Knowledge Tracking Based On Attention Mechanism

Knowledge tracking technology is an important research content in the online intelligent education system.Based on the learning trajectory of learners,it is used to dig and analyze it,and find that learners’ learning laws and the state of knowledge of knowledge are found.At present,research in the field of knowledge tracking is mainly based on deep learning models.Among them,the high degree of attention is circulating neural network.Through research,although circulating neural networks performed very well in the modeling of local characteristics,there is a problem of long-term dependence and interpreting poor in circulating neural networks.This problem will lead to the effect of circulating neural networks for long sequence models.The attention mechanism is mainly through the point of the point of the vector,so it has no limit to the length of the input sequence.On the other hand,because the quality of the data set in the real world cannot be guaranteed enough,there is a problem of continuous training in the same question in a large number of data sets.For this issue,the topic is used by increasing the convolution layer to improve the data to improve Model perception ability.Finally,in order to simulate the phenomenon of forgetting in the learning process,the calculation of the forgotten factor was added to the model to digitize the forgotten.In response to the above three problems,this thesis proposes a knowledge tracking model model,the specific content is as follows:(1)Aiming at the problem of long sequence dependencies in circulating neural networks,in order to maintain the continuity of the learning trajectory of the learner,the learning trajectory sequence of all length can be modeling and extracting features.It can be not limited by the length of the input sequence.(2)A large number of continuous repeated training data exist in the real world data set,which will cause Transformer to have a small amount of learning records that are too high in learning records when calculating the weight matrix.Potential connection between points.In order to ensure that Transformer better perceives the dependency between different knowledge points when calculating the weight matrix,and the convolutional data of the learner’s learning trajectory,so that the dependent relationship between different knowledge points can be better.Perception.(3)In response to the forgotten phenomenon of learners during the learning process,this article calculates the forgiveness factors through the following three time interval,which are the current topic distance of the current topic.The training interval between training and the current time of the last training.And further integrate the forgiveness factor and the weight matrix,which is more in line with the human learning process,and further increases the accuracy of prediction.Through the solution of the above problems,the accuracy of the knowledge tracking model prediction is further improved.The four data sets in the real world were verified,and the results showed that it was better than 4 classic open source algorithms.And through ablation experiments,the effectiveness of the algorithm independently operates.