From Language to Language-ish: How Brain-Like is an LSTM’s Representation of Nonsensical Language Stimuli?

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 645–656 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 645 From Language to Language-ish: How Brain-Like is an LSTM’s Representation of Nonsensical Language Stimuli? Maryam Hashemzadeh1,2, *, Greta Kaufeld3,4, Martha White1,2, Andrea E. Martin3,5, and Alona Fyshe1,2,6 1 Department of Computing Science, University of Alberta, Alberta, Canada 2Alberta Machine Intelligence Institute (Amii), Alberta, Canada 3Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands 4International Max Planck Research School for Language Sciences, Nijmegen, The Netherlands 5Donders Institute for Brain, Cognition, and Behaviour, Radboud University, Nijmegen, The Netherlands 6Department of Psychology, University of Alberta, Alberta, Canada *Corresponding author, hashemza@ualberta.ca {whitem,alona}@ualberta.ca, {greta.kaufeld, andrea.martin}@mpi.nl Abstract The representations generated by many mod- els of language (word embeddings, recurrent neural networks and transformers) correlate to brain activity recorded while people read. However, these decoding results are usually based on the brain’s reaction to syntactically and semantically sound language stimuli. In this study, we asked: how does an LSTM (long short term memory) language model, trained (by and large) on semantically and syntac- tically intact language, represent a language sample with degraded semantic or syntactic information? Does the LSTM representation still resemble the brain’s reaction? We found that, even for some kinds of nonsensical lan- guage, there is a statistically significant rela- tionship between the brain’s activity and the representations of an LSTM. This indicates that, at least in some instances, LSTMs and the human brain handle nonsensical data similarly. 1 Introduction When people read or listen to language, brain imag- ing studies have shown us that the brain’s activity correlates to LSTM (long short term memory) state representations for the same text (Jain and Huth, 2018; Toneva and Wehbe, 2019). In those stud- ies (and others like them) the stimuli used to test for this correlation was based on language with no errors.1 This implies that during the processing of within-distribution data (i.e. well-formed sen- tences/stories), LSTMs and the human brain show similar representational patterns. But what happens when language is out-of-distribution (e.g. nonsen- sical sentences or pseudo-words)? Can we expect 1Nonsensical language is often used when measuring Event Related Potentials. Here we speak of decoding studies only. that an LSTM will still compute contextual states in a way that resembles how the human brain reacts? I.e. is there a correlation between LSTM represen- tations and neural activity when the stimuli is not a predictable language sample? Answering these questions could provide evidence that an LSTM is able to generalize to new data in a human-like way, even when the new data is unlike anything it en- countered during training. Our answers could also help psycholinguists reason about the efficacy of nonsensical sentences and pseudo-words as syntax- only stimuli controls. Here we use brain imaging data (Electroen- cephalography, EEG) collected in three conditions, Sentence: well-formed grammatical sentences, Jab- berwocky: pseudo-word sentences that preserved word order, morphosyntax, and sentential prosody without lexical or compositional semantics, and Word-list: the words of the Sentence condition in a pseudo-random …

Content truncated to 500 words