Derivational probing

methods / Theoretical / Analytical / Geometric analysis / Derivational probing

Techniqueadvanced

Combines the structural probe's fitted tree-metric-probe distance with Tenney et al.'s scalar-mixing 'expected layer' metric to ask not just whether a syntactic subgraph is recoverable, but at which layer, on average, each specific subgraph (e.g. a subject noun phrase vs. the whole sentence's root-verb structure) is constructed — turning a single whole-tree recovery score into a per-subgraph construction timeline.

Used in (1 observation)

BERT derives local syntax before global; GPT-2 builds both in parallelmeasured

structure: Tree Metric Embedding · models: BERT-base-cased, BERT-large-cased, GPT-2-small, GPT-2-Medium · paper: Derivational Probing: Unveiling the Layer-wise Derivation of Syntactic Structures in Neural Language Models