The emergence of number and syntax units in LSTM language models
This provides mechanistic insights into grammatical encoding in LSTMs, addressing a key bottleneck in understanding neural language models.
The study investigated how LSTMs achieve long-distance number agreement in language modeling, revealing that specific 'number units' manage this information and are influenced by syntax-tracking units, indicating genuine syntactic processing mechanisms.
Recent work has shown that LSTMs trained on a generic language modeling objective capture syntax-sensitive generalizations such as long-distance 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 long-distance 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.