Do Neural Language Models Show Preferences for Syntactic Formalisms?

Do Neural Language Models Show Preferences for Syntactic Formalisms? Artur Kulmizev Uppsala University artur.kulmizev@lingfil.uu.se Vinit Ravishankar University of Oslo vinitr@ifi.uio.no Mostafa Abdou University of Copenhagen abdou@di.ku.dk Joakim Nivre Uppsala University joakim.nivre@lingfil.uu.se Abstract Recent work on the interpretability of deep neural language models has concluded that many properties of natural language syntax are encoded in their representational spaces. However, such studies often suffer from lim- ited scope by focusing on a single language and a single linguistic formalism. In this study, we aim to investigate the extent to which the semblance of syntactic structure captured by language models adheres to a surface-syntactic or deep syntactic style of analysis, and whether the patterns are consis- tent across different languages. We apply a probe for extracting directed dependency trees to BERT and ELMo models trained on 13 different languages, probing for two different syntactic annotation styles: Universal Depen- dencies (UD), prioritizing deep syntactic re- lations, and Surface-Syntactic Universal De- pendencies (SUD), focusing on surface struc- ture. We find that both models exhibit a pref- erence for UD over SUD — with interesting variations across languages and layers — and that the strength of this preference is correlated with differences in tree shape. 1 Introduction Recent work on interpretability in NLP has led to the consensus that deep neural language models trained on large, unannotated datasets manage to encode various aspects of syntax as a byproduct of the training objective. Probing approaches applied to models like ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2019) have demonstrated that one can decode various linguistic properties such as part-of-speech categories, dependency relations, and named-entity types directly from the internal hidden states of a pretrained model (Tenney et al., 2019b,b; Peters et al., 2018b). Another line of work has tried to tie cognitive measurements or theories of human linguistic processing to the machinations of language models, often establishing strong par- allels between the two (Prasad et al., 2019; Abnar et al., 2019; Gauthier and Levy, 2019). As is the case for NLP in general, English has served as the de facto testing ground for much of this work, with other languages often appearing as an afterthought. However, despite its ubiquity in the NLP literature, English is generally consid- ered to be atypical across many typological dimen- sions. Furthermore, the tendency of interpreting NLP models with respect to existing, canonical datasets often comes with the danger of conflat- ing the theory-driven annotation therein with sci- entific fact. One can observe this to an extent with the Universal Dependencies (UD) project (Nivre et al., 2016), which aims to collect syntactic an- notation for a large number of languages. Many interpretability studies have taken UD as a basis for training and evaluating probes, but often fail to mention that UD, like all annotation schemes, is built upon specific theoretical assumptions, which may not be universally accepted. Our research questions start from these concerns. When probing language models for syntactic de- pendency structure, is UD — with its emphasis …

Content truncated to 500 words