Recent work has shown that LSTMs trained on a generic language modeling objective capture syntax-sensitive generalizations such as longdistance number agreement. We have however no mechanistic understanding of how they accomplish this remarkable feat. Some have conjectured it depends on heuristics that do not truly take hierarchical structure into account. We present here a detailed study of the inner mechanics of number tracking in LSTMs at the single neuron level. We discover that longdistance number information is largely managed by two "number units". Importantly, the behaviour of these units is partially controlled by other units independently shown to track syntactic structure. We conclude that LSTMs are, to some extent, implementing genuinely syntactic processing mechanisms, paving the way to a more general understanding of grammatical encoding in LSTMs.

A sentence is more than the sum of its words: its meaning depends on how they combine with one another. The brain mechanisms underlying such semantic composition remain poorly understood. To shed light on the neural vector code underlying semantic composition, we introduce two hypotheses: First, the intrinsic dimensionality of the space of neural representations should increase as a sentence unfolds, paralleling the growing complexity of its semantic representation, and second, this progressive integration should be reflected in ramping and sentence-final signals. To test these predictions, we designed a dataset of closely matched normal and Jabberwocky sentences (composed of meaningless pseudo words) and displayed them to deep language models and to 11 human participants (5 men and 6 women) monitored with simultaneous magneto-encephalography and intracranial electro-encephalography. In both deep language models and electrophysiological data, we found that representational dimensionality was higher for meaningful sentences than Jabberwocky. Furthermore, multivariate decoding of normal versus Jabberwocky confirmed three dynamic patterns: (i) a phasic pattern following each word, peaking in temporal and parietal areas, (ii) a ramping pattern, characteristic of bilateral inferior and middle frontal gyri, and (iii) a sentence-final pattern in left superior frontal gyrus and right orbitofrontal cortex. These results provide a first glimpse into the neural geometry of semantic integration and constrain the search for a neural code of linguistic composition.

A sentence is more than the sum of its words: its meaning depends on how they combine with one another. The brain mechanisms underlying such semantic composition remain poorly understood. To shed light on the neural vector code underlying semantic composition, we introduce two hypotheses: First, the intrinsic dimensionality of the space of neural representations should increase as a sentence unfolds, paralleling the growing complexity of its semantic representation, and second, this progressive integration should be reflected in ramping and sentence-final signals. To test these predictions, we designed a dataset of closely matched normal and Jabberwocky sentences (composed of meaningless pseudo words) and displayed them to deep language models and to 11 human participants (5 men and 6 women) monitored with simultaneous magneto-encephalography and intracranial electro-encephalography. In both deep language models and electrophysiological data, we found that representational dimensionality was higher for meaningful sentences than Jabberwocky. Furthermore, multivariate decoding of normal versus Jabberwocky confirmed three dynamic patterns: (i) a phasic pattern following each word, peaking in temporal and parietal areas, (ii) a ramping pattern, characteristic of bilateral inferior and middle frontal gyri, and (iii) a sentence-final pattern in left superior frontal gyrus and right orbitofrontal cortex. These results provide a first glimpse into the neural geometry of semantic integration and constrain the search for a neural code of linguistic composition.Significance statementStarting from general linguistic concepts, we make two sets of predictions in neural signals evoked by reading multi-word sentences. First, the intrinsic dimensionality of the representation should grow with additional meaningful words. Second, the neural dynamics should exhibit signatures of encoding, maintaining, and resolving semantic composition. We successfully validated these hypotheses in deep Neural Language Models, artificial neural networks trained on text and performing very well on many Natural Language Processing tasks. Then, using a unique combination of magnetoencephalography and intracranial electrodes, we recorded high-resolution brain data from human participants while they read a controlled set of sentences. Time-resolved dimensionality analysis showed increasing dimensionality with meaning, and multivariate decoding allowed us to isolate the three dynamical patterns we had hypothesized.