Bilingual Language Models Capture Cross-Linguistic Influence in Second-Language Reading

Forum topic · 小凯 · 2026-10-25

Summary

This paper investigates whether bilingual language models can capture cross-linguistic influence in second-language (L2) reading. While monolingual LM surprisal is well established as a predictor of reading times, it cannot account for reading dynamics shaped by a reader's first language (L1). The authors evaluate bilingual models on eye-tracking data from readers of 12 diverse L1 backgrounds reading English texts. They find that matching the reader's L1 significantly improves the prediction of L2 reading times. The predictive advantage is concentrated in late-stage eye-tracking measures, is modulated by the degree of L1-L2 script overlap, and peaks among intermediate-proficiency readers. These results provide experimental evidence that bilingual LMs align with human L1-driven biases during L2 reading. Paper by Abderrahmane Issam, Irene E. Winther, and Stefan L. Frank, available on arXiv (2610.31620).

Paper Overview

Field: NLP Authors: Abderrahmane Issam, Irene E. Winther, Stefan L. Frank Published: 2026-10-22 arXiv: 2610.31620

Abstract

While monolingual LM surprisal is widely established as a predictor of reading times, it cannot capture non-native (L2) reading dynamics shaped by the reader's first language (L1). We investigate whether cross-lingual transfer within bilingual models aligns with human L1-driven biases during L2 reading. We evaluate bilingual models on eye-tracking data of readers from 12 diverse L1 backgrounds reading English texts. Matching the readers' L1 improves the prediction of L2 reading times. This predictive advantage is concentrated in late-stage eye-tracking measures, modulated by L1-L2 script overlap, and peaks among intermediate-proficiency readers. These findings provide evidence that bilingual LMs capture aspects of cross-linguistic influence in L2 reading behavior.

Key Findings

Matching the reader's L1 in a bilingual language model improves prediction of L2 (English) reading times.
The advantage appears mainly in late-stage eye-tracking measures (e.g., later processing stages).
The effect is modulated by L1-L2 script overlap.
The benefit peaks among intermediate-proficiency L2 readers.

Summary (Chinese source)

单语LM惊奇度虽被广泛确立为阅读时间的预测因子，但无法捕捉由读者第一语言（L1）塑造的非母语（L2）阅读动态。本文在12个不同L1背景读者阅读英文的眼动追踪数据上评估双语模型，发现匹配读者L1的双语模型显著改善L2阅读时间预测，这种优势集中在晚期眼动指标中，受L1-L2文字重叠度调节，并在中等熟练度读者中达到峰值。

*Auto-collected on 2026-10-26.*