Asking without Telling: Exploring Latent Ontologies in Contextual Representations

The success of pretrained contextual encoders, such as ELMo and BERT, has\nbrought a great deal of interest in what these models learn: do they, without\nexplicit supervision, learn to encode meaningful notions of linguistic\nstructure? If so, how is this structure encoded? To investigate this, we\nintroduce latent subclass learning (LSL): a modification to existing\nclassifier-based probing methods that induces a latent categorization (or\nontology) of the probe's inputs. Without access to fine-grained gold labels,\nLSL extracts emergent structure from input representations in an interpretable\nand quantifiable form. In experiments, we find strong evidence of familiar\ncategories, such as a notion of personhood in ELMo, as well as novel\nontological distinctions, such as a preference for fine-grained semantic roles\non core arguments. Our results provide unique new evidence of emergent\nstructure in pretrained encoders, including departures from existing\nannotations which are inaccessible to earlier methods.\n