Neural language models as psycholinguistic subjects: Representations of syntactic state

Proceedings of NAACL-HLT 2019, pages 32–42 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 32 Neural Language Models as Psycholinguistic Subjects: Representations of Syntactic State Richard Futrell1, Ethan Wilcox2, Takashi Morita3,4, Peng Qian5, Miguel Ballesteros6, and Roger Levy5 1Department of Language Science, UC Irvine, rfutrell@uci.edu 2Department of Linguistics, Harvard University, wilcoxeg@g.harvard.edu 3Primate Research Institute, Kyoto University, tmorita@alum.mit.edu 4Department of Linguistics and Philosophy, MIT 5Department of Brain and Cognitive Sciences, MIT, {pqian,rplevy}@mit.edu 6IBM Research, MIT-IBM Watson AI Lab, miguel.ballesteros@ibm.com Abstract We investigate the extent to which the behav- ior of neural network language models reflects incremental representations of syntactic state. To do so, we employ experimental method- ologies which were originally developed in the field of psycholinguistics to study syntac- tic representation in the human mind. We ex- amine neural network model behavior on sets of artificial sentences containing a variety of syntactically complex structures. These sen- tences not only test whether the networks have a representation of syntactic state, they also re- veal the specific lexical cues that networks use to update these states. We test four models: two publicly available LSTM sequence mod- els of English (Jozefowicz et al., 2016; Gulor- dava et al., 2018) trained on large datasets; an RNN Grammar (Dyer et al., 2016) trained on a small, parsed dataset; and an LSTM trained on the same small corpus as the RNNG. We find evidence for basic syntactic state representa- tions in all models, but only the models trained on large datasets are sensitive to subtle lexical cues signalling changes in syntactic state. 1 Introduction It is now standard practice in NLP to derive sen- tence representations using neural sequence mod- els of various kinds (Elman, 1990; Sutskever et al., 2014; Goldberg, 2017; Peters et al., 2018; De- vlin et al., 2018). However, we do not yet have a firm understanding of the precise content of these representations, which poses problems for inter- pretability, accountability, and controllability of NLP systems. More specifically, the success of neural sequence models has raised the question of whether and how these networks learn robust syntactic generalizations about natural language, which would enable robust performance even on data that differs from the peculiarities of the train- ing set. Here we build upon recent work studying neural language models using experimental techniques that were originally developed in the field of psy- cholinguistics to study language processing in the human mind. The basic idea is to examine language models’ behavior on targeted sentences chosen to probe particular aspects of the learned representations. This approach was introduced by Linzen et al. (2016), followed more recently by others (Bernardy and Lappin, 2017; Enguehard et al., 2017; Gulordava et al., 2018), who used an agreement prediction task (Bock and Miller, 1991) to study whether RNNs learn a hierarchical morphosyntactic dependency: for example, that The key to the cabinets. . . can grammatically con- tinue with was but not with were. This dependency turns out to be learnable from a language mod- eling …

Content truncated to 500 words