Thai-English target-based stock sentiment dataset for financial news with ICL-based evaluation

Sentiment analysis of financial documents provides insights into market dynamics and investment decision-making. Among various approaches, target-based sentiment analysis (TBSA) is widely studied for its ability to associate sentiment with specific stock entities. However, existing TBSA datasets are limited to a monolingual setting and lack support for low-resource languages, particularly Thai. While encoder-based models and Large Language Models (LLMs) have been widely applied in TBSA research, the effects of language setup and retrieval strategies on model performance remain underexplored. To address these gaps, we introduce a bilingual TBSA dataset comprising 10,295 Thai and 10,104 English financial news articles from 2018 to 2023. We describe the data collection, preparation, and annotation process, including investment-specific sentiment guidelines. Building on this dataset, we investigate two key research questions in financial TBSA: (1) how cross-lingual transfer and bilingual fine-tuning affect encoder-based models, and (2) how different in-context learning (ICL) settings influence LLMs. Our findings highlight the importance of language-aware modeling choices for encoder-based systems in financial NLP, with the best weighted F1 scores reaching 0.773 for Thai and 0.691 for English. For LLMs, we observe that performance benefits more from well-designed instructions and domain-specific criteria than from merely increasing the number of examples, with the best weighted F1 scores reaching 0.892 for Thai and 0.767 for English. These insights suggest that prompt quality–especially in terms of clarity, contextual alignment, and feedback-informed selection–is more impactful than quantity, offering practical guidance for designing effective LLM applications in bilingual financial analysis.