Inspiration
AI education is booming and the demand for education is still growing. While AI and chatbots provide valuable feedback, they sometimes fall short in addressing the fundamental need for a benchmark against which individuals can measure their writing proficiency. However, without a clear benchmark to gauge their performance, students may struggle to assess their writing objectively and set realistic goals for improvement. As students, we often feel the need to use graders.
For teachers, It's a time-consuming process that requires a large time and energy commitment. Grading essays involves reading them carefully, analyzing them thoughtfully, and providing constructive criticism. Human graders may have biases in the process of grading essays. The efficiency of manual correction is often relatively low. In order to solve this problem, we considered the possibility of improving the efficiency and credibility of an AI essay grader.
What it does
NivelMate is a service that uses machine-learning algorithms to grade essays. It takes in thousands of existing essays and scores and incorporates NLP models to create an accurate essay evaluator. This approach enables NivelMate to provide an accurate, consistent, and objective evaluation of essays across a wide range of topics and writing styles.
How we built it
We preprocessed training data (essays with scores) by removing stop words and clustering the words together. We used a BERT model to generate embeddings for the essays that could be used for training. We used a Long short-term memory (LSTM) network created in Torch to take in the embeddings and train.
Challenges we ran into
When trying to exchange data between the front and back ends, potential logic errors and URL errors were encountered, resulting in the inability to respond interactively. And for the market, we think about how to make our products more attractive compared to the products currently on the market.
Accomplishments that we're proud of
After our unremitting efforts, we successfully solved the front-end and back-end response problems and completed the project on time before the deadline.
What we learned
We learned that teamwork is very necessary. It is precisely because our classmates in charge of the front and back ends can actively communicate with each other that we can improve efficiency and discover problems. And our products still need to keep pace with the times, update the database, and improve our algorithms.
What's next for Best grader ever
First and foremost, our primary focus would be enhancing the functionality of the website application to offer a more comprehensive grading experience. One significant improvement would involve introducing multiple categories for evaluating writing, such as grammar, spelling, and coherence. By breaking down the grading process into distinct components, users can receive more detailed feedback on their strengths and areas for improvement. The final grade output would then be calculated as the average of scores across all categories, providing a holistic assessment of the writing.
Secondly, expanding the scope of our system to cater specifically to institutions like Purdue University would be paramount. To kickstart this process, we would initiate data collection efforts to gather a substantial dataset of essays from previous Purdue writing classes, paired with the grades assigned to them. Leveraging this dataset, we would train a specialized model tailored to Purdue's unique writing standards and preferences. This model would be fine-tuned to accurately evaluate Purdue-specific essays and assign scores on a standardized scale, such as out of 10 or 100.
Log in or sign up for Devpost to join the conversation.