Emotion Detection in Textual Records of Affective Correspondence in Spanish

Resumen

While addressing emotion detection, a specific textual-domain resource for this task may become a significant issue. There are multiple modern language corpora for emotion recognition, as well as language models trained with it. Nonetheless, how many resources for emotion detection are available for the historical textual domain in Spanish? To address this gap, we introduce a manually annotated corpus of historical correspondence in Spanish. Our work addresses the annotation challenges and the methodology for emotion labelling in this particular textual domain. Furthermore, we describe the results of the performance of existing pretrained LM and LLM in detecting emotions in the Spanish historical correspondence. Finally, we assess the performance of ensemble models based on features measured with TF-IDF relevance score. Our results show 1) moderate inter-annotator agreement; 2) LLM and pretrained models barely outperform a random classifier; 3) an ensemble model with a TF-IDF shows improvement in historical emotion detection.