Developed BERT-based model to tag multimedia Content (youtube, WebM, pdf, HTML, ECML) in multiple languages. Designed the architecture and lead the backend development in integrating the tags into Content Search for better discoverability of 65k+ Content for 150M+ app users.

Diksha is the largest Edtech platform in India that hosts 65K Content in 15 different languages with a userbase of 150M+ students. Given the scale, Content creation and consumption are two of the most important components of the platform.

• Reuse of appropriate existing Content by a teacher looking to cover a specific topic eases Content/course creation. Reuse also reduces onboarding time for ETB use-case (view content linked to the QR code printed in the textbook using the mobile app). This is enabled by building a semantic search.

• Tagging a Content with the right Textbook topic, Grade, Board, Medium of instruction is important for its discoverability by Students/Content consumer.

Both search and discovery hence require tagging a multimedia Content published on the platform to the right textbook topics/Concepts in addition to other tags provided by the author such as Grade, Board, Medium. A pipeline was designed for the purpose that gets triggered through a Content publish (Kafka) event and writes back to Content meta in ElasticSearch. The major components of the design are listed here:

1. A multimedia parsing pipeline for multilingual Content.
2. A taxonomy of keywords extracted from different textbooks with a confidence score for different keywords computed using DBpedia ontology.
3. A scalable biLSTM based similarity scoring model between parsed Content and textbook taxonomy to tag Content to textbook keywords.
4. Extending platform Search API for semantics search using pre-trained BERT embeddings of computed Content keywords

The pipeline is open sourced as part of Daggit:

• Content/Textbook keyword extraction pipeline
• Content to Textbook tagging