cd /news/natural-language-processing/does-typological-distance-affect-cro… · home › topics › natural-language-processing › article
[ARTICLE · art-157444] src=aclanthology.org ↗ pub= topic=natural-language-processing verified=true sentiment=· neutral

Does Typological Distance Affect Cross-lingual Transfer?

Pengfei Ren's BabyLM 2026 paper finds that coarse near/far typological distance between a first language (L1) and second language (L2) does not produce a uniform, target-independent advantage for the near L1 when GPT-2 Small models are trained from scratch on English, Dutch, or Chinese with a fixed Qwen2.5 tokenizer and then continued on the L2 target. Across five seeds, NL→EN and ZH→EN showed nearly identical Transfer Increment for English, and for Dutch the near-L1 condition's larger Transfer Increment came mostly from zero-shot starting scores, with terminal scores differing by only 0.9 percentage points. Headroom-normalized Transfer Increment gave the same target-dependent picture, the paper reports in Proceedings of the Second BabyLM Workshop, pages 65–74, Budapest, Hungary.

read2 min views2 publishedOct 24, 2026
Does Typological Distance Affect Cross-lingual Transfer?
Image: Aclanthology (auto-discovered)
Abstract

The Contrastive-Analysis Hypothesis (CAH) proposes that similarities between a learner’s first language (L1) and second language (L2) can facilitate L2 learning, while differences can make it more difficult. We ask a related computational question: under a limited, developmentally motivated data budget, does a coarse near/far relation between L1 and L2 predict grammatical performance when a language model is trained first on L1 and then on L2? We train GPT-2 Small models from scratch on English, Dutch, or Chinese with a fixed Qwen2.5 tokenizer, and continue training them on the corresponding L2 targets. Across five seeds, we compare zero-shot starting scores, terminal L2 scores, Transfer Increment (TI; terminal minus zero-shot), and Transfer Benefit (TB; terminal minus a same-budget target-only baseline). For English, NL→EN and ZH→EN have nearly identical TI. For Dutch, the near-L1 condition has a larger TI, but most of this difference comes from the zero-shot starting scores, while the terminal scores differ by only 0.9 percentage points. Headroom-normalized TI gives the same target-dependent picture. Across the tested pairs, the near/far contrast therefore does not produce a uniform, target-independent advantage for the near L1.

- Anthology ID:
- 2026.babylm-main.6
- Volume:
- [Proceedings of the Second BabyLM Workshop](https://aclanthology.org/volumes/2026.babylm-main/)
- Month:
- Venues:
- [BabyLM](https://aclanthology.org/venues/babylm/) |[WS](https://aclanthology.org/venues/ws/)
- SIG:
- Publisher:
  • Association for Computational Linguistics
- Note:
- Pages:
  • 65–74
- Language:
- URL:
- [https://aclanthology.org/2026.babylm-main.6/](https://aclanthology.org/2026.babylm-main.6/)
- DOI:
- Cite (ACL):
- Cite (Informal):
- [Does Typological Distance Affect Cross-lingual Transfer?](https://aclanthology.org/2026.babylm-main.6/) (Ren, BabyLM 2026)
- PDF:
- [https://aclanthology.org/2026.babylm-main.6.pdf](https://aclanthology.org/2026.babylm-main.6.pdf)
── more in #natural-language-processing 4 stories · sorted by recency
── more on @pengfei ren 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/does-typological-dis…] indexed:0 read:2min 2026-10-24 · —