Cross-Linguistic Effects in Bilingual Phoneme BabyLMs
Read the original on arXiv Computation and Language →The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The Flow has not summarised this story yet — read it at arXiv Computation and Language.
arXiv:2609.22633v1 Announce Type: new Abstract: Bilingual language models (LMs) offer a controlled setting for studying how training conditions shape second-language (L2) behaviour, but prior work ty...
The paper investigates whether language models can transfer hierarchical syntactic knowledge from a source language (L1) to a target language (L2) using a novel cross‑lingual poverty‑of‑stimulus (CPOS) experimental design. Results show that models can partially transfer L1 structural evidence to L2, producing a small but detectable hierarchical bias that varies across language pairs. The study further demonstrates that vocabulary overlap is crucial for successful transfer, indicating that cross‑lingual learning of hierarchical generalization is conditionally achievable.
The study trains small GPT‑2 models on child‑scale language data from dense home recordings (BabyView) and synthetic child‑directed dialogue (TinyDialogues), varying data amount, source, and diversity. Models trained on a single child’s environment predict that child’s language far better than others, a pattern largely driven by lexical and distributional factors as shown by a 5‑gram baseline. While training on multiple children improves predictions on pooled held‑out language, it does not yield a clear advantage on grammatical benchmarks, and naturalistic versus synthetic inputs produce distinct grammatical and lexical‑semantic patterns.
The study compares an English-only and a bilingual decoder-only model, each 310 M parameters, trained on eight diverse languages while controlling for English exposure, compute, and document overlap. After aligning on shared English vocabulary, the authors find that token embeddings appear similar, but the deeper hidden states used for prediction diverge across models. This hidden‑state mismatch grows through middle transformer layers and persists despite controls, indicating that contextual processing differs between the models. "whyItMatters":"The findings show that embedding alignment can conceal significant internal representation differences, which is crucial for any downstream work that assumes aligned multilingual models are interchangeable."
Large language models show strong cross‑lingual abilities, but previous studies have examined them only at isolated training points. This work pretrains a 1.7B multilingual model on nine languages, recording checkpoints at fine granularity to track translation development. Using word‑level translation and a new dataset, the authors find that early learning is dominated by token‑level copying and surface similarities, while later stages involve more general translation mechanisms and refined copying.
The study examines whether bilingual language models can reflect the influence of a reader’s first language (L1) on second-language (L2) reading. By comparing bilingual models trained on an L1 that matches the readers’ backgrounds to monolingual English models and bilingual models trained on other L1s, the authors find that matching L1 improves predictions of L2 reading times, especially in late-stage eye-tracking measures and among intermediate-proficiency readers. The effect is modulated by the overlap between L1 and L2 scripts, indicating that bilingual models capture cross-linguistic influence in L2 reading behavior.