PERC 2021 Abstract Detail Page
Previous Page | New Search | Browse All
| Abstract Title: | Using natural language processing to predict student problem solving performance |
|---|---|
| Abstract: | In this work we report on a pilot study where we used machine learning to predict whether students will correctly solve the classic "ballistic pendulum" problem based on an essay written by students elucidating their approach to solving the problem. Specifically, students were asked to describe the "principles, assumptions, and approximations" they used to solve the problem. Student essays were codified using the practices of natural language processing. Essays from two non-consecutive semesters were used for training/validation (N = 1441) and testing (N=1480). The final model used to make predictions was an ensemble classification scheme using random forest, eXtreme Gradient Boosting classifier (XGBoost), and logistic regression as estimators. Our accuracy in predicting students' correctness was around 80% with slightly higher accuracy in identifying students who incorrectly solved the problem and slightly lower in predicting student who correctly solved the problem. |
| Abstract Type: | Contributed Poster Presentation |
| Session Time: | Poster Session 2 Room C |
| Poster Number: | 2C-19 |
| Contributed Paper Record: | Contributed Paper Information |
| Contributed Paper Download: | Download Contributed Paper |
Author/Organizer Information | |
| Primary Contact: |
Jeremy Munsell Purdue University West Lafayette, IN 47906 Phone: 765-607-3517 |
| Co-Author(s) and Co-Presenter(s) |
N. Sanjay Rebello, Carina Rebello |
