Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866918210040758272 |
|---|---|
| author | de Seyssel, Maureen Chi, Jie Seto, Skyler ter Hoeve, Maartje Fedzechkina, Masha Schluter, Natalie |
| author_facet | de Seyssel, Maureen Chi, Jie Seto, Skyler ter Hoeve, Maartje Fedzechkina, Masha Schluter, Natalie |
| contents | We introduce a set of training-free ABX-style discrimination tasks to evaluate how multilingual language models represent language identity (form) and semantic content (meaning). Inspired from speech processing, these zero-shot tasks measure whether minimal differences in representation can be reliably detected. This offers a flexible and interpretable alternative to probing. Applied to XLM-R (Conneau et al, 2020) across pretraining checkpoints and layers, we find that language discrimination declines over training and becomes concentrated in lower layers, while meaning discrimination strengthens over time and stabilizes in deeper layers. We then explore probing tasks, showing some alignment between our metrics and linguistic learning performance. Our results position ABX tasks as a lightweight framework for analyzing the structure of multilingual representations. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_17747 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks de Seyssel, Maureen Chi, Jie Seto, Skyler ter Hoeve, Maartje Fedzechkina, Masha Schluter, Natalie Computation and Language We introduce a set of training-free ABX-style discrimination tasks to evaluate how multilingual language models represent language identity (form) and semantic content (meaning). Inspired from speech processing, these zero-shot tasks measure whether minimal differences in representation can be reliably detected. This offers a flexible and interpretable alternative to probing. Applied to XLM-R (Conneau et al, 2020) across pretraining checkpoints and layers, we find that language discrimination declines over training and becomes concentrated in lower layers, while meaning discrimination strengthens over time and stabilizes in deeper layers. We then explore probing tasks, showing some alignment between our metrics and linguistic learning performance. Our results position ABX tasks as a lightweight framework for analyzing the structure of multilingual representations. |
| title | Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2505.17747 |