Multilingual Nonce Dependency Treebanks: Understanding how Language Models represent and process syntactic structure

TL;DR

This work proposes Semantically Perturbed Universal Dependencies SPUD to create grammatically valid nonce treebanks that perturb lexical co occurrence while preserving syntactic structure. SPUD is applied to five languages and evaluated on two fronts: language model scoring with autoregressive and masked models and dependency probing with the DepProbe framework. The findings show that autoregressive perplexity is more sensitive to nonce content than masked pseudo perplexities, and that subword information mitigates some effects for MLMs; probing results reveal a robust amount of syntactic information despite semantic perturbations, with attachment predictions more affected than relation labeling. Overall, SPUD provides a principled way to study the interplay of syntax and semantics in multilingual LMs and offers data resources and a tutorial to support further research in this area.

Abstract

We introduce SPUD (Semantically Perturbed Universal Dependencies), a framework for creating nonce treebanks for the multilingual Universal Dependencies (UD) corpora. SPUD data satisfies syntactic argument structure, provides syntactic annotations, and ensures grammaticality via language-specific rules. We create nonce data in Arabic, English, French, German, and Russian, and demonstrate two use cases of SPUD treebanks. First, we investigate the effect of nonce data on word co-occurrence statistics, as measured by perplexity scores of autoregressive (ALM) and masked language models (MLM). We find that ALM scores are significantly more affected by nonce data than MLM scores. Second, we show how nonce data affects the performance of syntactic dependency probes. We replicate the findings of Müller-Eberstein et al. (2022) on nonce test data and show that the performance declines on both MLMs and ALMs wrt. original test data. However, a majority of the performance is kept, suggesting that the probe indeed learns syntax independently from semantics.

Figures (9)

• Figure 1: SPUD data creation
• Figure 2: Intrinsic evaluation results for English and Arabic. Plots for other languages are in App. \ref{['app:score-ratios']}.
• Figure 3: Arabic SPUD examples (in transliteration).
• Figure 4: German SPUD examples.
• Figure 5: English SPUD examples.