Request Correction

RO-ABSA: A Romanian Dataset and Baselines for Aspect-Based Sentiment Analysis

Despite the increasing use and applicability of sentiment analysis tools, a significant lack of datasets exists for low-resource or limited-resource languages, such as Romanian, which adequately address this task while considering language-specific traits. To overcome this limitation, we introduce a new dataset suitable for Aspect-Based Sentiment Analysis (ABSA) in Romanian, encompassing aspect term categorisation (ATC) and aspect-level sentiment classification (ALSC). Our dataset comprises approximately 6,250 annotated reviews with over 10,600 attributes and their corresponding polarities. We establish comprehensive baselines for each component and for the entire ABSA task. For ABSA, we evaluate two complementary strategies: (1) an end-to-end generative model that produces aspect–sentiment pairs, and (2) a pipeline combining encoder-based ATC and ALSC models. We fine-tune encoder, encoder–decoder, and decoder-only architectures and additionally test transfer learning from English for ATC. Few-shot prompting with LLaMA-3.3 and GPT-4o is also explored for comparison. Fine-tuned models consistently outperform few-shot setups: the best end-to-end ABSA model achieves an F1 score of 0.81, while the ATC and ALSC components reach 0.81 and 0.93 F1, respectively. These results highlight both the challenge of the RO-ABSA dataset and the benefits of supervised fine-tuning for Romanian ABSA.