Sentiment and Response Priority Detection in Latvian E-mails Using Large Language Models: A Case Study for Low-Resource Languages

This study investigates the capabilities of large language models (LLMs) in addressing sentiment analysis and response prioritization in Latvian texts, focusing on e-mails used in organizational workflows. In this work, we propose a response priority estimation schema based on detected severity, urgency, sentiment, and emotional level. To detect those attributes, we established a two-phase experimental approach for detecting urgency, severity, emotional intensity, and sentiment from e-mails using LLM with a zero-shot prompt technique. Considering the challenges associated with low-resource languages such as Latvian and the domain-specific requirement to maintain on-premise data processing, this study evaluates two LLMs: Meta’s Llama3 8B and Llama3.1 70B. Results highlight significant improvements in classification accuracy through refined prompt engineering, achieving up to 96-100% format accuracy, 74.6% for urgency, $\mathbf{7 4. 3 \%}$ for severity in two-class settings, and $\mathbf{9 3. 2 \%}$ and $\mathbf{9 5. 7 \%}$ for emotional intensity and sentiment detection, respectively. However, challenges persist, including discrepancies between human evaluations and model predictions. The findings emphasize the potential of LLMs for enhancing email management in lowresource language settings while outlining further improvement directions. This research advances the field of automated severity and urgency detection using LLMs, addressing a relatively underexplored area compared to the more extensively studied domains of sentiment analysis and affective computing, particularly in the context of low-resource languages.