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

TL;DR

This study investigates how an LSTM language model's representations of nonsensical language stimuli relate to human brain activity, revealing similarities even with degraded semantic or syntactic information.

Contribution

It demonstrates that LSTM representations maintain a significant correlation with brain activity even when processing nonsensical language stimuli, extending understanding of model-brain similarities.

Findings

LSTM representations correlate with brain activity for nonsensical language.

Brain-LSTM similarity persists despite semantic and syntactic degradation.

LSTMs handle nonsensical data in a brain-like manner.

Abstract

The representations generated by many models 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 syntactically 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 language, there is a statistically significant relationship 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.