Muravjov, Arseni (2026) Emotion Geometry across Languages: Projection-Based Emotion Transfer in Multilingual Language Models. Bachelor's Thesis, Artificial Intelligence.
|
Text
AI2026MuravjovA.pdf Download (5MB) | Preview |
Abstract
Multilingual emotion classification is challenging due to scarce emotion-labeled data in many languages. This thesis investigates whether multilingual transformer models contain transferable emotion geometry by constructing emotion subspaces from weakly supervised, emoji-labeled English and Arabic datasets. Emotion representations were extracted from frozen multilingual models, including mBERT, XLM-RoBERTa, mDeBERTa, and multilingual Llama models, and evaluated on German, Russian, Mandarin Chinese, Ukrainian, Estonian, and Swahili. Several projection-based transfer methods, including linear probes, centroid classification, cosine-similarity weighting, orthogonal Procrustes alignment, bias calibration, and low-dimensional emotion-space classifiers, were compared with machine-translation and lexicon baselines. The results show that raw embedding spaces do not consistently reveal clear emotional structure, while probe-derived probability and weight spaces reveal more interpretable patterns. Procrustes alignment improved cross-lingual transfer performance, with the six-dimensional emotion-space classifier often achieving the strongest results. The best projection-based methods were competitive with, and in some cases exceeded, machine-translation baselines while outperforming lexicon-based classification. Overall, the findings suggest that multilingual models encode a partially shared emotional geometry that supports data-efficient cross-lingual emotion classification.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Tashu, T.M. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 16 Jul 2026 11:57 |
| Last Modified: | 16 Jul 2026 11:57 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37424 |
Actions (login required)
 |
View Item |