Controlled Evaluation of Syntactic Knowledge in Multilingual Language Models

TL;DR

This paper evaluates the syntactic understanding of multilingual language models in low-resource languages, revealing strengths and weaknesses across different syntactic tasks and models, and highlighting biases and performance issues.

Contribution

It introduces targeted syntactic evaluation tests for Basque, Hindi, and Swahili, expanding understanding of LMs' capabilities in low-resource language contexts.

Findings

Some syntactic tasks are easy for LMs, others are challenging.

Multilingual BERT shows bias toward habitual aspect in Hindi.

XGLM-4.5B underperforms compared to similar-sized models.

Abstract

Language models (LMs) are capable of acquiring elements of human-like syntactic knowledge. Targeted syntactic evaluation tests have been employed to measure how well they form generalizations about syntactic phenomena in high-resource languages such as English. However, we still lack a thorough understanding of LMs' capacity for syntactic generalizations in low-resource languages, which are responsible for much of the diversity of syntactic patterns worldwide. In this study, we develop targeted syntactic evaluation tests for three low-resource languages (Basque, Hindi, and Swahili) and use them to evaluate five families of open-access multilingual Transformer LMs. We find that some syntactic tasks prove relatively easy for LMs while others (agreement in sentences containing indirect objects in Basque, agreement across a prepositional phrase in Swahili) are challenging. We additionally uncover issues with publicly available Transformers, including a bias toward the habitual aspect in Hindi in multilingual BERT and underperformance compared to similar-sized models in XGLM-4.5B.