L1 Influence in L2 Language Models: A Human-centric Approach

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A study investigates whether L2 language models (L2LMs) exhibit first language (L1) influence, mirroring patterns seen in human language learners. Researchers employed Native Language Identification (NLI) as a detection method, utilizing existing learner corpora and a novel L2 English dataset. They discovered that text length, but not proficiency, significantly affects NLI performance. Critically, instruction tuning L2LMs on human learner essays proved necessary for detectable L1 influence, achieving approximately 90% NLI accuracy. While NLI accuracy was similar for both L2LM and human essays, human evaluation revealed that LM-generated L1 influence remains distinguishable from authentic human writing.

Key takeaway

For NLP engineers developing L2 language models aiming for human-like L2 production, instruction tuning on human learner essays is critical. This approach enables your L2LMs to exhibit detectable L1 influence, achieving high NLI accuracy (~90%). However, be aware that human evaluators can still differentiate LM-generated L1 influence, indicating a need for further refinement to truly replicate human language acquisition nuances.

Key insights

L2LMs can exhibit human-like L1 influence, but instruction tuning on human data is essential for its detection via NLI.

Principles

• Human L2 production shows L1 influence.
• NLI effectively detects L1 influence in L2 text.
• Instruction tuning is vital for L2LM L1 mimicry.

Method

The study applied Native Language Identification (NLI) to L2LM-generated text under various instruction-tuning and prompting conditions, utilizing existing and novel L2 English datasets.

In practice

• Apply NLI to assess L2LM L1 influence.
• Instruction tune L2LMs with human learner essays.
• Consider text length for NLI accuracy.

Topics

• L2 Language Models
• L1 Influence
• Native Language Identification
• Instruction Tuning
• Learner Corpora
• Human Language Acquisition

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• The Sequence AI of the Week #859: Reading Claude’s Mind in English: A Note on Natural Language Autoencoders — Natural Language Autoencoders provide direct, unsupervised English explanations of LLM activations.
• LLMs Can Better Capture Human Judgments--With the Right Prompts — Simple prompting strategies, including eliciting statistical distributions and ensuring scenario clarity, significantly improve LLM alignment with human judgment.
• A 1B humanizer that matches human writing on an AI detector — A 1B-parameter humanizer can generate AI text that evades current AI detectors, matching human writing.
• Probing for Reading Times — Early language model layers predict initial human reading effort better than scalar surprisal, which excels at later stages.
• Dango: A Strictly L1-Only Large Language Model for Studying Second Language Acquisition — Dango is a 1.8B-parameter LLM for controlled L1→L2 transfer studies, using filtered L1 data and LLM-generated L2 lessons.
• Stanford AI Lab Papers and Talks at ACL 2022 — ACL 2022 showcased diverse NLP research from Stanford, spanning model pretraining, analysis, and evaluation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.