Context-Aware Video Recommendation via Transcript Embeddings and LLM-Based Hashtag Generation

Abstract

Video recommendation systems usually rely on user behavior patterns and collaborative filtering methods, so they are subject to popularity bias and create filter bubbles that result in content homogeneity. This article presents a complete system implementation using video transcript embeddings and large language model-generated hashtags to enable semantic recommendations. The system uses OpenAI Whisper to convert speech into text, Sentence-BERT to create detailed text representations, GPT-4 with special prompts to extract hashtags, and FAISS with IndexIVFPQ for fast similarity searches. Comprehensive prompt engineering experiments demonstrate that domain-adaptive prompts achieve superior precision and recall in computer science content, substantially outperforming baseline prompts and platform-generated tags in the F1 score. The complete system processes videos at real-time speeds on NVIDIA A100 GPUs, constructs indexes efficiently for large video collections, and delivers top-ranked recommendations with low latency. Reproducible experiments on educational videos across computer science, mathematics, physics, and biology demonstrate significant relevance improvement over description-based search in cold-start scenarios and substantial improvement in long-tail content exposure. All implementation details, prompt templates, evaluation datasets, and performance benchmarks are provided to enable replication and extension.