EmoMap
Emotion mapping and the comparability of computational emotion representations.
EmoMap (Emotion Mapping in Semantic Space) is a JSPS-funded project that investigates how computational representations of emotion can be mapped between different levels of granularity. An annotation scheme may distinguish annoyance from anger, while another groups both under anger or a broader negative category. Such mappings cannot be assumed to be interchangeable: the meaning and boundaries of emotion labels depend on data, context and annotation practices.
The project examines both overlap and distinctiveness among emotion categories, with the aim of improving how textual affect is represented and compared. Its empirical work also addresses the relationships between emotion and genre, register, text length and intentionality, and evaluates practical methods for multilingual emotion and intensity detection.
JSPS KAKENHI · Grant-in-Aid for Early-Career Scientists · 24K21058 · April 2024–March 2027 · Official KAKEN project record
- How can fine-grained emotion categories be mapped to broader categories or dimensions without erasing meaningful distinctions?
- How do genre, register, text length and communicative intentionality influence computational representations of emotion?
- How robust are multilingual emotion and emotion-intensity detection methods across datasets and contexts?
- What do annotation procedures and disagreement reveal about the limits of apparently shared emotion labels?
Context and textual form matter. Project outputs have investigated differences associated with text length and intentionality, including a contrastive analysis of Reddit communities. The project’s KAKEN progress report also identifies genre and register as relevant dimensions for emotion research.
Text Length and the Function of Intentionality: A Case Study of Contrastive Subreddits (NLP4DH 2024)
Multilingual emotion recognition can be approached with practical methods. The SemEval-2025 shared-task paper evaluates multilingual emotion detection and intensity prediction. It is a documented EmoMap project output, but does not by itself establish that emotion categories are equivalent across languages.
OZemi at SemEval-2025 Task 11: Multilingual Emotion Detection and Intensity
Annotation and evaluation choices affect what emotion models learn. Closely related published work compares hard labels with soft targets that retain annotator vote distributions. The findings show structured disagreement and a trade-off between conventional classification metrics and fidelity to human variation.
Recent preprints examine whether identically named emotion labels are comparable across datasets, the limitations of reducing mixed emotions to a single polarity, and the different structures underlying human annotation disagreement. These are public research outputs, but have not been presented here as peer-reviewed journal or conference findings.
Fear Is Not Always Fear (2026 preprint) · Mixed Feelings, Single Polarity (2026 preprint) · The Structure of Disagreement in Emotion Annotation (2026 preprint)
The KAKEN project record explicitly lists the following research publications among its outputs:
- Text Length and the Function of Intentionality (2024)
- OZemi at SemEval-2025 Task 11 (2025)
- Combining Qualitative and Quantitative Computational Approaches for Literary Analysis of Finnish Novels (2025), Scandinavian Studies
For the full bibliographic record, including related publications beyond the grant’s registered outputs, see Publications.