img-20161005-wa0003

English Duniya- A Personalized Learning Approach

( This blog post has been written by Ayush Sharma, Machine Learning Engineer at Zaya Learning Labs )

Tell me and I forget. Teach me and I remember. Involve me and I learn.

We, the Benjamins at Zaya, have developed English Duniya – a personalized English language learning app for kids. It is capable of helping the child by learning his/her learning and behaviour. ED places its four foundations on, marked by each necessary skill – Reading, Vocabulary, Listening and Writing.

Why Personalized

The incommensurate teacher – student ratio and inability of teachers to find every student’s weaknesses, enabled us to develop a system which can analyze and visualize the student’s performance and help him/her by recommending lessons/content, thereby making the app personalized for every student.

Moreover, pedagogically, the four skills mentioned above should be incorporated in a child’s learning concurrently.

The student grasps the essence of English when the skills go hand in hand, thereby, opening a path to analyse and recommend skill-lessons efficiently.

How did we make it personalized

With the help of the content team, we developed a knowledge map, which resembles the pedagogy of English language as a hierarchical tree. This knowledge map (k-maps, hereafter) was the first pivotal step in the direction of personalization, opening the floodgates to incorporate analysis, recommendations and inter-skill connections using Machine Learning techniques and Data Science insights.

  • ML gets its first entry as soon as the student takes a diagnosis test. The student behaviour and performance in the diagnosis test, helps the model to determine where the student is lacking, and starts recommending lessons, allowing the student to follow an ideal learning route in real time – the instant he has logged in.
  • The key task of ML, in here, is to recommend content. For a student, the model analyses his/her performance, backtracks (similar to genetic hierarchy) to lessons where the student is lacking, finds content/lessons which are prerequisites of the poorly performed lesson, and then recommends the suitable lessons (k-maps) to the student. The recommendations are in the form of a playlist of lessons, which should be exposed to the student.
  • Not only does the model analyse performance, but also is capable of defining an ideal learning path – based on the K-maps and analysis of all the student’s learning route. And when the student fails to progress in such an ideal route, the model kicks in, and recommends new routes or removes redundant content, (whatever necessary) in order to make the student attain mastery.

The ML Power

  1. With such a model, the system can help the student progress as well as can give us hidden insights which are very useful in pedagogy.
  2. The most common learning routes taken by students, will help us to validate the k-maps and define a whole new global efficient k-map altogether. This new k-map will then resemble an efficient learning route, having short learning paths.
  3. With such a model, we can apply state-of-the-art machine learning techniques facilitating cognitive science, recommendations based on similar students, finding the weak/difficult spots for students; which will enable us to help the student to learn in real time, without subjecting him/her to redundant contents which only increases the learning time.
  4. We can find the weaknesses of low performing groups, and analyse how they have done differently with respect to successful student groups. We can learn where the low performing students are lacking and why the successful students are learning better.

performance_across_units

Challenges we faced

The biggest challenge was finding out where the ML engine would work. As the app can be played offline, the ML model should analyse performance and recommend without communicating with our server. Hence, the need to execute complex algorithms arises on the client device itself.

We tackled it by simplifying much of the complex algorithms and dumped the k-maps and the ML model on the client device, so that the client doesn’t need to come online to get recommendations.

With the help of such a self-learning personalized system, which knows when/where a student goes wrong, we can change the conventional learning acutely. In this way, we can be very close to the student virtually, and can track his performance in real time which is not achieved in traditional learning. The ML model will learn more and gather dense, extensive insights when more students use this app. The day is not far, when the model will teach us how to teach.

Share this Post