Sort by Structure: Language Model Ranking as Dependency Probing

TL;DR

This paper introduces a probing method to rank pre-trained language models based on their ability to recover dependency structures, enabling efficient model selection for parsing tasks with minimal computational cost.

Contribution

It presents a novel dependency probing approach for LM ranking, covering diverse models and languages, and demonstrates its effectiveness in predicting the best models for parsing.

Findings

Probing predicts the best LM 79% of the time with less compute.

RemBERT contains less dependency info but performs well after fine-tuning.

Without RemBERT, the method identifies the best LM in 89% of cases.

Abstract

Making an informed choice of pre-trained language model (LM) is critical for performance, yet environmentally costly, and as such widely underexplored. The field of Computer Vision has begun to tackle encoder ranking, with promising forays into Natural Language Processing, however they lack coverage of linguistic tasks such as structured prediction. We propose probing to rank LMs, specifically for parsing dependencies in a given language, by measuring the degree to which labeled trees are recoverable from an LM's contextualized embeddings. Across 46 typologically and architecturally diverse LM-language pairs, our probing approach predicts the best LM choice 79% of the time using orders of magnitude less compute than training a full parser. Within this study, we identify and analyze one recently proposed decoupled LM - RemBERT - and find it strikingly contains less inherent dependency information, but often yields the best parser after full fine-tuning. Without this outlier our approach identifies the best LM in 89% of cases.