When Does Syntax Mediate Neural Language Model Performance? Evidence from Dropout Probes
This work addresses a methodological issue in causal probing for NLP researchers, offering incremental improvements in probe design to better assess syntax in language models.
The paper tackled the problem of false negative causality results in syntactic probing by demonstrating that language models encode syntactic information redundantly, and introduced a new probe design that guides probes to consider all syntactic information, leading to evidence of syntax use in models where prior methods failed and boosting model performance through syntactic injection.
Recent causal probing literature reveals when language models and syntactic probes use similar representations. Such techniques may yield "false negative" causality results: models may use representations of syntax, but probes may have learned to use redundant encodings of the same syntactic information. We demonstrate that models do encode syntactic information redundantly and introduce a new probe design that guides probes to consider all syntactic information present in embeddings. Using these probes, we find evidence for the use of syntax in models where prior methods did not, allowing us to boost model performance by injecting syntactic information into representations.