Identificação e interpretação de debates feministas no Instagram brasileiro: uma abordagem com BERTopic e LLaMA

Abstract

This study investigates the identification and interpretation of feminist themes addressed by Brazilian profiles on Instagram, using Natural Language Processing (NLP) techniques and to- pic modeling. Aiming to analyze recurring discussions and their variations, the study applied the BERTopic technique for topic modeling and the large language model (LLM) LLaMA to enrich the interpretation of identified themes. Data collection covered over 11,000 posts from six feminist profiles, and preprocessing was conducted in Python using the spaCy library for tokenization, lemmatization, and stopword removal. The modeling identified 90 topics, in- cluding prominent themes such as domestic violence, reproductive rights, mental health, and bodily autonomy, reflecting current and relevant debates within the Brazilian context. To vali- date topic interpretation, a comparative experiment was conducted, evaluating the consistency between human responses and interpretations generated by LLaMA. Although lexical simila- rity metrics showed low correspondence, the BERTScore indicated high semantic similarity, suggesting that LLaMA effectively captured the meaning of the topics, thus confirming the alternative hypothesis that NLP techniques and LLMs can accurately identify and interpret fe- minist themes on Instagram. The relevance of this research lies in its contribution to the field of social discourse analysis on digital networks, offering a structured and innovative approach to understanding feminist debates on a widely accessible platform like Instagram. The results un- derscore the importance of NLP techniques and large language models in interpreting complex social themes, providing a solid foundation for future studies on the impact of social media on awareness and the promotion of social change.