Experiment course-oriented intelligent tutoring device and method

Abstract

The invention discloses an intelligent tutoring device and method for experiment courses. The device comprises an acquisition unit, a monitoring unit, a tracking unit and a question and answer unit. The acquisition unit is used for acquiring, cleaning and classifying process data of student operation experiments; the monitoring unit is used for monitoring behaviors and progress of students in the experiment process in real time; the tracking unit analyzes experimental operation historical data and real-time data of students through an experimental level tracking model, and can predict correctness of subsequent experimental operation; the question and answer unit carries out semantic analysis and feature extraction on texts asked by the students and calculates similarity and correlation between the texts, so that questions encountered by the students in the experiment operation process are more accurately solved. The device provides auxiliary support for experimental process tutoring, experimental ability evaluation and other links of students, targeted auxiliary questions and answers can be performed according to different experimental ability levels of the students, and more intelligent and efficient whole-process experimental auxiliary guidance is realized.

Description

Technical Field

[0001] The present invention relates to an intelligent tutoring device and method, and in particular to an intelligent tutoring device and method for experimental courses. Background Art

[0002] With the continuous development of science and technology, experimental courses play a vital role in education. Experimental courses can help students consolidate theoretical knowledge, cultivate their practical ability and innovative thinking. However, since experimental courses have high requirements for hands-on ability, students may encounter various problems during the experiment and need timely and effective guidance.

[0003] In the past, the teaching system was basically based on textbooks. The teaching content was mainly the basic content such as the theory of the course. The teachers also explained the theoretical knowledge step by step, and rarely involved practical applications. Through this learning mode, it is difficult for students to form a deep understanding of the important technologies in the experimental courses, and it is even more impossible to consolidate these knowledge in practice. It does not match the specific learning requirements, which not only increases the learning burden, but also leads to frustration of enthusiasm. However, due to the shortage of teachers in colleges and universities, it is impossible to offer more experimental courses, and teachers are unable to participate in the experimental process of each student throughout the process. Therefore, the experimental content arranged by most universities is almost mainly for the digestion of the knowledge learned in the course, and rarely designs comprehensive applications. At the same time, although many teachers have introduced the online teaching mode, they still simply apply video conferencing and online courses to the delivery of teaching content, and have not really made full use of the interactive and diversified functions of the network platform to carry out teaching activities. These problems have led to the limited improvement of the quality of experimental teaching. In order to solve this problem, the intelligent tutoring device in the present invention can monitor students' behavioral performance in the experiment, predict students' current experimental operation status, provide students with real-time answers to questions, and help students quickly solve difficulties in the experiment. This not only saves teachers' time and energy, but also improves students' experimental operation capabilities. Summary of the invention

[0004] Purpose of the invention: The first purpose of the present invention is to provide an intelligent tutoring device and method for experimental courses that can monitor and track students' experimental operations throughout the entire process and provide timely tutoring and answers to students' questions. The second purpose of the present invention is to provide an intelligent question-answering unit based on a semantic vector model.

[0005] Technical solution: The present invention is an intelligent tutoring device for experimental courses, comprising a collection unit, a monitoring unit, a tracking unit and a question-answering unit;

[0006] The collection unit is used to collect data from the experimental operation page. It can identify and collect data related to students' experimental behavior, experimental progress, etc., and perform preliminary data cleaning and classification.

[0007] The monitoring unit is used to monitor students' behavior during the experiment in real time. The tracking unit can be called according to needs to predict students' experimental operations. If it is predicted that students are about to have problems, the question-and-answer unit will be called to assist in error correction.

[0008] The tracking unit is used to track and analyze students' previous experimental behavior data during the students' experiment using the experimental level tracking model, and to predict the accuracy of students' next experimental operation based on their current experimental behavior data.

[0009] The question-answering unit performs semantic analysis and feature extraction on the text based on the topic classification model and semantic vector model, extracts key topic features from the experimental question text, matches the extracted topic features with the existing experimental topic vocabulary, screens out questions that meet the experimental topic category, and then uses similarity matching to select candidate answers.

[0010] To sum up, the collection unit provides data reserves for the subsequent monitoring unit and tracking unit. The tracking unit discovers patterns and trends from the students' experimental behavior data and predicts the accuracy of the students' subsequent experimental operations. The question-and-answer unit is based on the topic classification model and semantic vector model, which can answer students' questions in a timely and accurate manner, while reducing the burden on experimental instructors.

[0011] The intelligent tutoring method for experimental courses specifically includes the following steps:

[0012] (1) Students complete the specified experimental operations in the experimental platform according to the requirements of the experimental task. This device will collect and record information such as students' experimental behavior, experimental steps, and experimental progress, pre-process the data, and provide a data source for the analysis and calculation of subsequent tracking units.

[0013] (2) The monitoring unit will monitor the students' experimental progress in real time, and combine the experimental behavior data and experimental progress information collected by the acquisition unit to predict the accuracy of the students' next step in the experimental process, and at the same time call the question-and-answer unit to provide auxiliary guidance.

[0014] (3) The tracking unit conducts deeper mining and processing on the data collected by the acquisition unit and analyzes the students’ previous data through the experimental level tracking model. The input of the experimental level tracking model is a student’s historical experimental record. It combines the current student’s historical experimental record to predict the student’s performance in the next experimental step.

[0015] The experiment level tracking model consists of three parts: Rasch embedding layer, learning layer and prediction layer. First, the experiment sequence (e1, e2, ..., e n ), the sequence of knowledge points covered by the experiment (c1, c2, ..., c n) and the student's experimental result sequence (r1, r2, ..., r n ) is input into the Rasch layer, and the experiment embedding and interaction embedding carrying historical information are obtained through the long short-term memory network and monotonic attention. Then, the embedded experiments and interactions are input into the learning layer, and the monotonic attention is used to simulate the students' learning behavior, thereby outputting the students' hidden knowledge state. Finally, the prediction layer passes the students' hidden knowledge state through the Stu-C decoder and the Stu-E decoder respectively, and outputs the students' e n Experimental predictions.

[0016] The input of the Rasch layer is the experimental sequence (e1, e2, ..., e n ), the sequence of knowledge points covered by the experiment (c1, c2, ..., c n ) and the student's experimental result sequence (r1, r2, ..., r n ), experimental embeddings with historical information obtained through long short-term memory networks and monotonic attention and interactive embedding The calculation process is shown in formulas (1) and (2):

[0017]

[0018] Among them, α n-i is the nith experimental pair C n The impact of is the value vector of the mapping matrix.

[0019]

[0020] Among them, α n-i is the nith interaction pair R n The impact of is the value vector of the mapping matrix.

[0021] The learning layer tracks the changes in students’ experimental levels based on historical experiments and completion status, and simulates the behavior of students’ experimental operations. The monotonic attention mechanism is used to simulate students’ operation behavior. The input is and The output is the student's hidden knowledge vector h n The formula principle is different from the embedding layer in that the learning layer uses Multiply a query matrix and a key matrix to get the query mapping and key mapping use Multiply a value matrix to get a value mapping vector

[0022] The role of the prediction layer is to predict the students' future experiments. nIn the prediction layer, two decoders are defined, Stu-C decoder and Stu-E decoder, which are used to extract the students’ mastery of each knowledge point and the probability of students making mistakes when they master the knowledge. n The inputs of the two decoders are Among them, H n The embedding vector h is the hidden knowledge state and future experiments of the student. n Hide the state of knowledge for students, For future experimental embedding, the Stu-C decoder transforms H through a linear layer and a sigmoid function. n Decoded as the degree of mastery of the knowledge points covered in the experiment L i The Stu-E decoder transforms H into n Decoded into prediction parameter G i The probability that a student will make a mistake in the experimental steps if they have mastered the knowledge. The probability that a student will make a correct experimental step p i ∈(0, 1) as shown in formula (3):

[0023] p i =sigmoid{(1-L i )×G i} (3)

[0024] Among them, L i is the mastery of knowledge points, G i It represents the probability that students make mistakes in experimental steps when they have mastered the knowledge.

[0025] (1) The question-answering unit includes the following steps: constructing a knowledge base through the data generated by the topic classification model and the semantic vector model, extracting key topic features from the question text, matching the extracted topic features with the existing topic vocabulary, screening out questions that meet the specific topic category, and then using similarity matching to select candidate answers. The steps for building the topic classification model are as follows: First, pre-train the model and initialize the topic features in combination with the labeled known topic classification task. At this stage, L is used as the loss function of the entire classification, L = α1×L1+α2×L2, where α1 and α2 are constants. The cross entropy loss function L1 calculates the loss of the predicted category label as shown in formula (4), and the mean square error loss function L2 calculates the loss of the continuous value corresponding to the predicted category label as shown in formula (5):

[0026]

[0027] Among them, φ is a linear classifier, φ(z i ) j Yes i The probability of belonging to the jth class, y i Theme zi The corresponding label;

[0028]

[0029] Among them, n is the number of samples, y_true is each actual experimental subject class, y_pred is the prediction result of the model, and the smaller the mean square error, the closer the prediction result of the model is to the true value.

[0030] Then, the pre-trained model is used to extract topic features to learn the decision boundary. It is assumed that each type of known topic sample is located in a spherical area. For each type of known topic cluster, the cluster center and radius are defined to determine the decision boundary, and then the decision boundary is learned in combination with the known topic feature distribution. After learning the cluster center and decision boundary radius of each type of known topic, the topics are classified according to the distance between the current topic and all cluster centers. The steps of constructing the semantic vector model are as follows: train the semantic vector model of the bidirectional encoder, calculate the loss of the function, determine the weight value of each sample according to the loss, and adjust the loss function so that the model in the following training pays more attention to difficult negative examples. The new loss function is shown in formula (6):

[0031]

[0032] Among them, L adjusted is the adjusted loss function, L i is the original loss of the ith sample, ω i is the weight of the sample. Through the topic classification model, an experimental course is divided into knowledge bases of different chapters, which speeds up the subsequent search for problems and improves the accuracy. By training the bidirectional encoder model, the training of difficult samples is emphasized to increase the accuracy of the model.

[0033] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: by establishing a systematic experimental course intelligent tutoring device, teachers can organize and manage experimental courses more effectively, without having to participate in the experimental tutoring of each student throughout the whole process, saving teachers' time and energy, and alleviating the plight of the shortage of teachers for experimental courses in some universities. The device can monitor students' behavioral performance in the experiment, predict students' current experimental operation status through the experimental level tracking model, provide students with real-time feedback on problems, help students quickly solve difficulties in the experiment, improve students' experimental operation ability, and further promote the intelligent development of practical teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the intelligent tutoring device of the present invention;

[0035] Figure 2 It is a workflow diagram of the question-answering unit of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is further described below in conjunction with embodiments.

[0037] The core of the present invention is an intelligent tutoring device and method designed for experimental courses, which can promptly answer questions encountered by students during the experiment and provide effective support for learning supervision, after-school tutoring and other links, thereby improving the efficiency of students' experimental operation and reducing the burden on teachers. Figure 1 As shown, the intelligent tutoring device for experimental courses of the present invention includes a collection unit, a monitoring unit, a tracking unit and a question-answering unit.

[0038] The acquisition unit is responsible for collecting data from the experimental operation page, identifying and collecting data related to students' experimental behavior, experimental progress, etc., and performing preliminary data cleaning and classification.

[0039] The monitoring unit monitors the students' behavior during the experiment in real time, and can call the tracking unit to predict the students' experimental operations according to needs. If it is predicted that the students are about to have problems, the question-and-answer unit will be called to assist in error correction.

[0040] During the students' experiment, the tracking unit uses the experimental level tracking model to track and analyze the students' previous experimental behavior data, and predicts the accuracy of the students' next experimental operation based on their current experimental behavior data.

[0041] The question-answering unit can perform semantic analysis and feature extraction on text based on the topic classification model and semantic vector model, extract key topic features from experimental question text, match the extracted topic features with the existing experimental topic vocabulary, screen out questions that meet the experimental topic category, and then use similarity matching to select candidate answers.

[0042] Figure 2 It is a specific implementation flow chart of the question-answering unit in the present invention, including the following steps:

[0043] Step 1: The collection unit uses tracking technology to collect students' behavior data on the experimental operation page, collect students' experimental operation process and record students' experimental process. In order to collect the tracking data generated by the experimental operation page, tracking technology provides services in the form of HTTP interface. The reporting request type of tracking data is HTTP POST, the type of request body is multipart / form-data, and the tracking data is stored in the request body in binary form.

[0044] After the request for reporting the tracking data arrives, the request will be forwarded to the downstream data receiving service. The key is built into the application in advance. After receiving the request, the service verifies the access key and starts parsing the request. When parsing, it is necessary to obtain the sender's IP address from the request header, obtain the compressed binary stream of the tracking data from the request body, and then assemble it together with the timestamp corresponding to the current time zone into a new data object, and finally store it in a blocking queue, waiting for data to be sent.

[0045] The receiving service first stores the processed data in a blocking queue and sends it to the message queue cluster through multi-threading. If the sending fails, it means that network congestion or message queue cluster crash occurs at this time. Then the message is placed in another blocking queue and stored locally in a multi-threaded manner. Data cleaning takes the byte array message in the message queue as input and outputs the processed fine-grained data to the data warehouse.

[0046] Step 2: The monitoring unit uses the experimental level tracking model to make real-time predictions on the accuracy of the student's next experimental step, sets automated reminder rules, and calls the question-and-answer unit to answer when it is predicted that the student will make a mistake in the next step. The question-and-answer unit is called in the device to provide auxiliary guidance to prompt the student. First, the student's next experimental step is predicted through the experimental level tracking model. If the prediction is wrong, the question-and-answer unit is called to answer the wrong question; then, the solution given by the question-and-answer unit is accepted through the prompt information in the prompt rule configuration; finally, the prompt message is pushed in the device: rules are defined in the prompt configuration file, and the prompt message is sent to the front end. The push function is integrated in the device to realize the message prompt.

[0047] Step 3: The tracking unit analyzes students’ past data through the experimental level tracking model, obtains experimental embedding and interaction embedding that carry historical information through long short-term memory network and monotonic attention, and uses monotonic attention to simulate students’ experimental operation behaviors, thereby outputting students’ hidden knowledge states. After passing through the Stu-C decoder and Stu-E decoder respectively, the prediction of students’ current experimental situation is output.

[0048] The input of the experimental level tracking model is a student's historical experimental record, defined as X = {x1, x2, ..., x n-1}, where x i represents the experimental situation of student i at time i, which is called an interaction and is represented by a triple, namely x i =(c i , e i , r i ), where c i represents the principles covered by the experimental steps performed by student i at time i, e i represents the experimental steps performed by student i at time i, ri Indicates the correctness of the experimental steps, r i =1 means the step is correct, r i = 0 means the step is wrong. The task of experimental level tracking is to predict the student’s step e in the next experiment based on the student’s historical experimental record X. n performance.

[0049] The experimental level tracking model consists of three parts: Rasch embedding layer, learning layer and prediction layer. The goal of the Rasch embedding layer is to mine more information through the experimental sequence, while solving the problem of data sparsity, so as to retain the relevant complexity of knowledge points and experiments as much as possible. In this process, the historical information of experimental operations is crucial. This historical information contains students' understanding of knowledge and experimental habits, and has a significant effect on predicting students' future experimental results. To this end, the Rasch embedding layer embeds the relevant complexity of experiments and knowledge points and the influence of historical information to maximize the embedding effect. The input of the Rasch layer is the experimental sequence (e1, e2, ..., e n ), the sequence of knowledge points covered by the experiment (c1, c2, ..., c n ) and the student's experimental result sequence (r1, r2, ..., r n ).

[0050] The initial embedded object is the knowledge point c n , and the interaction tuple (c n , r n ), the embedding process of the two is similar, but the model cannot process tuples directly, so the tuple is represented as a real number a n , as shown in formula (7). Then c n , a n and e n . Multiply each by an embedding matrix A1(|C|×d k ), A2(2|E|×1) and A3(2|C|×d k ) is mapped into a continuous vector space, and the dimension is d k The knowledge point vector of dimension k n , and the interaction vector Q n , and the experimental complexity parameter μ n ∈R. In order to incorporate the experimental complexity into k n and Q n In the process, long short-term memory network is used to extract knowledge points c n and interaction n The information in the vector d n and f nFinally, the experimental complexity parameter is used to calculate the correlation complexity between the experiment and the knowledge point and integrate it into the knowledge point and interaction vector to obtain the preliminary embedding C of the experiment containing the interaction between the experiment and the knowledge point. n and interactive preliminary embedding in R n , as shown in formulas (8) and (9):

[0051] a n =c n +|C|×r n (7)

[0052] C n =k n +μ n ·d n (8)

[0053] R n =Q n +μ n ·f n (9)

[0054] When predicting future experimental situations, it is necessary to capture students’ experimental habits from their past experimental sequences, that is, the impact of past experiments and their completion on the future. Therefore, an attention mechanism containing a kernel function is used to embed experiments and interactions twice, with C n and R n For example, in the secondary embedding process:

[0055] First, (C1, C2, ..., C n ) are respectively related to the three mapping matrices W1 Q 、W1 K and W1 V The query vector q C , key vector k C Sum value vector v C Multiply.

[0056] Then calculate the first n-1 experimental pairs C n The impact of the interaction sequence (R1, R2, ..., R n ) The same process is performed to obtain the past interaction pair R n As shown in formulas (10) and (11):

[0057]

[0058] in, is the nith experimental pair C n The impact of is the nith interaction pair R n The influence of k i-1 is the mapping matrix W1K The key vector of .

[0059] Finally, respectively and v C and v R Experimental embedding after weighted summation to obtain quadratic embedding and interactive embedding The calculation process is shown in formulas (12) and (13):

[0060]

[0061] in, is the nith experimental pair C n The impact of is the nith interaction pair R n The impact of and C n and R n The corresponding mapping matrix W1 V A vector of values ​​of .

[0062] The learning layer tracks the changes in students’ experimental levels based on historical experiments and completion status, simulating the behavior of students acquiring knowledge. The monotonic attention mechanism is used to simulate students’ experimental operation behavior. The input is and The output is the student's hidden knowledge vector h n , as shown in formulas (14) and (15), the principle of the formula is different from that of the embedding layer in that the learning layer uses Multiply a query matrix and a key matrix to get the query mapping and key mapping use Multiply a value matrix to get a value mapping vector

[0063]

[0064] in, and They are respectively the experimental embeddings after using secondary embedding in the learning layer The query mapping and key mapping are obtained by multiplying a query matrix and a key matrix respectively. For interactive embedding after secondary embedding A value mapping vector obtained by multiplying a value matrix.

[0065] The role of the prediction layer is to predict the students' future experiments. nIn the prediction layer, two decoders are defined, Stu-C decoder and Stu-E decoder, which are used to extract the students’ mastery of each knowledge point and the probability of students making mistakes when they master the knowledge. n The inputs of the two decoders are the embedding vectors of the student’s hidden knowledge state and future experiments, expressed as H n To embed the student’s hidden knowledge state and future experiments, the Stu-C decoder transforms H n Decoded as the degree of mastery of the knowledge points covered in the experiment L i The Stu-E decoder transforms H into n Decoded into prediction parameter G i The probability that a student will make a mistake in the experimental steps if they have mastered the knowledge. The probability that a student will make a correct experimental step p i ∈(0, 1) as shown in formula (16):

[0066] p i =sigmoid{(1-L i )×G i} (16)

[0067] Among them, L i G is the students’ mastery of the knowledge points covered in the experiment. i The probability that students make mistakes in experimental steps when they have mastered the knowledge.

[0068] Step 4: The question-answering unit includes the following steps: constructing a knowledge base through the data generated by the topic classification model and the semantic vector model.

[0069] First, extract the topic features. Given the i-th input sentence s i After all data passes through the Transformer encoding layer, the word embedding representation of the last layer output is extracted, and the embedding of all its tags [C, T1, ..., T N ]∈R (N+1)×Y , Y is the hidden layer size, and then the average vector representation x is obtained through the average pooling operation i ∈R Y As shown in formula (17):

[0070] x i =mean-pooling([C,T1,...,T N ]) (17)

[0071] Among them, C is the vector of text classification, N is the sequence length, T iIt is a sentence i To further enhance the feature extraction capability, we input a dense layer h to obtain the topic feature z. i ∈R D As shown in formula (18):

[0072] z i =h(x i )=σ(W h x i +b h ) (18)

[0073] Where D is the dimension of the topic representation, σ is the ReLU activation function, and W h ∈R Y×D and b h ∈R D They represent the weight and bias term of the dense layer h respectively.

[0074] Since the decision boundary depends on adaptive learning of the known topic feature distribution, it is necessary to pre-train the model and initialize the topic features in combination with the labeled known topic classification task. At this stage, L is used as the loss function of the entire classification, L = α1×L1+α2×L2, where the cross entropy loss function L1 calculates the loss of the predicted category label as shown in formula (19), and the mean square error loss function L2 calculates the loss of the continuous value corresponding to the predicted category label as shown in formula (20).

[0075]

[0076] Among them, φ is a linear classifier, φ(z i ) j Yes i The probability of belonging to the jth class, y i Theme z i The corresponding label;

[0077]

[0078] Among them, n is the number of samples, y_true is each actual experimental subject class, y_pred is the prediction result of the model, and the smaller the mean square error, the closer the prediction result of the model is to the true value.

[0079] Then, the pre-trained model is used to extract topic features to learn the decision boundary. Assuming that each type of known topic sample is located in the spherical region, the cluster center is determined by calculating the average feature vector of each cluster as shown in formula (21):

[0080]

[0081] Where S = {(z i ,yi ),...,(z N ,y N )} are known topic examples and their corresponding labels. k represents the set of examples labeled as class k. The cluster center c k ∈R D YesS k The average vector of the embedded samples in .

[0082] The radius parameter is initialized through the neural network. In order to ensure that the radius is differentiable, the SoftPlus activation function is used as the mapping between the radius and the radius parameter. The radius is defined as shown in formula (22):

[0083]

[0084] Among them, Δ k is the decision boundary relative to the cluster center c k The radius of .

[0085] Next, we learn the decision boundary based on the known topic feature distribution. After learning the cluster center and decision boundary radius of each known topic, we classify the topics according to the distance between the current topic and all cluster centers, and identify the topic corresponding to the closest cluster center by default.

[0086] Then, the semantic vector model of the bidirectional encoder is trained, the function loss is calculated, the weight value of each sample is determined according to the loss, and the loss function is adjusted so that the model in the following training pays more attention to difficult negative examples. The new loss function calculation is shown in formula (23):

[0087]

[0088] Among them, L adjusted is the adjusted loss function, L i is the original loss of the ith sample, ω i is the weight of the sample.

[0089] Through the topic classification model, an experimental course is divided into knowledge bases with different chapters, which speeds up the subsequent problem finding and improves the accuracy. The bidirectional encoder model is further trained, focusing on the training of difficult samples to increase the accuracy of the model.

[0090] The semantic vector data format constructed by the above two models is: Data 1: {Title: topic (topic classification model), Content: knowledge point, Content_vec: knowledge point vector (semantic vector model)}.

[0091] Finally, extract key topic features from the question text, identify the topic category to which the question belongs, match the extracted topic features with the existing topic word library, and determine the topic category to which the question belongs. Based on the topic matching results, filter out questions that meet the specific topic category, ignore questions that are not related to the target topic, and then use cosine similarity to compare question vectors, calculate the similarity between questions, and select the one with the highest similarity as the candidate answer.

Claims

1. An intelligent tutoring device for experimental courses, characterized by: It includes collection unit, monitoring unit, tracking unit and question-answering unit; The acquisition unit is used to collect data from the experimental operation page, identify and collect data related to students' experimental behavior, experimental progress, etc., and perform preliminary data cleaning and classification; The monitoring unit is used to monitor the student's behavior in real time during the experiment. The tracking unit can be called according to the needs to predict the student's experimental operation. If it is predicted that the student is about to have a problem, the question-and-answer unit is called to assist in error correction. The tracking unit is used to track and analyze the students' previous experimental behavior data using the experimental level tracking model during the students' experimental process, and predict the accuracy of the students' next experimental operation based on the students' current experimental behavior data; The question-answering unit performs semantic analysis and feature extraction on the text based on the topic classification model and the semantic vector model, extracts key topic features from the experimental question text, matches the extracted topic features with the existing experimental topic vocabulary, screens out questions that meet the experimental topic category, and then uses similarity matching to select candidate answers.

2. An intelligent tutoring method for experimental courses, characterized in that: The steps include: (1) The collection unit is responsible for collecting and identifying data related to students' experimental behavior, experimental progress, etc. from the experimental operation page, and performing preliminary data cleaning and classification; (2) The monitoring unit monitors the students' experimental progress in real time, and predicts each step of the students' experimental operation during the experiment based on the experimental behavior data collected and preprocessed by the acquisition unit. If it is predicted that the students are about to have problems, the question-and-answer unit is called to assist in error correction; (3) The tracking unit conducts a deeper level of analysis and mining on the experimental behavior and experimental progress data collected by the acquisition unit and the monitoring unit, analyzes the students’ previous experimental operation data through the experimental level tracking model, and predicts the accuracy of the students’ next experimental operation based on the current experimental process of the students; (4) The question-answering unit decomposes an experimental course into experimental knowledge bases of different topics through a topic classification model. The semantic vector model outputs the data with experimental topic classification as semantic vectors, constructs an experimental knowledge base, extracts key topic features from the question text, matches the extracted topic features with the existing experimental topic vocabulary, screens out questions that meet specific topic categories, and uses similarity matching to select candidate answers.

3. The intelligent tutoring method for experimental courses according to claim 2, characterized in that: The experimental level tracking model in step (3) consists of three parts: Rasch embedding layer, learning layer and prediction layer. The input of the Rasch layer is the experimental sequence (e1, e2, ..., e n ), the sequence of knowledge points covered by the experiment (c1, c2, ..., c n ) and the student's experimental result sequence (r1, r2, ..., r n ), experimental embeddings with historical information obtained through long short-term memory networks and monotonic attention and interactive embedding The calculation process is shown in formula (1) and formula (2): Among them, α n-i is the nith experimental pair C n The impact of is the value vector of the mapping matrix; where α n-i is the nith interaction pair R n The impact of is the value vector of the mapping matrix; The learning layer tracks the changes in students' experimental levels based on historical experiments and completion status, simulates students' behavior in performing experimental operations, and uses the monotonic attention mechanism to simulate students' learning behavior. The input is and The output is the student's hidden knowledge vector h n , which is different from the embedding layer, the learning layer uses Multiply a query matrix and a key matrix to get the query mapping and key mapping use Multiply a value matrix to get a value mapping vector The role of the prediction layer is to predict the students' future experiments. n In the prediction layer, two decoders are defined, Stu-C decoder and Stu-E decoder, which are used to extract the students’ mastery of each experimental knowledge point and the probability of students making mistakes in the experiment when they have mastered the experimental knowledge. n The probability of; the input of the two decoders is (where H n The embedding vector h is the hidden knowledge state and future experiments of the student. n Hide the state of knowledge from students, Embedded for future experiments), the Stu-C decoder transforms H through a linear layer and a sigmoid function. n Decoded as the degree of mastery of the knowledge points covered in the experiment L i , the Stu-E decoder transforms H through a linear layer and a sigmoid function n Decoded into prediction parameter G i It represents the probability that students make mistakes in experimental steps when they have mastered experimental knowledge, and the probability that students make correct experimental steps p i ∈(0, 1) as shown in formula (3): p i =sigmoid{(1-L i )×G i } (3) Among them, L i G is the mastery of experimental knowledge points, i It represents the probability that students make mistakes in experimental steps when they have mastered the experimental knowledge.

4. The intelligent tutoring method for experimental courses according to claim 2, characterized in that: The steps for constructing the topic classification model in step (4) are as follows: First, the model is pre-trained and the topic features are initialized in combination with the label known topic classification task. At this stage, L is used as the loss function of the entire classification, L = α1×L1+α2×L2, α1 and α2 are constants, where the cross entropy loss function L1 calculates the loss of the predicted category label as shown in formula (4), and the mean square error loss function L2 calculates the loss of the continuous value corresponding to the predicted category label as shown in formula (5): Among them, φ is a linear classifier, φ(z i ) j Yes i The probability of belonging to the jth class, y i Theme z i The corresponding label; Among them, n is the number of samples, y_true is each actual experimental subject class, y_pred is the prediction result of the model, and the smaller the mean square error, the closer the prediction result of the model is to the true value.

5. The intelligent tutoring method for experimental courses according to claim 2, characterized in that: The steps of constructing the semantic vector model in step (4) are as follows: train the semantic vector model of the bidirectional encoder, calculate the function loss, determine the weight value of each sample according to the loss, and adjust the loss function so that the model pays more attention to difficult negative examples in subsequent training. The new loss function is shown in formula (6): Among them, L adjusted is the adjusted loss function, L i is the original loss of the ith sample, ω i is the weight of the sample.

Images