What it does

This platform takes us closer to an end-to-end automated answer correctness checking system. The interface enables teachers to upload text comprehension tests (in the form of texts and questions about them) and enables students to be evaluated on their submitted answers.

How we built it

We have used data sets available online to produce a new data set with around 250 thousand entries, where each entry is a quadruple consisting of a paragraph, a question about it, a possible answer and a true-false label evaluating the correctness of the answer. We have built a deep neural network, trained it with the data and integrated it to the web. Python was used for the extraction of data from two SQuAD data sets. TensorFlow and Keras models were used to create and train a deep neural network (DNN). The input to DNN is a paragraph and an answer, output - true or false whether the answer is correct. Python, CSS, JavaScript, HTML, angular.js were used to create a website which integrates all bits of the application.

Challenges we ran into

It takes around 8 hours to train a neural network. Moreover, many services (Google, Amazon, Microsoft) do not allow to use given credits to train a neural network on GPUs, so we had to train the network on CPUs.

What's next for LearnIt

We have planned to use Amazon Alexa, Amazon Lex, Microsoft Azure and other to:

  1. Provide voice input and voice output to make the interaction with the website easier
  2. Use personal assistants which can read the text, ask a question, listen and evaluate the answer.
  3. Use Vision APIs to allow questions not only about texts but also about images.
  4. Train a model which can perform paraphrasing a text to augment the data set used for training and testing.
  5. Experiment with deeper neural networks architectures

Built With

Share this project:
×

Updates