Semantic Search Engine for Video Content Powered by SBERT
Watch the demo here: https://github.com/adelelwan24/Semantic-Search-project/assets/72202195/cb0a3074-9812-4d23-a2ad-5cbd500e5a5d
The aim of this project is to simplify the process of searching within videos by leveraging semantic search using the SBERT model. Rather than requiring users to watch an entire video or listen to a full podcast, we allow them to search semantically for the exact segment they need.
Our target users include students, educators, and researchers, especially those interacting with educational content. By enabling efficient and intelligent search, we aim to enhance the learning experience, increase accessibility, and save valuable time.
-
Built using SBERT, our system delivers highly relevant search results based on meaning, not just keywords.
-
Includes support for:
- Keyword search
- Phrase-based queries
- Advanced filters to refine results
-
Users can browse transcripts with synchronized timestamps, enabling fast location of the desired video snippet.
-
Users can preview video segments before watching, making it easier to verify relevance.
-
The interface is fully responsive and works seamlessly across desktop, tablet, and mobile devices.
-
Robust user authentication system ensures data privacy and protection for all users.
We curated two specialized datasets from the ArXiv research archive, focusing on the Artificial Intelligence domain within computer science:
-
Title-based dataset
- 📁 Kaggle
- 🤗 Hugging Face
-
Synthetic query-based dataset
- 📁 Kaggle
We fine-tuned two models on our AI-specific datasets. The model trained on the title-based dataset achieved the best performance in semantic relevance and search accuracy.
This project significantly enhances the way users interact with video content—especially in academic settings—by introducing intelligent semantic search. Whether you're studying, teaching, or researching, this tool makes it easier to find what matters most, quickly and efficiently.
Let me know if you’d like a version formatted for a GitHub README, a website landing page, or technical documentation!
| Model Name | v2-Titles-wiht_150000_samples | Base | v2-Query_150000_105000_105000 | Base |
|---|---|---|---|---|
| Corpus Size | 150k | 150k | 150k | 150k |
| Queries Size | 22.5k | 22.5k | 22.5k | 22.5k |
| Acc@5 | 83.45% | 75.05% | 63.49% | 60.09% |
| Acc@10 | 87.78% | 79.98% | 71.58% | 67.41% |
| Acc@1oo | 96.06% | 91.04% | 90.79% | 85.61% |
| Precision@1 | 65.53% | 58.27% | 41.06% | 39.79% |
| Precision@3 | 26.43% | 23.67% | 18.94% | 18.09% |
| Precision@5 | 16.69% | 15.01% | 12.70% | 12.02% |
| Precission@10 | 8.78% | 8.00% | 7.16% | 6.74% |
| Recall@3 | 79.30% | 71.01% | 56.82% | 54.26% |
| Recall@5 | 83.45% | 75.05% | 63.49% | 60.09% |
| Recall@10 | 87.78% | 79.98% | 71.58% | 67.41% |
| Recall@100 | 96.06% | 91.04% | 90.79% | 85.61% |
| MRR@10 | 0.7327 | 0.6557 | 0.5062 | 0.4854 |
| MRR@100 | 0.7364 | 0.6604 | 0.5143 | 0.4929 |
| NDCG@10 | 0.768 | 0.6905 | 0.5564 | 0.5307 |
| NDCG@100 | 0.7858 | 0.7139 | 0.597 | 0.5688 |