Skip to content

Latest commit

 

History

History
120 lines (80 loc) · 9.33 KB

README.md

File metadata and controls

120 lines (80 loc) · 9.33 KB
Raw

Interactive Tool in Regional Language to Inculcate Coding Skills in Rural Students (ITRL)

Intuition

In the modern world, programming abilities—which include critical thinking and problem-solving—are essential. However, because internet materials are primarily written in English, students in rural areas frequently lack access to high-quality instructional resources, particularly in programming. They also encounter linguistic obstacles. In order to overcome language hurdles and give rural students fair access to programming, analytical thinking, and logical reasoning education, we came up with a web application that teaches programming in Kannada.

Key Features

With the many interactive features on this platform, students can test their knowledge and experiment with different ideas.

  1. Learning Materials: A series of meticulously crafted tutorials have been developed, each focusing on essential coding concepts like variables, print statements, operators, data types, conditional statements, loops and error handling. These tutorials are created in Kannada enabling the user to understand the concepts in their regional language. Drive Link

  1. Coding Platform: The primary feature is, of course, the coding platform, where students can dictate or type algorithmic statements in Kannada and have the matching Python code generated for them. Students can use the embedded interpreter on the right side of the interface to run their code once all of the program's statements have been completed.

  1. Error detection: The embedded interpreter recognises any logical, syntactic, or other errors and users can translate the error message into Kannada so that students can comprehend it.

  1. Interactive quiz: The fourth element is an interactive quiz platform where students can solve problems to see how much they know before moving on to writing programs.

Code generation using NLP

Existing code-generation models don't meet this application's needs. We fine-tuned a pre-trained model on a custom dataset to ensure accurate results, focusing on the required input-output format.

Dataset:

In order to optimise the model, the dataset was carefully prepared by the team to fine-tune the model. There are 1250 items in this dataset, covering a variety of approaches. To make preparation simpler, the original dataset—which had been saved in CSV format—was transformed to JSON format. The JSON file had the items under the labels "text" and "code." The NL inputs were listed under the "text" label, and the corresponding and anticipated Python code output was listed under the "code" label. For every algorithm statement, two to five similar input statements with slightly different wording but the same meaning were added to the dataset, with the expected output being the same, in order to make sure the model understood the requirements of the input and related syntax automatically. Dataset

Fine Tuning

Through fine-tuning Salesforce's CodeT5-base model, the NLP model processes natural language inputs to produce programming language outputs. The model uses PyTorch for training, TensorFlow for data processing, and Hugging Face's transformers to generate comprehensive code using Google's T5 encoder-decoder architecture. 1125 elements from the custom dataset—hosted on GitHub—were designated as training and 125 as testing. TensorFlow and Roberta Tokenizer were used for byte-level Byte-Pair Encoding during the preprocessing stage. Hugging Face's Trainer class was used for Google Colab training. After it was adjusted, the model was uploaded to the Hugging Face Hub so that it could be integrated with the website.

Web application

Frontend

The frontend of this project is built using React and Chakra UI, providing a modern, responsive, and accessible user interface. Chakra UI's comprehensive set of accessible components and design principles ensures a seamless user experience across different devices and screen sizes. The application features an interactive Kannada keyboard and a voice recognition system to facilitate input in Kannada, enhancing user engagement. Users can input text either by typing or speaking, with the application translating their input into Python code. This interactive interface promotes coding education in Kannada, making it accessible to native speakers. The use of Chakra UI allows for quick development and easy customization, ensuring the application remains user-friendly and visually appealing.

Backend

The backend of this project is developed using Django, a high-level Python web framework that promotes rapid development and clean, pragmatic design. It handles several key functionalities including translating Kannada algorithms to English using the deep_translator library, converting translated text to Python code via a Hugging Face model, and providing error message translations. Secure handling of POST requests with CSRF exemption is implemented for endpoints requiring external API interactions. The backend includes a model activation function to ensure the Hugging Face model is ready for inference. All responses are managed through Django's JsonResponse to facilitate smooth communication with the frontend, which is built with React and Chakra UI. Additionally, the backend is configured to handle CORS to allow interaction with the frontend hosted on different origins during development.

Embedded interpreter

A versatile online code editor and compiler that works great in a web environment is Trinket.io. With its intuitive interface, both novices and experts can experiment with a variety of computer languages, such as HTML, CSS, JavaScript, and Python. For our application, we set it up to run in Python. Developing, publishing, and embedding code in webpages or instructional materials is made simple for users of Trinket.io, an intuitive platform for collaborative coding projects. It served as a coding interface by being integrated into the online application. The user needs to copy and paste each generated Python code expression or snippet from the text box into the editor.

Deployment

KalikeMate

API Endpoints

  1. Translate Algorithm

    • Endpoint: /api/translate/
    • Method: POST
    • Description: Translates a given Kannada algorithm to English and converts the translated text to Python code.
    • Request Body: 
      {
  "text": "<kannada_algorithm>"
}

  2. Translate Error Message

    • Endpoint: /api/translateError/
    • Method: POST
    • Description: Translates an error message from English to Kannada.
    • Request Body: 
      {
  "text": "<error_message>"
}

  3. Activate Model

    • Endpoint: /api/activate-model/
    • Method: POST
    • Description: Activates the Hugging Face model for inference.
    • Request Body: 
      {}

Demo Video

English Tutorial

Conclusion

Possessing coding and logical thinking skills is a much-needed skill today. However, students hailing from rural areas often do not know English, which hinders them from developing these skills because of the lack of resources available in their regional language. This study proposes an interactive tool that aims to solve this issue by creating a web application in Kannada, a regional language of India. The aim was to build the coding and logical thinking skills of students who did not know the English language and, thus, could not utilize traditional resources. The application consists of features, such as custom-made tutorials, quizzes, algorithm testing, and code generation features, all of which are available in Kannada. The application leverages a code-generation NLP model to support algorithm-building and testing features. The proposed application has significant scope for improvement. Future improvements could include the extension of the application to be available in other regional languages apart from Kannada and in programming languages other than Python.