Sistema de extração de padrões de conteúdo viral em vídeos curtos via PLN e geração assistida por LLM

Abstract

This work presents a system for automated extraction of viral content patterns in short videos, segmented by niche. From these patterns, the system generates strategic content guides in an assisted manner. It collects metadata and transcripts from YouTube, refining them through natural language processing and unsupervised clustering of narrative format. Patterns for each niche are extracted through density-based clustering. A semantic similarity mechanism resolves niches not yet present in the base. The final guide, structured into opening, development, and closing, is generated by a language model and delivered through a full-stack web application. End-to-end operation was qualitatively validated for three distinct niches. The guides reflect each niche's actual pattern, not a generic recommendation. The work identifies limitations related to dependency on free third-party services, occasional translation errors, and language model instability. The absence of validation with real creators is pointed out as a priority for future work.