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

2026-02-19 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

GPT-2 models trained on synthetic corpora demonstrate human-like typological preferences for natural markedness direction in differential argument marking (DAM) systems. Researchers extended a controlled synthetic learning method to DAM, a semantic licensing system where morphological marking depends on semantic prominence. They trained GPT-2 models on 18 distinct DAM systems and evaluated their generalization using minimal pairs. The models consistently favored systems where overt marking targets semantically atypical arguments, aligning with human language regularities. However, the models did not reproduce the strong object preference observed in human languages, where DAM more frequently targets objects over subjects. This dissociation suggests that various typological tendencies in language models may stem from different underlying mechanisms.

Key takeaway

For research scientists investigating language model typological alignment, you should note that current models replicate some human linguistic preferences, like natural markedness direction, but not others, such as the strong object preference in differential argument marking. This implies that your research into the origins of typological tendencies in LMs may need to explore distinct underlying sources for different linguistic phenomena.

Key insights

LMs show human-like markedness preferences but not object preference in differential argument marking.

Principles

• Typological preferences can emerge in LMs.
• Semantic prominence influences morphological marking.

Method

GPT-2 models were trained on 18 synthetic corpora implementing distinct differential argument marking (DAM) systems, then evaluated for generalization using minimal pairs.

In practice

• Use synthetic corpora for typological studies.
• Evaluate LM generalization with minimal pairs.

Topics

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

Best for: Research Scientist, AI Researcher, AI Scientist, NLP Engineer

Related on AIssential

• 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.
• Can I guess where you are from? Modeling dialectal morphosyntactic similarities in Brazilian Portuguese — Clustering morphosyntactic features can reveal regional dialectal patterns in Brazilian Portuguese.
• Floating or Suggesting Ideas? A Large-Scale Contrastive Analysis of Metaphorical and Literal Verb-Object Constructions — Metaphorical and literal language differences are largely construction-specific, not universally consistent.
• When Similar Means Different: Evaluating LLMs on Arabic--Hebrew Cognates — Current LLMs struggle with cross-lingual semantic disambiguation, especially false friends and loanwords, due to over-reliance on surface forms.
• Disparities In Negation Understanding Across Languages In Vision-Language Models — VLMs exhibit affirmation bias, with negation understanding varying significantly across typologically diverse languages.
• G-IdiomAlign: A Gloss-Pivoted Benchmark for Cross-Lingual Idiom Alignment — Cross-lingual idiom alignment challenges LLMs due to non-compositionality, with glosses offering semantic grounding but literal bias remaining dominant.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.