Naruse, Hina (2026) Emotion Without Borders? Cross-Lingual Transfer and the Geometry of Emotion-Label Representations in Multilingual Language Models. Bachelor's Thesis, Artificial Intelligence.
|
Text
bAI2026NaruseH.pdf Download (2MB) | Preview |
Abstract
Multilingual language models are increasingly used in applications where recognizing expressed emotions is important, yet it remains unclear how emotions are represented across languages. This study examines frozen representations from mBERT and XLM-RoBERTa across four typologically diverse languages: English, Spanish, Arabic, and Japanese. We evaluate monolingual emotion classification and cross-lingual transfer for eight shared emotion labels, and analyze embedding geometry using centroid similarity, principal component analysis, and Procrustes alignment. We also compare encoder-based probing with zero-shot prompting using Qwen2.5-7B-Instruct. Results show that XLM-R outperforms mBERT in all monolingual settings and 11 of 12 cross-lingual directions. Cross-lingual performance remains below monolingual baselines and is particularly weak when Japanese is the target language. Geometric analyses reveal recurring emotion relationships alongside substantial language clustering, with only partial correspondence between geometry and transfer performance. Because the Japanese dataset differs in domain, size, and annotation procedure, typological effects cannot be isolated. Overall, the findings suggest that emotion representations are partially transferable across languages but remain strongly influenced by language- and dataset-specific factors.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Tashu, T.M. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 28 Jul 2026 07:01 |
| Last Modified: | 28 Jul 2026 07:01 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37472 |
Actions (login required)
 |
View Item |