Subject Classification and Difficulty Ranking of Math Problems

Education is a stage that everyone is bound to go through. With the development of technology, the quality of education has also advanced a lot. But the improvement in quality does not improve the learning performance of students equally. E-learning tools such as mobile language learning apps and mock test sites often only provide a uniform set of learning material and study plan. As a result, one popular research trend is to identify student' learning progress with the aim of customizing instruction to better meet their needs. In this paper, deep neural network models are used to classify the subjects (e.g. algebra, geometry) of math problems and rank their difficulty levels. To address the problem of classification, we extend deep learning models based on LSTM and BERT. The experiment results show that the LSTM model initially has only 59% accuracy, but after exploiting structure of math expressions through feature engineering, the LSTM model can accomplish up to 75% accuracy, close to the accuracy (79%) of the more sophisticated BERT model. Moreover, it was found that these classification models can learn feature representation that is useful in the task of ranking difficulty level of math problems. This is verified through an experiment comparing transfer learning with direct learning. This work shows that the deep learning models can handle the subject classification and difficulty level ranking of math problems. This has the potential to generate a more refined assessment of the student's domain knowledge, and help to recommend more focused review exercises to teachers and students.