Automatic grading method and system for physical, chemical and student practical operation test based on visual language model
By adopting an automatic scoring system based on visual language model in the physical, chemical and biology practical examination, combining visual features and text description for practical behavior recognition and scoring, the problems of poor recognition ability and low scoring credibility in the prior art are solved, and higher scoring accuracy and credibility are achieved.
Patent Information
- Application Number
- CN202411801174.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-04-01
AI Technical Summary
The existing automatic scoring system for physical, chemical and biology practical examinations has poor recognition ability with few samples and zero samples, which is prone to misjudgment, resulting in poor scoring accuracy and low credibility.
An automatic scoring system based on visual language model is adopted, including practical behavior recognition module, scoring module and confidence evaluation module. The visual language model combines visual features and text descriptions to identify and score practical behaviors, and improves the credibility of scores through the confidence evaluation module.
It improves the scoring accuracy and credibility of practical examinations, reduces manual intervention, and ensures fairness and consistency of scoring.
Smart Images
Figure CN119740919A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of experimental operation detection, and in particular relates to an automatic scoring method and system for practical operation examinations in physics, chemistry and biology based on a visual language model. Background Art
[0002] In traditional physics, chemistry and biology practical exams, grading is usually done manually, which not only requires a lot of manpower and time, but is also easily affected by subjective factors, resulting in the fairness and consistency of the grading results being questioned. At the same time, students cannot obtain the grading results and detailed feedback information in a timely manner, making it difficult to improve and enhance their experimental operation capabilities. The existing automated grading system uses deep learning models as the core to build the system, which reduces manpower requirements, improves the efficiency of grading, and can provide timely feedback to students. However, the robustness of deep learning models is low, and their recognition ability is poor in the case of few samples and zero samples. They are prone to misjudgment, which affects the scores of candidates. At the same time, scoring errors may occur, so it is necessary to conduct confidence assessment on the scores to improve the credibility and persuasiveness of the scores. In the field of artificial intelligence, the visual language model is a new model that combines computer vision and natural language processing technology. It can understand the correspondence between images and texts. Compared with deep learning models, it has a stronger ability to handle few samples and zero samples, and is more suitable for the diversity and complexity of experimental operations. Its output results are more accurate and can be used for confidence assessment of scores. Summary of the invention
[0003] In view of the problems of poor scoring accuracy and low scoring credibility in current practical examinations, the present invention provides a method and system for automatic scoring of practical examinations in physics, chemistry and biology based on a visual language model.
[0004] In order to achieve the above object, the present invention adopts the following technical solutions:
[0005] An automatic scoring system for practical exams in physics, chemistry and biology based on a visual language model, including a practical behavior recognition module, a scoring module and a confidence evaluation module;
[0006] The practical behavior recognition module is used to analyze and recognize the examinee's behavior;
[0007] The scoring module is used to compare the candidate's experimental operation with the standard experimental operation and score the candidate's practical test;
[0008] The confidence evaluation module is used to evaluate the rationality of the practical test score.
[0009] Furthermore, the practical operation behavior recognition module is implemented by a visual language model, and the visual language model includes: a visual encoder layer, a projection layer layer and an LLM layer;
[0010] The Visual encoder layer is used to convert visual information into a representation that can be understood and processed by the model, extract features from the original visual data, learn the potential representation of visual information, and enhance the visual perception ability of the model; the Visual encoder layer uses the pre-trained visual encoder CLIP-ViT-L / 14 to extract visual features of each frame.
[0011] The Projection layer is used to map the features extracted by the visual encoder into a representation space that matches the language model. The Projection layer maps the visual feature vector to a lower-dimensional vector space by learning a weight matrix, so that the visual features and language features can be better aligned and interacted.
[0012] The LLM layer is used to learn the representation of visual information provided by the visual encoder layer, obtain the ability to generate tasks, and generate corresponding text outputs based on the input image. After LLM identifies a key experimental step of the examinee's practice through the visual feature vector, it outputs a text description of the step. Behaviors not related to the experiment are not described in text, and a record file is generated to record the location information of the behavior in the video clip.
[0013] Furthermore, the behaviors identified by the practical behavior identification module include: examinee's practical test behavior, cheating behavior and behavior unrelated to the test.
[0014] Furthermore, the scoring module includes an experiment comparison unit and a scoring unit; the experiment comparison unit is used to compare the video and text description of the candidate's experimental operation with the video and text description of the standard experimental operation, and use the difference between the candidate's experimental operation and the standard experimental operation as the basis for scoring; the scoring unit integrates the candidate's gain and loss points based on the difference between the visual language features, and then generates the corresponding score based on the scoring rules. The experiment comparison unit not only focuses on the specific practical operations in the video and text description of the standard experimental operation, but also pays attention to the global semantic information to ensure that when understanding each specific experimental step, the significance of each experimental step in the entire experimental process can also be considered.
[0015] The automatic scoring method for physics, chemistry and biology practical examination based on visual language model includes the following steps:
[0016] Step 1: film the process of students' practical examination from multiple perspectives and save the video data;
[0017] Step 2: Identify the student's practical test behavior through the practical behavior recognition module; send the identified video clips and marked text descriptions to the scoring module;
[0018] Step 3: The scoring module compares the features of the videos, videos and texts, and texts and removes the features of the video clips that are not relevant to the test based on the marks, records the reasons for the examinees' loss of points, and scores the examinees' practical test.
[0019] Step 4, using the confidence evaluation module, the rationality of the practical test score is calculated based on the representation characteristics of the candidate's practical test, the representation characteristics of the standard practical test, and the test score;
[0020] Step 5: The teacher will correct the practical scores of candidates with low confidence scores. For the scores questioned by the candidates, the teacher will re-correct them through manual verification. Candidates can check their scores immediately after the practical test, and can also check the reasons for their lost scores and get feedback in time. The use of confidence assessment method reflects the fairness of scoring, reduces the influence of subjective factors of the invigilator, and reflects the fairness of the examination.
[0021] In the step 2, the practical test behavior of the students is identified through the practical behavior identification module. When cheating behavior of the students is detected, the video clip containing the cheating behavior is marked, the cheating behavior is recorded in the text description, and the test score is recorded as 0 points; when the students are detected to have behavior unrelated to the test, the video clip containing the unrelated behavior is marked, and the behavior unrelated to the test is not recorded in the text description.
[0022] The step 2 is to identify the student's practical test behavior through the practical behavior recognition module; send the identified video clips and marked text descriptions to the scoring module; specifically:
[0023] The practical behavior recognition module divides the video saved in step 1 into several segments, each of which represents a key experimental step, aggregates the significant features in each segment, and generates a compact representation of each step;
[0024] Generate corresponding representations according to the time sequence of the experimental steps to maintain the coherence of the practical process, so that the model can understand and recognize the actual process of the practical test; at the same time, pay attention to the global semantic information representing the entire video, combine the global information with the characteristics of each clip, and ensure that when understanding each specific experimental step, the significance of each experimental step in the entire experimental process is taken into account;
[0025] The calculation formula of the feature V of the video clip of the candidate's practical test is: V = Visual encoder (AV); the calculation formula of the feature T of the text description of the candidate's practical test is: T = LLM (Projection layer (V));
[0026] Establish the connection between the video clips and text descriptions of the key experimental steps, and generate the features Hs = {(V1, T1), (V2, T2), …, (V k ,T k )}, where V k Represents the characteristics of the video clip, T k represents the features of the text description, k represents the number of video clips, and the text description can be regarded as the label of the video clip; the practical behavior recognition module integrates the text description of each video clip in chronological order to form a description of the student's practical examination process.
[0027] The comparison of the features of the video and the video, the video and the text, and the text and the text by the scoring module is specifically: V is represented as the features of the video clip of the candidate's practical test, T is represented as the features of the text description of the candidate's practical test, V' is represented as the features of the video clip of the standard practical test, and T' is represented as the features of the text description of the standard practical test;
[0028] Compare the features of the video clips of the candidate's practical test with the features of the video clips of the standard practical test, denoted as (V, V′);
[0029] The comparison method of (V, V′) is to obtain the similarity score by calculating the cosine similarity between the global feature vectors of the two video clips.
[0030] Compare the features of the video clip of the candidate's practical test with the features of the text description of the standard practical test, denoted as (V,T');
[0031] Compare the features of the text description of the candidate's practical test with the features of the video clip of the standard practical test, denoted as (T, V′);
[0032] (V,T′) is compared to (T,V′) by mapping text features to the same dimensional space as visual features through a trainable multi-layer perceptron (MLP). For text, it uses BERT or a similar pre-trained language model to obtain an embedded representation of the text. The features of the video and text are fed into their own Transformer models. These Transformer models can be independent or share some or all parameters. The Transformer model processes the input sequence through a self-attention mechanism to generate high-level representations of the video and text. The hidden state of each video and text token output by the Transformer is aggregated by average pooling to obtain a global feature representation. The similarity score is obtained by calculating the cosine similarity between the global feature vectors of the video and text, which quantifies the degree of match between the video content and the text description. A similarity below 50% is considered irrelevant.
[0033] The features of the text description of the candidate's practical test are compared with the features of the text description of the standard practical test, denoted as (T, T′); the comparison method of (T, T′) is to use BERT or a similar pre-trained language model to obtain the text features, and the similarity score is obtained by calculating the cosine similarity between the two text features.
[0034] The practical behavior module connects the representations of these visual features in the chronological order of these experimental steps to maintain the coherence of the practical process. At the same time, it also pays attention to the global semantic information representing the entire video, combines this global information with the features of each segment, and ensures that when understanding each specific experimental step, the significance of each experimental step in the entire experimental process can also be considered;
[0035] In step 3, the scoring module compares the features of the videos, videos and texts, and texts and texts, removes the features of the video clips that are not related to the test according to the marks, records the reasons for the examinee's loss of points, and scores the examinee's practical test, specifically:
[0036] After the scoring module obtains the candidate's practical test video and text description saved in step 1, it extracts the video, text description and fragment information of the standard practical test related to this test, slices the standard practical test video and text description according to the fragment information, and generates the feature Hn = {(V′1, T′1), (V′2, T′2), …, (V′ k ,T′ k )}; Then, evenly sample video frames from the video clips of the standard practical test feature Hn, extract the visual features of each frame, and extract the text features from the corresponding text description to obtain hn = (V′ y ,T′ y );
[0037] The scoring module obtains the video clips and the marked text descriptions hs=(V x ,T x ), remove the features of the video clips that are not related to the test according to the labels; then evenly sample the features of the video clips of hs, and extract the text features corresponding to the visual features in the corresponding text description;
[0038] V x ,V′ y )、(V x ,T′ y )、(V′ y ,T x ) and (T x ,T′ y) are compared respectively, and the similarity is calculated, and the results of the four are integrated to generate (V x ,T x ) and (V′ y ,T′ y ) represents the difference between
[0039] The scoring unit gives the corresponding scores according to the differences and the scoring table. The scores are calculated after the test to get the final scores of this test.
[0040] Score=100-Grading_units(D).
[0041] The confidence evaluation module uses the trust score as the calculation method; the calculation formula is:
[0042]
[0043] Among them, x is the input, s is the predicted score, m is the scoring model, and h is the feature of the standard answer; d c (x,h) represents the shortest Euclidean distance between the standard answer feature with the same score and the output feature h′ obtained after input into m; d p (x,h) represents the shortest Euclidean distance between the standard answer features under different scores and the output feature h′ obtained after input into m.
[0044] Compared with the prior art, the present invention has the following advantages:
[0045] The system of the present invention uses a module composed of a visual encoder layer, a projection layer, and an LLM layer to perform practical behavior recognition, which is more suitable for the diversity and complexity of experimental operations compared with the recognition method of the prior art. The scoring module uses a multi-angle comparison method of visual features and text features for scoring, which is more scientific. The confidence evaluation evaluates the confidence of the score, which improves the credibility and persuasiveness of the score.
[0046] This method divides automatic scoring into three processes: practical behavior recognition, scoring, and confidence assessment. This method is applicable to a variety of experimental scenarios. The scoring method is more objective, using the comparison of the video and text features of the student's practical test with the video and text features of the standard practical test. Teachers can judge whether human intervention is needed for this scoring based on the results given by the confidence assessment module. This method increases scoring efficiency and also protects students' rights and interests to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0048] Figure 1 A schematic diagram of the process structure of a method for automatic scoring of practical examinations proposed by the present invention;
[0049] Figure 2 A schematic diagram of the system structure for an automatic scoring method for practical examinations proposed for this discovery;
[0050] Figure 3 Schematic diagram of the scoring module structure of the automatic scoring method for practical examinations proposed in this discovery. DETAILED DESCRIPTION
[0051] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.
[0052] It should be understood that when used in this specification and the appended claims, the term "pre-training" refers to training a model on a large amount of data so that it can learn common language patterns, knowledge or features; "salient features" refers to the experimental operation behavior of students in a certain frame of the video; "compact representation" refers to the use of key frames within a video clip to represent the features of the entire video clip; "standard practical test video" refers to the experimental demonstration video of teachers or professionals, which is the standard process of the experiment. "Standard practical test text description" refers to the description of the standard process of the experiment.
[0053] Example 1
[0054] like Figure 1 As shown, Figure 1 A schematic diagram of the process structure of a method for automatic scoring of practical examinations proposed by the present invention;
[0055] The automatic scoring method for physics, chemistry and biology practical examination based on visual language model includes the following steps:
[0056] Step 1: Use a camera to capture the process of students' practical examination from multiple perspectives and save the video data;
[0057] Step 2: Identify the student's practical test behavior through the practical behavior recognition module; send the identified video clips and marked text descriptions to the scoring module;
[0058] Step 3: The scoring module compares the features of the videos, videos and texts, and texts and removes the features of the video clips that are not relevant to the test based on the marks, records the reasons for the examinees' loss of points, and scores the examinees' practical test.
[0059] Step 4, using the confidence evaluation module, the rationality of the practical test score is calculated based on the representation characteristics of the candidate's practical test, the representation characteristics of the standard practical test, and the test score;
[0060] Step 5: The teacher will conduct practical score correction for candidates with low score confidence. For scores questioned by candidates, the teacher will re-correct them through manual verification.
[0061] In the step 2, the practical test behavior of the students is identified through the practical behavior identification module. When cheating behavior of the students is detected, the video clip containing the cheating behavior is marked, the cheating behavior is recorded in the text description, and a request to terminate the test is issued, and the test score is recorded as 0 points. When the students are detected to have behavior unrelated to the test, the video clip containing the unrelated behavior is marked, and the behavior unrelated to the test is not recorded in the text description.
[0062] The step 2 is to identify the student's practical test behavior through the practical behavior recognition module; send the identified video clips and marked text descriptions to the scoring module; specifically:
[0063] The practical behavior recognition module divides the video saved in step 1 into several segments, each of which represents a key experimental step, aggregates the significant features in each segment, and generates a compact representation of each step;
[0064] Generate corresponding representations according to the time sequence of the experimental steps to maintain the coherence of the practical process, so that the model can understand and recognize the actual process of the practical test; at the same time, pay attention to the global semantic information representing the entire video, combine the global information with the characteristics of each clip, and ensure that when understanding each specific experimental step, the significance of each experimental step in the entire experimental process is taken into account;
[0065] The calculation formula of the feature V of the video clip of the candidate's practical test is: V = Visual encoder (AV); the calculation formula of the feature T of the text description of the candidate's practical test is: T = LLM (Projection layer (V));
[0066] Establish the connection between the video clips and text descriptions of the key experimental steps, and generate the features Hs = {(V1, T1), (V2, T2), …, (V k ,T k )}, where V kRepresents the characteristics of the video clip, T k represents the features of the text description, k represents the number of video clips, and the text description can be regarded as the label of the video clip; the practical behavior recognition module integrates the text description of each video clip in chronological order to form a description of the student's practical examination process.
[0067] The comparison of the features of videos and videos, videos and texts, and texts and texts by the scoring module is specifically as follows:
[0068] Compare the features of the video clips of the candidate's practical test with the features of the video clips of the standard practical test, denoted as (V, V′);
[0069] The comparison method of (V, V′) is to obtain the similarity score by calculating the cosine similarity between the global feature vectors of the two video clips.
[0070] Compare the features of the video clip of the candidate's practical test with the features of the text description of the standard practical test, denoted as (V,T′);
[0071] Compare the features of the text description of the candidate's practical test with the features of the video clip of the standard practical test, denoted as (T, V′);
[0072] (V,T′) is compared to (T,V′) by mapping text features to the same dimensional space as visual features through a trainable multi-layer perceptron (MLP). For text, it uses BERT or a similar pre-trained language model to obtain an embedded representation of the text. The features of the video and text are fed into their own Transformer models. These Transformer models can be independent or share some or all parameters. The Transformer model processes the input sequence through a self-attention mechanism to generate high-level representations of the video and text. The hidden state of each video and text token output by the Transformer is aggregated by average pooling to obtain a global feature representation. The similarity score is obtained by calculating the cosine similarity between the global feature vectors of the video and text, which quantifies the degree of match between the video content and the text description. A similarity below 50% is considered irrelevant.
[0073] The features of the text description of the candidate's practical test are compared with the features of the text description of the standard practical test, denoted as (T, T′); the comparison method of (T, T′) is to use BERT or a similar pre-trained language model to obtain the text features, and the similarity score is obtained by calculating the cosine similarity between the two text features.
[0074] In step 3, the scoring module compares the features of the videos, videos and texts, and texts and texts, removes the features of the video clips that are not related to the test according to the marks, records the reasons for the examinee's loss of points, and scores the examinee's practical test, specifically:
[0075] After the scoring module obtains the candidate's practical test video and text description saved in step 1, it extracts the video, text description and fragment information of the standard practical test related to this test, slices the standard practical test video and text description according to the fragment information, and generates the feature Hn = {(V′1, T′1), (V′2, T′2), …, (V′ k ,T′ k )}; Then, evenly sample video frames from the video clips of the standard practical test feature Hn, extract the visual features of each frame, and extract the text features from the corresponding text description to obtain hn = (V′ y ,T′ y );
[0076] The scoring module obtains the video clips and the marked text descriptions hs=(V x ,T x ), remove the features of the video clips that are not related to the test according to the labels; then evenly sample the features of the video clips of hs, and extract the text features corresponding to the visual features in the corresponding text description;
[0077] V x ,V′ y )、(V x ,T′ y )、(V′ y ,T x ) and (T x ,T′ y ) are compared respectively, and the similarity is calculated, and the results of the four are integrated to generate (V x ,T x ) and (V′ y ,T′ y ) represents the difference between
[0078] The scoring unit gives the corresponding scores according to the differences and the scoring table. The scores are calculated after the test to get the final scores of this test.
[0079] Score=100-Grading_units(D).
[0080] The confidence evaluation module uses the trust score as the calculation method; the calculation formula is:
[0081]
[0082] Among them, x is the input, s is the predicted score, m is the scoring model, and h is the feature of the standard answer; d c (x,h) represents the shortest Euclidean distance between the standard answer feature with the same score and the output feature h′ obtained after input into m; d p (x,h) represents the shortest Euclidean distance between the standard answer features under different scores and the output feature h′ obtained after input into m.
[0083] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.
Claims
1. An automatic scoring system for practical examinations in physics, chemistry and biology based on a visual language model, characterized by: Including practical behavior recognition module, scoring module, and confidence assessment module; The practical behavior recognition module is used to analyze and recognize the examinee's behavior; The scoring module is used to compare the candidate's experimental operation with the standard experimental operation and score the candidate's practical test; The confidence evaluation module is used to evaluate the rationality of the practical test score.
2. The automatic scoring system for practical examinations in physics, chemistry and biology based on a visual language model according to claim 1 is characterized by: The practical operation behavior recognition module is implemented by a visual language model, and the visual language model includes: a visual encoder layer, a projection layer layer and an LLM layer; The Visual encoder layer is used to convert visual information into a representation that can be understood and processed by the model, extract features from the original visual data, learn the potential representation of visual information, and enhance the visual perception ability of the model; The Projectionlayer layer is used to map the features extracted by the visual encoder into a representation space that matches the language model; The LLM layer is used to learn the representation of visual information provided by the visual encoder layer, obtain the generation capability on the task, and generate corresponding text output according to the input image.
3. The automatic scoring system for practical examinations in physics, chemistry and biology based on visual language model according to claim 1 is characterized by: The behaviors identified by the practical behavior identification module include: examinee's practical test behaviors, cheating behaviors, and behaviors unrelated to the test.
4. The automatic scoring system for practical examinations in physics, chemistry and biology based on a visual language model according to claim 1 is characterized by: The scoring module includes an experiment comparison unit and a scoring unit; the experiment comparison unit is used to compare the video and text description of the candidate's experimental operation with the video and text description of the standard experimental operation, and use the difference between the candidate's experimental operation and the standard experimental operation as the basis for scoring; the scoring unit integrates the candidate's gain and loss points based on the differences between the visual language features, and then generates corresponding scores based on the scoring rules.
5. The automatic scoring method for practical examinations in physics, chemistry and biology based on a visual language model according to any one of claims 1 to 4, characterized in that: The following steps are involved: Step 1: film the process of students' practical examination from multiple perspectives and save the video data; Step 2: Identify the student's practical test behavior through the practical behavior recognition module; send the identified video clips and marked text descriptions to the scoring module; Step 3: The scoring module compares the features of the videos, videos and texts, and texts and removes the features of the video clips that are not relevant to the test based on the marks, records the reasons for the examinees' loss of points, and scores the examinees' practical test. Step 4, using the confidence evaluation module, the rationality of the practical test score is calculated based on the representation characteristics of the candidate's practical test, the representation characteristics of the standard practical test, and the test score; Step 5: The teacher will conduct practical score correction for candidates with low score confidence. For scores questioned by candidates, the teacher will re-correct them through manual verification.
6. The automatic scoring method for practical examinations in physics, chemistry and biology based on a visual language model according to claim 5 is characterized by: In the step 2, the practical test behavior of the students is identified through the practical behavior identification module, specifically: when the student is detected to have cheating behavior, the video clip containing the cheating behavior is marked, the cheating behavior is recorded in the text description, and the test score is recorded as 0 points; when the student is detected to have behavior unrelated to the test, the video clip containing the unrelated behavior is marked, and the behavior unrelated to the test is not recorded in the text description.
7. The automatic scoring method for practical examinations in physics, chemistry and biology based on a visual language model according to claim 5 is characterized by: The step 2 is to identify the student's practical test behavior through the practical behavior recognition module; send the identified video clips and marked text descriptions to the scoring module; specifically: The practical behavior recognition module divides the video saved in step 1 into several segments, each of which represents a key experimental step, aggregates the significant features in each segment, and generates a compact representation of each step; Generate corresponding representations according to the time sequence of the experimental steps to maintain the coherence of the practical process, so that the model can understand and recognize the actual process of the practical test; at the same time, pay attention to the global semantic information representing the entire video, combine the global information with the characteristics of each clip, and ensure that when understanding each specific experimental step, the significance of each experimental step in the entire experimental process is taken into account; The calculation formula of the feature V of the video clip of the candidate's practical test is: V = Visual encoder (AV); the calculation formula of the feature T of the text description of the candidate's practical test is: T = LLM (Projection layer (V)); Establish the connection between the video clips and text descriptions of the key experimental steps, and generate the features Hs = {(V1, T1), (V2, T2), …, (V k ,T k )}, where V k Represents the characteristics of the video clip, T k represents the features of the text description, k represents the number of video clips, and the text description can be regarded as the label of the video clip; the practical behavior recognition module integrates the text description of each video clip in chronological order to form a description of the student's practical examination process.
8. The automatic scoring method for practical examinations in physics, chemistry and biology based on a visual language model according to claim 5 is characterized by: In step 3, the features of the videos, videos and texts, and texts are compared by the scoring module as follows: Compare the features V of the video clip of the candidate's practical test with the features V′ of the video clip of the standard practical test, denoted as (V, V′); Compare the features V of the video clip of the candidate's practical test with the features T' of the text description of the standard practical test, denoted as (V, T'); Compare the features T of the text description of the candidate's practical test with the features V' of the video clip of the standard practical test, denoted as (T, V'); The feature T of the text description of the candidate's practical test is compared with the feature T' of the text description of the standard practical test, denoted as (T, T').
9. The automatic scoring method for practical examinations in physics, chemistry and biology based on a visual language model according to claim 5 is characterized by: In step 3, the scoring module compares the features of the videos, videos and texts, and texts and texts, removes the features of the video clips that are not related to the test according to the marks, records the reasons for the examinee's loss of points, and scores the examinee's practical test, specifically: After the scoring module obtains the candidate's practical test video and text description saved in step 1, it extracts the video, text description and fragment information of the standard practical test related to this test, slices the standard practical test video and text description according to the fragment information, and generates the feature Hn = {(V′1, T′1), (V′2, T′2), …, (V′ k ,T′ k )}; Then, evenly sample video frames from the video clips of the standard practical test feature Hn, extract the visual features of each frame, and extract the text features from the corresponding text description to obtain hn = (V′ y ,T′ y ); The scoring module obtains the video clips and the marked text descriptions hs=(V x ,T x ), remove the features of the video clips that are not related to the test according to the labels; then evenly sample the features of the video clips of hs, and extract the text features corresponding to the visual features in the corresponding text description; V x ,V′ y )、(V x ,T′ y )、(V′ y ,T x ) and (T x ,T′ y ) are compared respectively, and the similarity is calculated, and the results of the four are integrated to generate (V x ,T x ) and (V′ y ,T′ y ) represents the difference between The scoring unit gives the corresponding scores according to the differences and the scoring table. The scores are calculated after the test to get the final scores of this test. Score=100-Grading_units(D).
10. The automatic scoring method for practical examinations in physics, chemistry and biology based on a visual language model according to claim 5 is characterized by: The confidence evaluation module uses a trust score as the calculation method; The calculation formula is: Among them, x is the input, s is the predicted score, m is the scoring model, and h is the feature of the standard answer; d c (x,h) represents the shortest Euclidean distance between the standard answer feature with the same score and the output feature h′ obtained after input into m; d p (x,h) represents the shortest Euclidean distance between the standard answer features under different scores and the output feature h′ obtained after input into m.