Differences in Typological Alignment in Language Models’ Treatment of Differential Argument Marking

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

Summary

A study published in the Proceedings of the 30th Conference on Computational Natural Language Learning in July 2026 investigates language models' (LMs) typological alignment with Differential Argument Marking (DAM), a semantic licensing system. Researchers trained GPT-2 models on 18 distinct synthetic corpora, each implementing a unique DAM system, and evaluated their generalization using minimal pairs. The findings reveal that models reliably exhibit human-like preferences for natural markedness direction, where overt marking targets semantically atypical arguments. However, the models do not reproduce the strong object preference observed in human languages, where DAM more frequently targets objects over subjects. This dissociation suggests that different typological tendencies in LMs may originate from distinct underlying sources.

Key takeaway

For research scientists investigating linguistic universals in LMs, this work highlights that models may selectively acquire typological preferences. You should recognize that LMs can align with natural markedness direction but might miss other human language regularities, such as the strong object preference in Differential Argument Marking. This suggests a need to refine synthetic training data or evaluation methods to fully capture and assess the nuanced linguistic biases present in LMs.

Key insights

LMs align with some, but not all, human typological preferences for Differential Argument Marking.

Principles

• LMs can learn human-like typological preferences.
• Typological tendencies may have distinct origins.
• Semantic prominence influences LM marking.

Method

Trained GPT-2 models on 18 synthetic corpora implementing distinct Differential Argument Marking (DAM) systems, evaluating generalization with minimal pairs.

In practice

• Design synthetic corpora for linguistic studies.
• Test LM alignment with specific linguistic phenomena.
• Investigate semantic licensing systems in LMs.

Topics

• Language Models
• Typological Alignment
• Differential Argument Marking
• GPT-2
• Synthetic Corpora
• Linguistic Universals

Best for: AI Scientist, Research Scientist

Related on AIssential

• Differences in Typological Alignment in Language Models' Treatment of Differential Argument Marking — LMs show human-like markedness preferences but not object preference in differential argument marking.
• Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models — Automated metrics can identify and quantify lexical overuse and preference learning shifts in LLMs without manual curation.
• Presupposition and Reasoning in Conditionals: A Theory-Based Study of Humans and LLMs — LLMs often match human linguistic judgments via surface patterns, not deep pragmatic reasoning.
• Disparities In Negation Understanding Across Languages In Vision-Language Models — VLMs exhibit affirmation bias, with negation understanding varying significantly across typologically diverse languages.
• Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer — Cross-lingual fine-tuning primarily improves task-format alignment, not language-specific knowledge transfer, especially for weaker models.
• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.

Open in AIssential →

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