Inspired by cognitive neuroscience studies, we introduce a novel “decoding probing” method that uses minimal pairs benchmark (BLiMP) to probe internal linguistic characteristics in neural language models layer by layer. By treating the language model as the brain and its representations as “neural activations”, we decode grammaticality labels of minimal pairs from the intermediate layers’ representations. This approach reveals: 1) Self-supervised language models capture abstract linguistic structures in intermediate layers that GloVe and RNN language models cannot learn. 2) Information about syntactic grammaticality is robustly captured through the first third layers of GPT-2 and also distributed in later layers. As sentence complexity increases, more layers are required for learning grammatical capabilities. 3) Morphological and semantics/syntax interface-related features are harder to capture than syntax. 4) For Transformer-based models, both embeddings and attentions capture grammatical features but show distinct patterns. Different attention heads exhibit similar tendencies toward various linguistic phenomena, but with varied contributions.

The Alice Datasets are a set of datasets based on magnetic resonance data and electrophysiological data, collected while participants heard a story in English. Along with the datasets and the text of the story, we provide a variety of different linguistic and computational measures ranging from prosodic predictors to predictors capturing hierarchical syntactic information. These ecologically valid datasets can be easily reused to replicate prior work and to test new hypotheses about natural language comprehension in the brain.

John Hale, Adhiguna Kuncoro, Keith Hall, Chris Dyer, Jonathan Brennan. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Recurrent neural network grammars (RNNGs) are generative models of (tree, string) pairs that rely on neural networks to evaluate derivational choices.Parsing with them using beam search yields a variety of incremental complexity metrics such as word surprisal and parser action count.When used as regressors against human electrophysiological responses to naturalistic text, they derive two amplitude effects: an early peak and a P600-like later peak.By contrast, a non-syntactic neural language model yields no reliable effects.Model comparisons attribute the early peak to syntactic composition within the RNNG.This pattern of results recommends the RNNG+beam search combination as a mechanistic model of the syntactic processing that occurs during normal human language comprehension.