Nudge

On entering the website, the user will be greeted with a home page that will prompt them to enter a topic or some text
After entering the topic, the user will receive 5 questions that will test their understanding of the topic and they can answer them.
The AI will then give them feedback on their understanding of the topic.
Once finished they can choose to retry with another text.

Inspiration

Are any of these learning-related situations relatable to you as they are to us ?

When teachers ask question but no one is answering and looking at each other instead. Teachers like to ask question and encourage question asking during class to test student understanding about a subject and keep the class engaged, but do all students actually want to answer? Not likely. Among several paper, a 2021 (Nadiel et. al) research indicated that almost half of the surveyed undergrads were afraid to answer or ask questions with common underlying factor to both phenomena being the fear of negative evaluation or feedback.
When you are learning a new topic or reading new articles, usually ones with abstract concept, and not sure whether you actually grasps the entire thing you have read but were too exhausted to reread it again from the beginning. Immediately rereading a chapter of textbook or even just an article could be exhaustive and time consuming. In fact, a 2009 (Callender & McDaniel) research shows that "immediate rereading may have little or no benefit for improving performance on educationally relevant summative assessments."

Based on these two scenarios, we came up with an important question: Is there a better, more convenient, and a faster way than everybody can test their own understanding on any topics of their choices? With this question in our mind, we have sought out an ambitious and innovative mission: to revolutionize the way people of all ages and educational backgrounds engage with and comprehend any concepts effectively and concisely without being judged on their level of understanding of a topic. Let us present you ... Nudge!

What Nudge does?

Nudge invites users to input any given topic or even an excerpt of text (it can be just simply a newspaper article you just read), and in return, it provides them with a thoughtful selection of five questions. These questions serve as prompts for users to articulate their understanding of the topic, fostering a deep and comprehensive grasp of the subject matter. After answering each question, a feedback will be prompted with possible link entities and recommended external resources if any.

How we built Nudge

We crafted a dynamic web interface utilizing React JS with Tailwind for a polished frontend experience. Employing the Fetch API, we seamlessly integrated Azure Language, Azure Bing Search, and Text Davinci, specifically for question generation and feedback generation. Text Davinci's succinct responses proved invaluable, driving our choice over GPT-3.5. Azure's tools facilitated recommendation generation and entity marking within feedback. Users can engage by inputting text, generating questions, providing answers, and receiving tailored feedback and recommendations. Project deployment was swift with Vercel, while our codebase's readability was enhanced through multiple function abstractions. Finally, rigorous user testing, including usage of try...catch blocks and console logs, ensured a robust application foundation for this project.

For full code reference, please refer to our git repository below: link.

Why Nudge when there is ChatGPT?

As explained briefly above, the choice of using Davinci model is because of its superior succinct response make it faster and more convenient for users. Moreover, our app bypass all the need of prompt-engineering just to make the ChatGPT return the satisfied format of response and interaction. Moreover, we provide a more user friendly interface designed specifically for question and answer interaction without the need of creating an account(As of the version on 6 December 2023).

Challenges we ran into when building Nudge

AI bugs / wrong information with inconsistencies from different AI APIS that we use Time crunch There is limited free use of openAI API

What we learned

How to use Azure AI Services as well as the OpenAI API

What's next for Nudge

Customized number of questions with max limit
A possible upgrade to better version if needed of GPT model
New features such as tracking progress, strengths, weaknesses, etc
Adding more LLM models so that the application is not dependent on just one model.
Scaling if enough customers.

Research References

Callender, Aimee & Mcdaniel, Mark. (2009). The limited benefits of rereading educational texts. Contemporary Educational Psychology. 34. 30-41. 10.1016/j.cedpsych.2008.07.001.

Nadile EM, Alfonso E, Barreiros BM, Bevan-Thomas WD, Brownell SE, Chin MR, Ferreira I, Ford SA, Gin LE, Gomez-Rosado JO, Gooding G, Heiden A, Hutt AE, King ML, Perez SG, Rivera Camacho YI, Salcedo F, Sellas CF, Sinda KA, Stahlhut KN, Stephens MD, Wiesenthal NJ, Williams KD, Zheng Y, Cooper KM. Call on me! Undergraduates' perceptions of voluntarily asking and answering questions in front of large-enrollment science classes. PLoS One. 2021 Jan 12;16(1):e0243731. doi: 10.1371/journal.pone.0243731. PMID: 33434226; PMCID: PMC7802933.