Asking without Telling: Exploring Latent Ontologies in Contextual Representations

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6792–6812, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6792 Asking without Telling: Exploring Latent Ontologies in Contextual Representations Julian Michael1∗, Jan A. Botha2, and Ian Tenney2 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Google Research julianjm@cs.washington.edu {jabot,iftenney}@google.com Abstract The success of pretrained contextual encoders, such as ELMo and BERT, has brought a great deal of interest in what these models learn: do they, without explicit supervision, learn to encode meaningful notions of linguistic struc- ture? If so, how is this structure encoded? To investigate this, we introduce latent subclass learning (LSL): a modification to classifier- based probing that induces a latent categoriza- tion (or ontology) of the probe’s inputs. With- out access to fine-grained gold labels, LSL extracts emergent structure from input repre- sentations in an interpretable and quantifiable form. In experiments, we find strong evi- dence of familiar categories, such as a notion of personhood in ELMo, as well as novel on- tological distinctions, such as a preference for fine-grained semantic roles on core arguments. Our results provide unique new evidence of emergent structure in pretrained encoders, in- cluding departures from existing annotations which are inaccessible to earlier methods. 1 Introduction The success of self-supervised pretrained models in NLP (Devlin et al., 2019; Peters et al., 2018a; Rad- ford et al., 2019; Lan et al., 2020) on many tasks (Wang et al., 2018, 2019b) has stimulated interest in how these models work, and what they learn about language. Recent work on model analysis (Belinkov and Glass, 2019) indicates that they may learn a lot about linguistic structure, including part of speech (Belinkov et al., 2017a), syntax (Blevins et al., 2018; Marvin and Linzen, 2018), word sense (Peters et al., 2018a; Reif et al., 2019), and more (Rogers et al., 2020). Many of these results are based on predictive methods, such as probing, which measure how well a linguistic variable can be predicted from inter- mediate representations. However, the ability of ∗Work performed while at Google. Figure 1: LSL overview. A probing classifier over con- textual embeddings produces multi-class latent logits, which are marginalized into a single logit trained on binary classification. In this example, “Pierre Vinken” is identified as a named entity and assigned to latent class 2, which aligns well with the PERSON label. We treat the classes as clusters representing a latent ontol- ogy that describes the underlying representation space. Figure 2 visualizes latent logits in more detail. supervised probes to fit weak features makes it difficult to produce unbiased answers about how those representations are structured (Saphra and Lopez, 2019; Voita et al., 2019). Descriptive meth- ods like clustering and visualization explore this structure directly, but provide limited control and often regress to dominant categories such as lexical features (Singh et al., 2019) or word sense (Reif et al., 2019). This leaves open many questions: how are linguistic features like entity types, syn- tactic dependencies, or semantic …

Content truncated to 500 words