Ellie Yang

Author directory

2026

Prior work evaluating emotion and affective understanding in large language models (LLMs) typically rely on predetermined label sets or focus on a singular evaluation task (e.g., emotion detection). We consider affective states, referring to the much broader variety of terms people use to label their emotional experiences. We evaluate multilingual language models’ understanding of affective states in English and Spanish through three different tasks: 1) identification, where models predict an affective state given text, 2) expression, where models generate text expressing a given affective state, and 3) verification, where models report whether a given term refers to an affective state. We show that performance on one task is not necessarily predictive of performance on another. Using these three tasks, we then begin to explore when and why models struggle to understand particular affective states compared to others. We examine systematic patterns in the affective state terms that are well and poorly understood by models, characterizing the working emotion vocabulary of LLMs.
While in-context machine translation approaches lead to notable progress for low-resource languages, it is not yet clear what kinds of linguistic information in prompts can consistently and reliably lead to performance gains. We evaluate whether English semantic representations, specifically Abstract Meaning Representation (AMR) and its extension Uniform Meaning Representation (UMR), can improve LLM-based MT via prompting. Across 12 languages, 7 models from 5 families, and multiple prompting strategies, we find that conclusions about their effectiveness are highly sensitive to the setting. We further observe that prompts using UMR tend to outperform those using AMR, particularly in few-shot scenarios and for certain models. These results highlight the need for comprehensive evaluation for in-context learning work.