Do Syntactic Probes Probe Syntax? Experiments with Jabberwocky Probing

TL;DR

This paper questions whether syntactic probes truly measure syntax encoding in neural language models by testing them on semantically nonsensical yet syntactically correct Jabberwocky sentences, revealing limitations in current probing methods.

Contribution

The study introduces a novel Jabberwocky-based evaluation dataset to critically assess the validity of syntactic probing in neural models, highlighting potential confounds from semantic cues.

Findings

Probes perform worse on Jabberwocky data, indicating they may not isolate syntax effectively.

Probes still outperform baselines but with significantly reduced accuracy, e.g., 53% decrease for BERT.

Results challenge the assumption that high probe accuracy directly reflects syntactic knowledge.

Abstract

Analysing whether neural language models encode linguistic information has become popular in NLP. One method of doing so, which is frequently cited to support the claim that models like BERT encode syntax, is called probing; probes are small supervised models trained to extract linguistic information from another model's output. If a probe is able to predict a particular structure, it is argued that the model whose output it is trained on must have implicitly learnt to encode it. However, drawing a generalisation about a model's linguistic knowledge about a specific phenomena based on what a probe is able to learn may be problematic: in this work, we show that semantic cues in training data means that syntactic probes do not properly isolate syntax. We generate a new corpus of semantically nonsensical but syntactically well-formed Jabberwocky sentences, which we use to evaluate two probes trained on normal data. We train the probes on several popular language models (BERT, GPT, and RoBERTa), and find that in all settings they perform worse when evaluated on these data, for one probe by an average of 15.4 UUAS points absolute. Although in most cases they still outperform the baselines, their lead is reduced substantially, e.g. by 53% in the case of BERT for one probe. This begs the question: what empirical scores constitute knowing syntax?