Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks

TL;DR

The paper introduces ABX-style, training-free discrimination tasks to separately assess language form and semantic meaning in multilingual representations, applied to XLM-R across checkpoints and layers. It shows that language identity cues fade and localize in lower layers over training, while semantic content becomes more pronounced in deeper layers, revealing a gradual decoupling of form and meaning. The study links ABX discrimination to downstream linguistic probing and cross-linguistic transfer, finding that higher language-discriminability often correlates with poorer POS/NER transfer, while meaning-discrimination shows weaker or inconsistent ties to tasks like NLI. The ABX framework provides a lightweight, interpretable diagnostic for representational structure and offers practical hints for checkpoint selection and cross-lingual strategies, with future work extending to more architectures and typologically diverse languages.

Abstract

We introduce a set of training-free ABX-style discrimination tasks to evaluate how multilingual language models represent language identity (form) and semantic content (meaning). Inspired from speech processing, these zero-shot tasks measure whether minimal differences in representation can be reliably detected. This offers a flexible and interpretable alternative to probing. Applied to XLM-R (Conneau et al, 2020) across pretraining checkpoints and layers, we find that language discrimination declines over training and becomes concentrated in lower layers, while meaning discrimination strengthens over time and stabilizes in deeper layers. We then explore probing tasks, showing some alignment between our metrics and linguistic learning performance. Our results position ABX tasks as a lightweight framework for analyzing the structure of multilingual representations.

Figures (15)

• Figure 1: Illustration of the ABX discrimination task. $A$ and $X$ share the target variable, whereas $B$ differs. Control variables may be included, with $A$ and $B$ sharing the same control variable.
• Figure 2: Language and meaning ABX discrimination scores across checkpoints (averaged over layers and all language pairs). Baseline score is 0.5.
• Figure 3: Language and meaning ABX discrimination scores across layers (averaged over all language pairs) for the last checkpoint (step 150,000)
• Figure 4: Evolution of ABX discrimination scores across model checkpoints and layers. Dark regions indicate higher discrimination scores. (Left: LD; right: MD)
• Figure 5: Illustration of the Language Discrimination (left) and Meaning Discrimination (right) ABX tasks.