Artificial intelligence learning-based user knowledge tracing system and operating method thereof

1 . An operating method for a user knowledge tracing system configured to receive exercise information and response information and output correct answer probability information and including a first embedding execution unit, a second embedding execution unit, a plurality of encoder neural networks and a plurality of decoder neural networks,

wherein the plurality of encoder neural networks include a first encoder neural network and a second encoder neural network,

wherein the plurality of decoder neural networks include a first decoder neural network and a second decoder neural network,

wherein the first encoder neural network has a transformer model architecture and includes an exercise processing unit and a first non-linearization performing unit,

wherein the first decoder neural network has a transformer model architecture and includes a first response processing unit, a second response processing unit and a second non-linearization performing unit,

wherein the operating method comprises:

receiving, by the first encoder neural network, only the exercise information as input among the exercise information and the response information, wherein the receiving of only the exercise information as input comprises performing, by the first embedding execution unit, embedding of only the exercise information among the exercise information and the response information, the exercise information being composed of a plurality of exercises E1, E2, . . . , Ek and inputting, by the first embedding execution unit, the embedded exercise information composed of the plurality of exercises E1, E2, . . . , Ek to the exercise processing unit of the first encoder neural network, wherein the exercise information includes (i) exercise identification information being a unique value assigned to each of the plurality of exercises, (ii) exercise category information indicating a type of each of the plurality of exercises, and (iii) position information indicating a position of corresponding data in a sequence of an entire input data to separately indicate where the corresponding data is located in the sequence of the entire input data, and wherein the position information is embedded together with the input data, added to the embedded input data, and input to the exercise processing unit of the first encoder neural network,

generating, by the exercise processing unit, prediction data by performing operations of self-attention on the embedded exercise information,

performing, by the first non-linearization performing unit, non-linearization on the prediction data to generate non-linearized prediction data, the generated non-linearized prediction data being first attention information,

performing, by the second encoder neural network, operations of self-attention and non-linearization on the first attention information to generate second attention information, wherein the operations of self-attention and non-linearization performed by the second encoder neural network correspond to the operations of self-attention and non-linearization performed by the first encoder neural network,

receiving, by the first decoder neural network, only the response information as input among the exercise information and the response information, wherein the receiving of only the response information as input comprises performing, by the second embedding execution unit, embedding of only the response information among the exercise information and the response information, the response information being composed of a plurality of responses R1, R2, . . . , Rk−1 of a user, and inputting, by the second embedding execution unit, the embedded response information composed of the plurality of responses R1, R2, . . . , Rk−1 to the first response processing unit of the first decoder neural network, wherein the response information includes (i) position information indicating a position of corresponding data in a sequence of an entire input data to separately indicate where the corresponding data is located in the sequence of the entire input data, and (ii) response accuracy information indicating whether the user's response is a correct answer or an incorrect answer, and wherein the position information is embedded together with the input data, added to the embedded input data, and input to the first response processing unit of the first decoder neural network,

generating, by the first response processing unit, prediction data by performing operations of self-attention on the embedded response information, the prediction data generated by the first response processing unit being query data,

generating, by the second response processing unit, a correct answer probability of the exercise Ek based on the embedded response information and the second attention information by applying the second attention information as a weight for the query data,

performing, by the second non-linearization performing unit, non-linearization on information output from the second response processing unit,

training the user knowledge tracing system by using the second attention information as the weight for the query data, and

performing, by the second decoder neural network, operations of self-attention and non-linearization based on the correct answer probability of the exercise Ek to generate information corresponding to the correct answer probability of the exercise Ek, and

wherein the operations of self-attention and non-linearization performed by the second decoder neural network correspond to the operations of self-attention and non-linearization performed by the first decoder neural network, and

wherein the operating method comprises:

performing, by the exercise processing unit of the first encoder neural network, a first upper triangular masking on only the exercise information among the exercise information and the response information, and

performing, by the first response processing unit of the first decoder neural network, a second upper triangular masking on only the response information among the exercise information and the response information,

wherein the first upper triangular masking and the second upper triangular masking are respectively performed in the first encoder neural network and the first decoder neural network,

wherein each of the first upper triangular masking and the second upper triangular masking is an operation for preventing attention from being performed on information corresponding to a future position for prediction of a next exercise.

2 . The operation method of claim 1 , wherein the plurality of encoder neural networks are stacked and further include at least one encoder neural network in addition to the first and second encoder neural networks, and the plurality of decoder neural networks are stacked and further include at least one decoder neural network in addition to the first and second decoder neural networks, and an operation of generating attention information is repeatedly performed in each of the plurality of encoder neural networks.

3 . The operation method of claim 2 , further comprising:

completing training of the user knowledge tracing system and outputting information corresponding to the correct answer probability of the exercise Ek, which is a probability that the user correctly answers the exercise Ek.

4 . The operation method of claim 1 ,

wherein the plurality of exercises are expressed as vectors,

wherein the plurality of responses are expressed as vectors.

5 . A non-transitory computer-readable recording medium having recorded thereon a computer program for executing the operation method of claim 1 .

6 . A user knowledge tracing system configured to receive exercise information and response information and output correct answer probability information and including a first embedding execution unit, a second embedding execution unit, a plurality of encoder neural networks and a plurality of decoder neural networks, the system comprising:

wherein the plurality of encoder neural networks include a first encoder neural network and a second encoder neural network,

wherein the plurality of decoder neural networks include a first decoder neural network and a second decoder neural network,

wherein the first encoder neural network has a transformer model architecture and includes an exercise processing unit and a first non-linearization performing unit,

wherein the first decoder neural network has a transformer model architecture and includes a first response processing unit, a second response processing unit and a second non-linearization performing unit,

wherein the first encoder neural network is configured to receive only the exercise information as input among the exercise information and the response information,

wherein the first embedding execution unit is configured to:

perform embedding of only the exercise information among the exercise information and the response information, the exercise information being composed of a plurality of exercises E1, E2, . . . , Ek and input the embedded exercise information composed of the plurality of exercises E1, E2, . . . , Ek to the exercise processing unit of the first encoder neural network, wherein the exercise information includes (i) exercise identification information being a unique value assigned to each of the plurality of exercises, (ii) exercise category information indicating a type of each of the plurality of exercises, and (iii) position information indicating a position of corresponding data in a sequence of an entire input data to separately indicate where the corresponding data is located in the sequence of the entire input data, and wherein the position information is embedded together with the input data, added to the embedded input data, and input to the exercise processing unit of the first encoder neural network,

wherein the exercise processing unit is configured to:

generate prediction data by performing operations of self-attention on the embedded exercise information,

wherein the first non-linearization performing unit is configured to:

perform non-linearization on the prediction data to generate non-linearized prediction data, the generated non-linearized prediction data being first attention information,

wherein the second encoder neural network is configured to:

perform operations of self-attention and non-linearization on the first attention information to generate second attention information, wherein the operations of self-attention and non-linearization performed by the second encoder neural network correspond to the operations of self-attention and non-linearization performed by the first encoder neural network,

wherein the first decoder neural network is configured to receive only the response information as input among the exercise information and the response information,

wherein the second embedding execution unit is configured to:

perform embedding of only the response information among the exercise information and the response information, the response information being composed of a plurality of responses R1, R2, . . . , Rk−1 of a user and input the embedded response information composed of the plurality of responses R1, R2, . . . , Rk−1 to the first response processing unit of the first decoder neural network, wherein the response information includes (i) position information indicating a position of corresponding data in a sequence of an entire input data to separately indicate where the corresponding data is located in the sequence of the entire input data, and (ii) response accuracy information indicating whether the user's response is a correct answer or an incorrect answer, and wherein the position information is embedded together with the input data, added to the embedded input data, and input to the first response processing unit of the first decoder neural network,

wherein the first response processing unit is configured to:

generate prediction data by performing operations of self-attention on the embedded response information, the prediction data generated by the first response processing unit being query data,

wherein the second response processing unit is configured to:

generate a correct answer probability of the exercise Ek based on the embedded response information and the second attention information by applying the second attention information as a weight for the query data,

wherein the second non-linearization performing unit is configured to:

perform non-linearization on information output from the second response processing unit,

wherein the user knowledge tracing system is trained by using the second attention information as the weight for the query data, and

wherein the second decoder neural network is configured to:

perform operations of self-attention and non-linearization based on the correct answer probability of the exercise Ek to generate information corresponding to the correct answer probability of the exercise Ek,

wherein the operations of self-attention and non-linearization performed by the second decoder neural network correspond to the operations of self-attention and non-linearization performed by the first decoder neural network,

wherein the exercise processing unit of the first encoder neural network is configured to perform a first upper triangular masking on only the exercise information among the exercise information and the response information,

wherein the first response processing unit of the first decoder neural network is configured to perform a second upper triangular masking on only the response information among the exercise information and the response information,

wherein the first upper triangular masking and the second upper triangular masking are respectively performed in the first encoder neural network and the first decoder neural network,

wherein each of the first upper triangular masking and the second upper triangular masking is an operation for preventing attention from being performed on information corresponding to a future position for prediction of a next exercise, and

wherein at least one of the first embedding execution unit, the second embedding execution unit, the plurality of encoder neural networks and the plurality of decoder neural networks is implemented by a hardware component of the user knowledge tracing system.