This is my first post on the forum.
My graduation project is about using deep learning and nlp to auto grade exams with short answer questions. Our idea is to have students take a picture of the their answers using a mobile app and the app will extract the answers from the picture and feed it to the model which will then grade the exam.
I don’t really know where to begin when building the model, my knowledge about deep learning is very basic; I am still watching the fastai course.
I would appreciate it very much if someone can tell me where I should start and what I need to know when building the model.
I am also a beginner, but I think one of the biggest challenge for training your model is obtaining data. If you already have graded tests, that is great. Another problem is it would be very difficult to find out why AI gave someone A or B. If students complain about their score, there is no explanation.
Therefore, it would be best to use AI grading as a guideline, and teachers still have to grade as they did before.
For more about production, I recommend you read chapter 2 of the FastBook.
As you said that the students can take a picture using a mobile app, so I think its best to start from OCR ( Optical Character Recognition) to convert images into text. I am new here so I don’t know if I am allowed to refer other libraries but in my experience PaddleOCR works great. Tesseract is also other famous library.
After converting images into text, you can use that text for grading by
Text Classification or
Named Entity Recognition.
As @galopy also mentioned that the biggest challenge is to make sure how tests are graded fairly. I think you can explore current and past Feedback Essay Competition on Kaggle as they are also using ML for grading essays and how do they mitigate the risks of biases.
For the data we plan to use the ASAG data set from North Texas university.
And as you have said we plan to use it as a guideline to help teachers with grading.
Thank you for your help!
I will check them out. Thank you!