segmentation point to obtain the third block and the fourth block. If the third block and the fourth block contain the first word and the second word that satisfy the distribution rule, the third block is retained as the subject text. If there is no segmentation point, the combined block and the blocks following it are continued to be compared.

[0046] First, the first block and the second block are merged into a combined block, and the frequency of occurrence of each word in the combined block is counted. Then, the end of each sentence is used as a potential segmentation point, and the combined block is tried to be segmented multiple times from the beginning to the end. If a suitable segmentation point is found, the combined block is segmented into the third block and the fourth block at the segmentation point. The first word and the second word described above exist in the third block and the fourth block, and the distribution pattern meets the preset conditions, then the third block is retained as a topic text. If no segmentation point that meets the conditions is found after traversing the combined block, the original division of the combined block is maintained. Then, the combined block and the subsequent blocks continue the above process until the last block after keyword segmentation is reached.

[0047] In this embodiment, extracting the teaching summary includes the following steps: The second deep learning model is a neural network based on the Transformer mechanism. It splits the topic text into multiple first independent sentences, generates an initial summary in the topic text based on the second deep learning model, obtains the first independent sentence corresponding to each second independent sentence in the initial summary, calculates the second score of the first independent sentence based on the second independent sentence, screens the second score greater than the third threshold, extracts the corresponding first independent sentences and combines them into a teaching summary.

[0048] Obtaining the first independent statement corresponding to the second independent statement includes the following steps: The weights of the first independent statements used by the second deep learning model to generate the second independent statement are obtained, and the first independent statement corresponding to the largest weight is defined as corresponding to the second independent statement.

[0049] Specifically, the Transformer model is first trained using a training set, which includes multiple topic texts, and each topic text corresponds to an initial summary to be generated. After the training is completed, the topic text is input into the Transformer model as a whole, and the Transformer model outputs an initial summary of the topic text. The initial summary is often different from the topic text, and is a condensation and summary of the meaning in the topic text. Since the present invention is applied to teaching scenarios, it will involve many professional knowledge terms, and the final output result of the Transformer model depends on the result of adding the topic text embedding vector to the attention head output vector, which is defined here as the result vector, and then the result vector is converted into text characters through the vector dictionary. In high-dimensional space, a slight deviation in the calculated result vector may output text characters with similar meanings but different names.

[0050] To solve the above problems, after generating the initial summary, the present invention splits the initial summary into multiple first independent sentences, splits the subject text into multiple second independent sentences, and then determines the correlation between the first independent sentence and the second independent sentence, that is, which second independent sentence generates the first independent sentence. When determining, the text similarity of the first independent sentence and the second independent sentence can be compared two by two, or when the Transformer model generates the first independent sentence, the weight of each second independent sentence in the first independent sentence is determined, and the first independent sentence with the largest weight is defined as corresponding to the second independent sentence. In addition, this embodiment calculates the second score through the second formula, and the second formula is: , where Y is the second score of the first independent sentence, R1 is the number of repeated characters between the first independent sentence and the second independent sentence, and R2 is the number of characters in the first independent sentence. The teaching summaries generated by the above method are all derived from the original subject text, so there will be no professional terminology errors.

[0051] In this embodiment, matching the first knowledge point corresponding to the teaching summary includes the following steps: A second dictionary library is established to extract multiple focus words in the teaching summary, and the superordinate words of the focus words are obtained based on the second dictionary library. The subject texts are grouped based on the superordinate words to obtain multiple first categories, and the subject texts are grouped into second categories again based on the focus words in each first category. When matching the first knowledge point, the superordinate words corresponding to the first knowledge point are first matched to locate the corresponding first category, and then the second category is matched according to the focus words.

[0052] The second dictionary library is manually established in advance, which includes multiple hypernyms, for example, the hypernym of matrix is ​​linear algebra. When obtaining the focus words in the teaching summary, the words with the highest frequency of occurrence can be extracted as the focus words, and then the corresponding hypernyms are obtained from the hypernym dictionary library based on the focus words. According to the hypernyms in each teaching summary, the subject text is divided into multiple first categories to achieve a preliminary division of the subject text. After that, the first category is further divided into multiple second categories according to the focus words; in the final matching, for example, if the first knowledge point is a set, then the subject text under the first category of function will be matched first, and then a more accurate match will be performed based on the word set, thereby greatly improving the matching speed.

[0053] It should be understood that the various technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A classroom diagnostic analysis method based on deep learning, characterized in that: include: Acquire multiple classroom voices, locate turning points in each classroom voice, and cut the classroom voice into multiple voice segments based on the turning points, each of the voice segments having a first label; Using a first deep learning model to convert the speech segment into a language text, cutting the language text into a plurality of subject texts, and marking the first label in the subject text according to the speech segment to which the subject text belongs; Extracting the teaching summary in each of the subject texts based on the second deep learning model, generating a plurality of first knowledge points based on the teaching materials, matching the teaching summary with the corresponding first knowledge points, and marking them as second labels in the subject texts; Analyze the first knowledge point tested by each test question in the test paper, calculate the second error frequency of the first knowledge point based on the first error frequency of the test question, and obtain the top N first knowledge points with higher second error frequencies as second knowledge points; Based on the second tag matching the subject text corresponding to the second knowledge point, multiple evaluation indicators are preset, the score of each evaluation indicator in the subject text is calculated, and improvement suggestions are generated according to the score.

2. The method according to claim 1, characterized in that Locating the turning point in the classroom speech comprises the following steps: Splitting the classroom speech into a plurality of speech intervals, calculating a first score for each of the speech intervals, generating a fluctuation curve of the first score changing over time, selecting a plurality of reference points at intervals on the fluctuation curve, and obtaining the slopes of the reference points; Set multiple numerical intervals, divide the slope into corresponding numerical intervals according to the value of the slope, count the number of coverages including the slope in the numerical intervals, obtain M numerical intervals with smaller coverage numbers, define them as target intervals, select X slopes from the target intervals under a constraint condition, the constraint condition is that the horizontal coordinate interval of the corresponding coordinate points of the selected slopes is greater than a first threshold, determine the corresponding reference point based on the selected slope, obtain the first time point when the reference point appears in the classroom speech, and define the first time point as the turning point.

3. The method according to claim 2, characterized in that Drawing the fluctuation curve comprises the following steps: Counting a first number of voiceprints appearing in each of the voice intervals, generating a corresponding speech text based on each voiceprint, counting a second number of characters included in the speech text, setting a character limit for the voice interval, and calculating the first score corresponding to the voice interval based on the first number, the second number, and the character limit; A coordinate system is established with the speech duration of the classroom speech as the horizontal axis and the first score as the vertical axis to locate the existence interval of each speech interval in the speech duration, and coordinate points corresponding to each speech interval are drawn in the coordinate system. The vertical coordinate of the coordinate point is the first score corresponding to the speech interval, and the horizontal coordinate is within the existence interval of the speech interval. The coordinate points are fitted to generate the fluctuation curve.

4. The method according to claim 1, characterized in that: Cutting the language text into the subject text comprises the following steps: A first dictionary library is set, wherein the first dictionary library includes a plurality of keywords, and the language text is divided into a plurality of initial language blocks by taking the keywords appearing in the language text as segmentation points, and the initial language blocks adjacent to each other are defined as a first language block and a second language block, respectively. If a first word and a second word exist in the first language block and the second language block, and a distribution rule of the first word and the second word in the first language block and the second language block meets a preset condition, the first language block is retained as the subject text, and the second language block and the language block located thereafter are continuously compared, otherwise the first language block and the second language block are re-cut based on the distribution rule until the distribution rule meets the preset condition.

5. The method according to claim 4, characterized in that The distribution rule includes a first frequency and a second frequency of the first word appearing in the first language block and the second language block, and a third frequency and a fourth frequency of the second word appearing in the first language block and the second language block. If the absolute value of the difference between the first frequency and the second frequency, and the absolute value of the difference between the third frequency and the fourth frequency are both greater than a second threshold, it is determined that the distribution rule meets the preset condition.

6. The method according to claim 4, characterized in that During re-segmentation, the first language block and the second language block are merged into a combined language block. If a segmentation point exists in the combined language block, the combined language block is segmented based on the segmentation point to obtain a third language block and a fourth language block. If the first words and the second words satisfying the distribution rule exist in the third language block and the fourth language block, the third language block is retained as the subject text. If the segmentation point does not exist, the combined language block and the language blocks following it are continuously compared.

7. The method according to claim 1, characterized in that Extracting the teaching summary includes the following steps: The second deep learning model is a neural network based on the Transformer mechanism, which splits the subject text into multiple first independent sentences, generates an initial summary in the subject text based on the second deep learning model, obtains the first independent sentence corresponding to each second independent sentence in the initial summary, calculates the second score of the first independent sentence according to the second independent sentence, filters the second score greater than a third threshold, extracts the corresponding first independent sentences and combines them into the teaching summary.

8. The method according to claim 7, characterized in that Acquiring the first independent statement corresponding to the second independent statement includes the following steps: Obtain the weights of each of the first independent statements used by the second deep learning model to generate the second independent statement, and define the first independent statement corresponding to the largest weight as corresponding to the second independent statement.

9. The method according to claim 1, characterized in that: Matching the first knowledge point corresponding to the teaching summary includes the following steps: A second dictionary library is established to extract multiple focus words in the teaching summary, and the superordinate words of the focus words are obtained based on the second dictionary library. The subject text is grouped based on the superordinate words to obtain multiple first categories, and in each of the first categories, the subject text is grouped again into second categories based on the focus words. When matching the first knowledge point, the superordinate words corresponding to the first knowledge point are first matched to locate the corresponding first category, and then the second category is matched according to the focus words.

10. The method according to claim 1, characterized in that The first deep learning model is an LSTM model.

Images