Life after BERT: What do Other Muppets Understand about Language?

TL;DR

This study evaluates 29 diverse pre-trained transformer models on linguistic tasks, revealing that current models struggle with compositional questions and that architecture and training data do not predict linguistic abilities.

Contribution

It provides a comprehensive analysis of various transformer models' linguistic understanding, highlighting limitations and the lack of correlation between architecture and capabilities.

Findings

Models fail to resolve compositional questions in zero-shot settings.

Model architecture and training data size do not predict linguistic performance.

Current pre-training objectives do not enable learning compositional skills.

Abstract

Existing pre-trained transformer analysis works usually focus only on one or two model families at a time, overlooking the variability of the architecture and pre-training objectives. In our work, we utilize the oLMpics benchmark and psycholinguistic probing datasets for a diverse set of 29 models including T5, BART, and ALBERT. Additionally, we adapt the oLMpics zero-shot setup for autoregressive models and evaluate GPT networks of different sizes. Our findings show that none of these models can resolve compositional questions in a zero-shot fashion, suggesting that this skill is not learnable using existing pre-training objectives. Furthermore, we find that global model decisions such as architecture, directionality, size of the dataset, and pre-training objective are not predictive of a model's linguistic capabilities.