CN-121999655-A - Test question explanation method and device, electronic equipment and product

CN121999655ACN 121999655 ACN121999655 ACN 121999655ACN-121999655-A

Abstract

The application provides a test question explanation method, a device, electronic equipment and a product, wherein the method is used for generating a shell-clamping point guide list corresponding to a test question to be explained based on test question information of the test question to be explained, and the shell-clamping point guide list comprises explanation information and guide questions corresponding to at least one shell-clamping point of the explanation question; inputting the test question information of the test questions to be explained and the card case point guide list corresponding to the test questions to be explained into a pre-constructed teaching and guiding system, and carrying out test question explanation interaction of the test questions to be explained with a user by utilizing the teaching and guiding system according to the card case point guide list. By adopting the technical scheme, the card case point guiding list of the test questions to be explained can be generated in advance, and the teaching and tutoring system can explain and ask questions according to the card case point guiding list aiming at the problem solving card case points in the test questions to be explained when the test questions to be explained are taught to the user, so that heuristic guidance of students is realized, active thinking of the students is promoted, and further, the teaching effect of the test questions explanation is improved.

Inventors

LI XIXI
LI SONG
SHA JING
SHENG ZHICHAO
WANG SHIJIN

Assignees

科大讯飞股份有限公司

Dates

Publication Date: 20260508
Application Date: 20251230

Claims (13)

1. The test question explanation method is characterized by comprising the following steps of: Based on the test question information of the test questions to be explained, generating a card shell point guide list corresponding to the test questions to be explained, wherein the card shell point guide list comprises explanation information and guide questions corresponding to at least one explanation card shell point; inputting the test question information of the test questions to be explained and the card shell point guide list corresponding to the test questions to be explained into a pre-constructed teaching and guiding system, and carrying out test question explanation interaction of the test questions to be explained with a user by utilizing the teaching and guiding system according to the card shell point guide list.
2. The method for teaching questions as claimed in claim 1, wherein generating the list of guidance of the stuck point corresponding to the questions to be taught based on the questions information of the questions to be taught comprises: Inputting test question information of a test question to be explained into a pre-trained cartooning point guide generation model to obtain a cartooning point guide list corresponding to the test question to be explained; the shell point guide generation model is obtained by training a shell point guide list generation of the large language model through a pre-collected training sample.
3. The method of claim 1, wherein the cartooning point guides a training process of generating a model, comprising: based on a first training sample acquired in advance, performing supervision fine tuning on the first large language model to obtain a supervision fine tuning model; and performing reinforcement learning training on the supervised fine tuning model based on a second training sample acquired in advance to obtain a reinforcement learning model, and taking the reinforcement learning model as a card shell point guide generation model.
4. The method for teaching a test question according to claim 3, wherein the first training sample comprises a stuck point guide generation instruction corresponding to a first sample test question and a desired stuck point guide list corresponding to the first sample test question, and the stuck point guide generation instruction corresponding to the first sample test question consists of a preset stuck point guide generation template and test question information of the first sample test question; Based on a first training sample acquired in advance, performing supervision fine tuning on the first large language model to obtain a supervision fine tuning model, wherein the supervision fine tuning model comprises the following components: inputting a first pre-collected training sample into a first large language model, so that the first large language model generates a predicted card case point guide list corresponding to the first sample test question according to a card case point guide generation instruction corresponding to the first sample test question; And carrying out parameter adjustment on the first large language model by taking the same target of the predicted card shell point guide list corresponding to the first sample test question and the expected card shell point guide list corresponding to the first sample test question as the target, so as to obtain a supervision fine-tuning model.
5. The method for teaching questions according to claim 3, wherein the second training samples comprise a stuck point guide generation instruction corresponding to a second sample question, a preferred stuck point guide list and an inferior stuck point guide list corresponding to the second sample question; based on a second training sample acquired in advance, performing reinforcement learning training on the supervised fine tuning model to obtain a reinforcement learning model, wherein the reinforcement learning model comprises: Inputting a second training sample acquired in advance into the supervision fine adjustment model so that the supervision fine adjustment model generates a prediction card shell point guide list corresponding to the second sample test question according to a card shell point guide generation instruction corresponding to the second sample test question; and carrying out parameter adjustment on the supervision fine adjustment model by taking the difference between the predicted card case point guide list corresponding to the second sample test question and the preferred card case point guide list corresponding to the second sample test question as a target, wherein the difference between the predicted card case point guide list corresponding to the second sample test question and the preferred card case point guide list corresponding to the second sample test question is the smallest, and the difference between the predicted card case point guide list corresponding to the second sample test question and the inferior card case point guide list corresponding to the second sample test question is the largest.
6. The method for teaching questions according to claim 3, wherein the second training sample comprises a stuck point guide generation instruction corresponding to a third sample question, and the stuck point guide generation instruction corresponding to the third sample question consists of a preset stuck point guide generation template and the question information of the third sample question; based on a second training sample acquired in advance, performing reinforcement learning training on the supervised fine tuning model to obtain a reinforcement learning model, wherein the reinforcement learning model comprises: Inputting a second training sample acquired in advance into a historical iteration model, so that the historical iteration model generates at least one predicted card shell point guide list corresponding to the third sample test question according to a card shell point guide generation instruction corresponding to the third sample test question; Determining a reward value of a predicted card shell point guide list corresponding to the third sample test question by utilizing a pre-trained reward model; based on the predicted card shell point guide list corresponding to the third sample test question and the rewarding value of the predicted card shell point guide list corresponding to the third sample test question, carrying out parameter adjustment on the iteration model of the current iteration process to obtain an iteration model of the next iteration process; and if the current iteration process reaches the preset iteration condition, taking an iteration model of the next iteration process as a reinforcement learning model.
7. The method of claim 6, wherein performing parameter adjustment on the iteration model of the current iteration process based on the predicted card case point guide list corresponding to the third sample test question and the reward value of the predicted card case point guide list corresponding to the third sample test question to obtain the iteration model of the next iteration process comprises: Determining a first generation probability of generating a prediction card case point guide list corresponding to the third sample test question by using the historical iteration model, determining a second generation probability of generating the prediction card case point guide list corresponding to the third sample test question by using an iteration model of the current iteration process, and taking the ratio of the second generation probability to the first generation probability as a strategy ratio between the iteration model and the historical iteration model of the current iteration process; calculating the average value and standard deviation of the rewarding values of the prediction card shell point guide list corresponding to all the third sample test questions; Determining a relative dominance value of the predicted card case point guide list corresponding to the third sample test question based on the reward value of the predicted card case point guide list corresponding to the third sample test question, the mean value and the standard deviation; and carrying out parameter adjustment on the iteration model of the current iteration process based on the strategy ratio between the iteration model and the historical iteration model of the current iteration process and the relative advantage value of the prediction karst point guide list to obtain the iteration model of the next iteration process.
8. The method of claim 6, wherein the training process of the reward model comprises: Collecting a third training sample, wherein the third training sample comprises a stuck point guide generation instruction corresponding to a fourth sample test question, a stuck point guide list of the fourth sample test question and a quality score label corresponding to the stuck point guide list, and the stuck point guide generation instruction corresponding to the fourth sample test question consists of a preset stuck point guide generation template and test question information of the fourth sample test question; inputting the third training sample into a pre-constructed reward model to obtain a predicted mass fraction corresponding to a shell-clamping point guide list of the fourth sample test question; And carrying out parameter adjustment on the reward model by taking the mass fraction label corresponding to the shell point guiding list as a target, wherein the mass fraction label is the same as the predicted mass fraction corresponding to the shell point guiding list.
9. The method of claim 3, wherein the reinforcement learning training is performed on the supervised fine tuning model based on a second training sample acquired in advance, and further comprising, after obtaining the reinforcement learning model: and carrying out model compression on the reinforcement learning model to obtain a compression model, and guiding the compression model as a card shell point to generate a model.
10. The method of claim 9, wherein performing model compression on the reinforcement learning model to obtain a compressed model comprises: Performing supervision fine tuning on a second large language model to obtain a student model, wherein the parameter scale of the second large language model is smaller than that of the first large language model; Collecting a fourth training sample, wherein the fourth training sample comprises a shell clamping point guide generation instruction corresponding to a fifth sample test question and an expected shell clamping point guide list corresponding to the fifth sample test question, and the shell clamping point guide generation instruction corresponding to the fifth sample test question consists of a preset shell clamping point guide generation template and test question information of the fifth sample test question; inputting a fourth training sample acquired in advance into the student model to obtain a predicted card shell point guide list corresponding to the fifth sample test question, and obtaining a first predicted probability distribution of the predicted card shell point guide list and a first expected probability distribution of the expected card shell point guide list, which are determined by the student model; Inputting a predicted card case point guide list corresponding to the fifth sample test question into the reinforcement learning model, and acquiring a second predicted probability distribution of the predicted card case point guide list and a second expected probability distribution of the expected card case point guide list, which are determined by the reinforcement learning model; and carrying out parameter adjustment on the student model by taking the first prediction probability distribution and the second prediction probability distribution as targets and the first expected probability distribution and the second expected probability distribution as targets.
11. The utility model provides a test question explanation device which characterized in that includes: The system comprises a guide generation module, a guide generation module and a display module, wherein the guide generation module is used for generating a card shell point guide list corresponding to a test question to be explained based on the test question information of the test question to be explained, and the card shell point guide list comprises explanation information and guide questions corresponding to at least one explanation card shell point; And the interaction module is used for inputting the test question information of the test questions to be taught and the card shell point guide list corresponding to the test questions to be taught into a pre-constructed teaching and guiding system, and carrying out test question teaching interaction of the test questions to be taught with a user by utilizing the teaching and guiding system according to the card shell point guide list.
12. An electronic device is characterized by comprising a memory and a processor; the memory is connected with the processor and used for storing programs; The processor is configured to implement the method for explaining a test question according to any one of claims 1 to 10 by running a program in the memory.
13. A computer program product comprising computer program instructions which, when executed by a processor, cause the processor to implement the method of question interpretation of any one of claims 1 to 10.

Description

Test question explanation method and device, electronic equipment and product Technical Field The application relates to the technical field of intelligent education, in particular to a test question explanation method, a device, electronic equipment and a product. Background Along with the rapid development of artificial intelligence technology, in the field of teaching, education and intelligence changes are gradually advanced, and a teaching and coaching model is gradually and widely focused. The existing teaching and tutoring model generally gives corresponding test question answers and test question analysis directly aiming at test question questions, but gives the test question answers and analysis directly, so that students are easy to steal and do not actively think, and the teaching effect is affected. Disclosure of Invention Based on the requirements, the application provides a test question explanation method, a device, electronic equipment and a product, which can promote the active thinking of students and improve the teaching effect of test question explanation. In order to achieve the above purpose, the present application proposes the following technical scheme: According to a first aspect of an embodiment of the present application, there is provided a method for explaining a test question, including: Based on the test question information of the test questions to be explained, generating a card shell point guide list corresponding to the test questions to be explained, wherein the card shell point guide list comprises explanation information and guide questions corresponding to at least one explanation card shell point; inputting the test question information of the test questions to be explained and the card shell point guide list corresponding to the test questions to be explained into a pre-constructed teaching and guiding system, and carrying out test question explanation interaction of the test questions to be explained with a user by utilizing the teaching and guiding system according to the card shell point guide list. Optionally, based on the test question information of the test questions to be explained, generating the card shell point guide list corresponding to the test questions to be explained includes: Inputting test question information of a test question to be explained into a pre-trained cartooning point guide generation model to obtain a cartooning point guide list corresponding to the test question to be explained; the shell point guide generation model is obtained by training a shell point guide list generation of the large language model through a pre-collected training sample. Optionally, the cartooning point guides a training process of generating a model, including: based on a first training sample acquired in advance, performing supervision fine tuning on the first large language model to obtain a supervision fine tuning model; and performing reinforcement learning training on the supervised fine tuning model based on a second training sample acquired in advance to obtain a reinforcement learning model, and taking the reinforcement learning model as a card shell point guide generation model. Optionally, the first training sample comprises a card case point guide generation instruction corresponding to a first sample test question and an expected card case point guide list corresponding to the first sample test question, wherein the card case point guide generation instruction corresponding to the first sample test question consists of a preset card case point guide generation template and test question information of the first sample test question; Based on a first training sample acquired in advance, performing supervision fine tuning on the first large language model to obtain a supervision fine tuning model, wherein the supervision fine tuning model comprises the following components: inputting a first pre-collected training sample into a first large language model, so that the first large language model generates a predicted card case point guide list corresponding to the first sample test question according to a card case point guide generation instruction corresponding to the first sample test question; And carrying out parameter adjustment on the first large language model by taking the same target of the predicted card shell point guide list corresponding to the first sample test question and the expected card shell point guide list corresponding to the first sample test question as the target, so as to obtain a supervision fine-tuning model. Optionally, the second training sample comprises a card shell point guide generation instruction corresponding to a second sample test question, a preferred card shell point guide list and an inferior card shell point guide list corresponding to the second sample test question; based on a second training sample acquired in advance, performing reinforcement learning training on the supervised fine tuning model to obtai