From traditional to large language models: a novel nlp-based model for sentiment analysis in social media

Identifier:  TFM:2338

Handle: https://hdl.handle.net/20.500.11797/TFM2338

Authors:  Arias Cámara, Daniel

Abstract:
This thesis addresses the need for Catalan sentiment analysis tools by creating a new 23,000-sample balanced corpus and developing a specialized classification model, the CSXSC. This fine-tuned, 125M-parameter, encoder-only RoBERTa model was benchmarked against two 7-billion-parameter decoder-only models trained efficiently using QLoRA. The results demonstrate the superiority of the specialized model, which achieved a final test set accuracy of 83.69\% while being over 24 times more computationally efficient. This study concludes that for this discriminative task, a smaller, architecturally appropriate model provides a more accurate and practical solution than larger, general-purpose LLMs.