Using LLM for Learning Objects Metadata Extraction on Educational Robotics Content

• Gabriel Batistuta Urbano Lopes UERN
• Ryllari Raianne Marques De Santana Costa UERN / UFERSA
• Sebastião Emidio Alves Filho UERN / UFERSA

Resumo

RepositORE is a collaborative web platform for registering learning objects and educational robotics activities. Manually adding new objects - with their 32 metadata fields each - quickly becomes a tedious and discouraging process for users. This complexity discourages educators and developers from incorporating new resources and maintaining the repository up-to-date. Automating the extraction and completion of metadata would reduce errors, save time, and encourage sharing new content, thereby strengthening the robotics education ecosys-tem. Large Language Models (LLMs), pre-trained on vast text corpora, are capable of understanding the nuances of human language and generating contextualized responses. Owing to their prior training across diverse domains, they can handle questions and commands on a wide variety of subjects, thereby enhancing educational applications and efficiently automating complex tasks. This article presents a methodology for automatically evaluating and capturing educational robotics content from text content. This methodology employs Prompt Engineering (PE) to guide the LLMs in locating and extracting from heterogeneous sources, improving the precision and consistency of the outputs. The developed subsystem, RobotFinder v3, extracts metadata from web sites that contain relevant material and incorporates it into RepositORE. It was also applied to extract robotics-specific metadata through the analysis of YouTube video subtitles using LLMs. For this purpose, we employed the Llama 3.1, DeepSeek R-l, and Gemma 3 models for relevance assessment and data extraction. In our experiments, Llama 3.1 and Gemma 3 demonstrated superior performance in this highly specialized context.

Palavras-chave: Training, Video on demand, Large language models, Conferences, Metadata, Web sites, Prompt engineering, Data mining, Robots, Videos, Learning Objects, Educational Robotics, Prompt Engineering, Large Language Models