Large-model-assisted student ability level adaptive test method
By fine -tuning the large language model, building a student ability reasoning model, solving the problem of initialization of students' ability initialization in computer adaptive testing, and improving the accuracy of testing and user experience.
Patent Information
- Application Number
- CN202411282631.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2025-01-24
AI Technical Summary
The existing computer adaptive testing methods have shortcomings in the initialization of the ability of cold-start students, resulting in a large number of students being assigned unsuitable questions, increasing the number of evaluation rounds and poor user experience.
Make fine -tuning with a large language model (LLM) to build a student ability reasoning model, and train the interactive matrix through the student's auxiliary information and exercise answers to generate a fine -tuning large language model LLM 'for adaptive testing students' ability.
Effectively solve the problem of initialization of cold starting students, improve the accuracy of testing and user experience, and significantly improve the accuracy of the cognitive diagnostic module and the effectiveness of the topic selector.
Smart Images
Figure CN119359066A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of smart education, large language models and data mining, and in particular to a method for solving the cold start problem in computer adaptive student ability testing based on fine-tuning a large language model. Background Art
[0002] Computerized Adaptive Testing (CAT) is an intelligent testing technology that aims to accurately assess students' abilities by adaptively selecting the most appropriate questions. In the CAT system, there are two important modules that are executed in steps, the cognitive diagnosis module and the question selector module. In each step, the cognitive diagnosis module estimates the current student's ability level based on the student's previous answers; while the question selector module selects the appropriate question as the student's next question based on the obtained student's ability and question characteristics.
[0003] Existing CAT methods mainly focus on the design of two modules, but ignore the exploration of how to initialize the abilities of cold-start students. The abilities of students in these methods are initialized by random vectors. Random initialization leads to a large number of cold-start students being assigned questions that are not suitable for them, which not only increases the number of rounds of assessment, but is also unfriendly to the user experience. Summary of the invention
[0004] The present invention aims to address the deficiencies of the above-mentioned prior art and proposes a large model-assisted adaptive testing method for student ability levels, in order to effectively estimate the ability parameters of cold-start students, thereby improving test accuracy and user experience.
[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:
[0006] The invention provides a method for adaptively testing student ability levels assisted by a large model, which is carried out according to the following steps:
[0007] Step 1: Build a large language model training text set ,in, represents the i-th training text, |S| represents the total number of training students;
[0008] Step 2: Based on D, fine-tune the large language model LLM to obtain the fine-tuned large language model LLM';
[0009] Step 3: Use LLM' as the student ability reasoning model to adaptively test student ability and obtain the final estimated student ability parameters.
[0010] The large model-assisted adaptive testing method for student ability levels described in the present invention is also characterized in that step 1 is performed as follows:
[0011] Step 1.1: Construct the student-question answer record interaction matrix:
[0012] Definition Problem Set , where e j represents the jth exercise, |E| represents the total number of exercises, 1≤j≤|E|;
[0013] Define the training student set , where s i represents the i-th training student, |S| represents the total number of training students, 1≤i≤|S|;
[0014] Defining the test student set ,in, represents the kth test student, |U| represents the total number of test students, 1≤k≤|U|;
[0015] Let the interactive matrix of the training students' answer records for the exercises be ,in, represents the i-th training student s i Answer exercise j The score of i Answer the jth exercise correctly , then let , if s i Wrong answer to question j , then let , if s i Unanswered Exercise j , then let ;
[0016] Step 1.2, order represents a set of auxiliary information, where Indicates Information, Indicates the total amount of auxiliary information, 1≤ ≤|T|;
[0017] Step 1.3: Let the i-th training student The records on all auxiliary information are ,in, represents the i-th training student In the Auxiliary information If the i-th student In the Auxiliary information If there is no record, ;
[0018] Step 1.4: Let the kth test student The records on all auxiliary information are ,in, represents the kth test student In the Auxiliary information If the records on In the Auxiliary information If there is no record, ;
[0019] Step 1.5: represents the i-th student s i Capability representation, represents the jth exercise The parameter representation of the exercise is and Input to cognitive diagnostic model Processed in, get the i-th training student s i Answer exercise j The predicted score ; Then use formula (1) to calculate the cross entropy loss loss:
[0020] (1)
[0021] Step 1.6: Cognitive diagnostic model by minimizing cross entropy loss Conduct training to obtain a trained cognitive diagnostic model and its optimal model parameters, and the optimal ability parameters of all training student sets in the optimal model parameters are recorded as And the optimal parameters for all exercise sets are recorded as ;
[0022] Step 1.7: Use formula (2) to convert Capacity parameters Mapped to a score between 0-100 :
[0023] (2)
[0024] In formula (2), yes The maximum value in yes The minimum value in ;
[0025] Step 1.8: Generate the i-th training text using formula (3) , thus obtaining the training text set :
[0026] (3)
[0027] In formula (3), LLM represents the large language model; prompt represents the set prompt.
[0028] Further, the step 2 is performed as follows:
[0029] Step 2.1: Tokenize D to obtain the processed training text set ;
[0030] Step 2.2: Define the pre-trained weights of the large language model LLM as , and use Gaussian to initialize the first low-rank matrix , use zero to initialize the second low-rank matrix , thereby initializing the fine-tuning incremental weights of the large language model LLM ;
[0031] Step 2.3: Freeze the pre-trained weights Finally, use the Adam optimizer to train the large language model LLM and save the fine-tuning incremental weights corresponding to the lowest loss value during training. ;
[0032] Step 2.4: and After adding, assign to After that, we get the fine-tuned large language model LLM'.
[0033] Further, the step 3 is performed as follows:
[0034] Step 3.1: Test students Auxiliary information Input LLM' for processing, and then output using formula (4) Estimated ability score :
[0035] (4)
[0036] Step 3.2: Use formula (5) to convert Mapping back to estimated capacity parameters , and as the tth iteration Estimated capability parameters :
[0037] (5)
[0038] Step 3.3: Use formula (6) to get the tth iteration The most suitable topic :
[0039] (6)
[0040] In formula (6), Indicated in Down The amount of Fisher information contained, and:
[0041] Guidance Report (7)
[0042] In formula (7), represent exist In the case of Predicted correct answer The probability of express The partial derivative of ;
[0043] Step 3.4: middle The corresponding optimal parameters Input to Then, the equation (8) is used to obtain the value of Estimated capability parameters :
[0044] (8)
[0045] Step 3.5, after assigning t+1 to t, return to step 3.3 and execute sequentially until the maximum number of iterations is reached, thus obtaining Final estimated student ability parameters.
[0046] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the student ability level adaptive testing method, and the processor is configured to execute the program stored in the memory.
[0047] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the student ability level adaptive testing method when the computer program is executed by a processor.
[0048] Compared with the prior art, the present invention has the following beneficial effects:
[0049] 1. This paper uses a lightweight fine-tuning method Lora to adjust the base large model to solve the cold start problem in computer adaptive testing (CAT). This method uses the powerful ability of large language models (LLMs) to effectively estimate the ability level of students without any initial data on the students. This innovation fills a gap in existing research and is the first time that LLMs have been applied to solve the cold start problem in CAT.
[0050] 2. The present invention considers the impact of various auxiliary information of students on their learning and integrates it into the task of estimating student ability parameters. This approach not only helps to more accurately assess students' ability levels, but also provides data support for personalized education, improves educational equity, and significantly improves the accuracy of the cognitive diagnosis module and the effectiveness of the question selector by combining contextual features with fine-tuning of the LLM. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 is a flow chart of computer adaptive testing of the present invention;
[0052] Figure 2 This is a flow chart of the computer adaptive testing method assisted by a large model according to the present invention. DETAILED DESCRIPTION
[0053] In this embodiment, a large model-assisted adaptive test method for student ability level is based on the significant difference between the training task of the large language model and the task of estimating student ability parameters. By adopting the lightweight fine-tuning method Lora and constructing a student ability parameter estimation task instruction data set to adjust the base large model, at the same time, by merging the large model after Lora weighting, it is possible to effectively estimate the ability parameters of cold-start students. Specifically, Figure 2 As shown, the method comprises the following steps:
[0054] Step 1: Build a training text set for a large language model:
[0055] Step 1.1: Construct the student-question answer record interaction matrix:
[0056] Definition Problem Set , where e j represents the jth exercise, |E| represents the total number of exercises, 1≤j≤|E|;
[0057] Define the training student set , where s i represents the i-th training student, |S| represents the total number of training students, 1≤i≤|S|;
[0058] Defining the test student set ,in, represents the kth test student, |U| represents the total number of test students, 1≤k≤|U|;
[0059] Let the interactive matrix of the training students' answer records for the exercises be ,in, represents the i-th training student s i Answer exercise j The score of i Answer the jth exercise correctly , then let , if s i Wrong answer to question j , then let , if s i Unanswered Exercise j , then let .
[0060] Step 1.2, order represents a set of auxiliary information, where Indicates Information, Indicates the total amount of auxiliary information, 1≤ ≤|T|, auxiliary information such as family economics, parents’ education level, IQ test, etc.;
[0061] Step 1.3: Let the i-th training student The records on all auxiliary information are ,in, represents the i-th training student In the Auxiliary information If the i-th student In the Auxiliary information If there is no record, ;
[0062] Step 1.4: Let the kth test student The records on all auxiliary information are ,in, represents the kth test student In the Auxiliary information If the records on In the Auxiliary information If there is no record, .
[0063] Step 1.5: represents the i-th student s i Capability representation, represents the jth exercise The parameter representation of the exercise is and Input to cognitive diagnostic model Processed in, get the i-th training student s i Answer exercise j The predicted score ; Then use formula (1) to calculate the cross entropy loss loss:
[0064] (1)
[0065] Step 1.6: Cognitive diagnostic model by minimizing cross entropy loss Conduct training to obtain a trained cognitive diagnostic model and its optimal model parameters, and the optimal ability parameters of all training student sets in the optimal model parameters are recorded as , and the optimal parameters for all exercise sets are recorded as .
[0066] Step 1.7: Use formula (2) to convert Capacity parameters Mapped to a score between 0-100 :
[0067] (2)
[0068] In formula (2), yes The maximum value in yes The minimum value in ;
[0069] Step 1.8: Generate the i-th training text using formula (3) , thus obtaining the training text set :
[0070] (3)
[0071] In formula (3), LLM represents the large language model; prompt represents the set prompt; the complete For: {"instruction": "Based on the student's auxiliary information, consider the impact of this information on learning ability and predict the student's ability value, which is between 0 and 100", "input": "{ }No explanation or steps are required. The answer format is: The student's ability value is approximately <ability value>. ","output": "The student's ability value is approximately <ability value> >"}.
[0072] Step 2: Fine-tune the large language model LLM:
[0073] Step 2.1: Tokenize D to obtain the processed training text set ;
[0074] Step 2.2: Define the pre-trained weights of the large language model LLM as , and use Gaussian to initialize the first low-rank matrix , use zero to initialize the second low-rank matrix , thereby initializing the fine-tuning incremental weights of the large language model LLM ;
[0075] Step 2.3: Freeze the pre-trained weights Finally, use the Adam optimizer to train the large language model LLM and save the fine-tuning incremental weights corresponding to the lowest loss value during training. ;
[0076] Step 2.4: and After adding, assign to After that, we get the fine-tuned large language model LLM'.
[0077] Step 3: Use LLM' as a student ability reasoning model to adaptively test student abilities:
[0078] Step 3.1: Test students Auxiliary information Input LLM' for processing, and then output using formula (4) Estimated ability score :
[0079] (4)
[0080] Step 3.2 Use formula (5) to convert Mapping back to estimated capacity parameters , and as the tth iteration Estimated capability parameters :
[0081] (5)
[0082] Step 3.3: Use formula (6) to get the tth iteration The most suitable topic :
[0083] (6)
[0084] In formula (6), Indicated in Down The amount of Fisher information contained is specifically:
[0085] (7)
[0086] In formula (7), represent exist In the case of Predicted correct answer The probability of Express The partial derivative process of ;
[0087] Step 3.4: middle The corresponding optimal parameters Input to Then, the equation (8) is used to obtain the value of Estimated capability parameters :
[0088] (8)
[0089] Step 3.5, after assigning t+1 to t, return to step 3.3 and execute sequentially until the maximum number of iterations is reached, thus obtaining Final estimated student ability parameters.
[0090] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.
[0091] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.
[0092] In this embodiment, auxiliary information and academic performance data of 500 students were extracted from the public data set SLP. The method of the present invention first preprocessed these data and extracted auxiliary information of each student. Then, the real ability parameters of the trained students were obtained through the cognitive diagnosis model. Next, the training data set was constructed using the prompt word template using these auxiliary information and real ability parameter values, and the ChatGLM-6B model was fine-tuned using Lora technology on GPU 4090. After the training was completed, the fine-tuned model was used to evaluate the ability of the cold-start students in the test set, and the results were compared with the traditional random initialization method. The results show that the method of the present invention is significantly superior to the traditional method (random initialization) in the accuracy of the initial test question selection.
[0093] Computer adaptive testing Figure 1 As shown, at the beginning of the test, the student's ability parameters are randomly initialized. According to the student's current ability estimate, the question selection algorithm is used to select the most suitable question for the student. After the student finishes answering the question, the cognitive diagnosis model is used to update the student's ability parameters according to the answer situation. This step is repeated many times until the test step length is met or the test accuracy reaches a certain requirement, and the test is terminated.
[0094] In order to verify the effectiveness of the method of the present invention, the method of the present invention is experimentally verified on the SLP dataset and the math dataset. The SLP dataset has been made public.
[0095] First, the data set is cleaned, mainly from the following aspects:
[0096] 1) Filter out students with incomplete background information;
[0097] 2) Filter questions with less than 10 answers;
[0098] 3) Filter students who answered less than 10 questions based on 1;
[0099] Secondly, compared with the Base method, that is, randomly initializing the student ability values;
[0100] Evaluation indicators: ACC and AUC;
[0101] Table 1 Experimental results
[0102]
[0103] The experimental results of this method are shown in Table 1, where the underlying cognitive diagnosis model is IRT and the question selection algorithm is the maximum Fisher information. Because the generation of large models is unstable, this method was used for 5 tests, namely test1, test2, test3, test4, and test5. It can be seen from Table 1 that the AUC and ACC indicators of the technical solution of this application are better than the baseline Base solution at the beginning of the test, and have obvious advantages, which is consistent with the essential goal of computer adaptive testing, which is to accurately measure the true ability value of students through a small number of questions.
Claims
1. A large model-assisted adaptive testing method for student ability levels, characterized in that: The steps are as follows: Step 1: Build a large language model training text set ,in, represents the i-th training text, |S| represents the total number of training students; Step 2: Based on D, fine-tune the large language model LLM to obtain the fine-tuned large language model LLM'; Step 3: Use LLM' as the student ability reasoning model to adaptively test student ability and obtain the final estimated student ability parameters.
2. A large model-assisted student ability level adaptive testing method according to claim 1, characterized in that: The step 1 is carried out as follows: Step 1.1: Construct the student-question answer record interaction matrix: Definition Problem Set , where e j represents the jth exercise, |E| represents the total number of exercises, 1≤j≤|E|; Define the training student set , where s i represents the i-th training student, |S| represents the total number of training students, 1≤i≤|S|; Defining the test student set ,in, represents the kth test student, |U| represents the total number of test students, 1≤k≤|U|; Let the interactive matrix of the training students' answer records for the exercises be ,in, represents the i-th training student s i Answer exercise j The score of i Answer the jth exercise correctly , then let , if s i Wrong answer to question j , then let , if s i Unanswered Exercise j , then let ; Step 1.2, order represents a set of auxiliary information, where Indicates Information, Indicates the total amount of auxiliary information, 1≤ ≤|T|; Step 1.3: Let the i-th training student The records on all auxiliary information are ,in, represents the i-th training student In the Auxiliary information If the i-th student In the Auxiliary information If there is no record, ; Step 1.4: Let the kth test student The records on all auxiliary information are ,in, represents the kth test student In the Auxiliary information If the records on In the Auxiliary information If there is no record, ; Step 1.5: represents the i-th student s i Capability representation, represents the jth exercise The parameter representation of the exercise is and Input to cognitive diagnostic model Processed in, get the i-th training student s i Answer exercise j The predicted score ; Then use formula (1) to calculate the cross entropy loss loss: (1) Step 1.6: Cognitive diagnostic model by minimizing cross entropy loss Conduct training to obtain a trained cognitive diagnostic model and its optimal model parameters, and the optimal ability parameters of all training student sets in the optimal model parameters are recorded as And the optimal parameters for all exercise sets are recorded as ; Step 1.7: Use formula (2) to convert Capacity parameters Mapped to a score between 0-100 : (2) In formula (2), yes The maximum value in yes The minimum value in ; Step 1.8: Generate the i-th training text using formula (3) , thus obtaining the training text set : (3) In formula (3), LLM represents the large language model; prompt represents the set prompt.
3. A large model-assisted adaptive testing method for student ability levels according to claim 2, characterized in that: Described step 2 is carried out as follows: Step 2.1: Tokenize D to obtain the processed training text set ; Step 2.2: Define the pre-trained weights of the large language model LLM as , and use Gaussian to initialize the first low-rank matrix , use zero to initialize the second low-rank matrix , thereby initializing the fine-tuning incremental weights of the large language model LLM ; Step 2.3: Freeze the pre-trained weights Finally, use the Adam optimizer to train the large language model LLM and save the fine-tuning incremental weights corresponding to the lowest loss value during training. ; Step 2.4: and After adding, assign to After that, we get the fine-tuned large language model LLM'.
4. A large model-assisted student ability level adaptive testing method according to claim 3, characterized in that: The step 3 is carried out as follows: Step 3.1: Test students Auxiliary information Input LLM' for processing, and then output using formula (4) Estimated ability score : (4) Step 3.2: Use formula (5) to convert Mapping back to estimated capacity parameters , and as the tth iteration Estimated capability parameters : (5) Step 3.3: Use formula (6) to get the tth iteration The most suitable topic : (6) In formula (6), Indicated in Down The amount of Fisher information contained, and: (7) In formula (7), represent exist In the case of Predicted correct answer The probability of express The partial derivative of ; Step 3.4: middle The corresponding optimal parameters Input to Then, the equation (8) is used to obtain the value of Estimated capability parameters : (8) Step 3.5, after assigning t+1 to t, return to step 3.3 and execute sequentially until the maximum number of iterations is reached, thus obtaining Final estimated student ability parameters.
5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the student ability level adaptive testing method described in any one of claims 1-4, and the processor is configured to execute the program stored in the memory.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the student ability level adaptive testing method described in any one of claims 1 to 4 are executed.