The structural-probe embedding recovers BERT syntax trees; heads track coreference

Manning et al. synthesize and extend the structural-probe metric embedding (a learned linear transform under which squared Euclidean distance approximates parse-tree edge distance) on pretrained BERT base and large [manning-etal-2020-emergent-linguistic-structure] The transform recovers parse-tree distances and depths from BERT's activation geometry layer-by-layer [manning-etal-2020-emergent-linguistic-structure] A separate attention-head analysis shows specific heads track coreference clusters [manning-etal-2020-emergent-linguistic-structure] The entry is a broader synthesis (syntax plus coreference) applying the established structural probe, not a new method, and is purely descriptive [manning-etal-2020-emergent-linguistic-structure]