On the Representational Capacity of Recurrent Neural Language Models
It addresses the theoretical expressivity of RLMs for language modeling, but is incremental as it builds on prior Turing completeness results.
This work extends Turing completeness to probabilistic recurrent neural language models (RLMs), showing they can simulate deterministic probabilistic Turing machines with rational weights, and provides lower bounds for real-time computation.
This work investigates the computational expressivity of language models (LMs) based on recurrent neural networks (RNNs). Siegelmann and Sontag (1992) famously showed that RNNs with rational weights and hidden states and unbounded computation time are Turing complete. However, LMs define weightings over strings in addition to just (unweighted) language membership and the analysis of the computational power of RNN LMs (RLMs) should reflect this. We extend the Turing completeness result to the probabilistic case, showing how a rationally weighted RLM with unbounded computation time can simulate any deterministic probabilistic Turing machine (PTM) with rationally weighted transitions. Since, in practice, RLMs work in real-time, processing a symbol at every time step, we treat the above result as an upper bound on the expressivity of RLMs. We also provide a lower bound by showing that under the restriction to real-time computation, such models can simulate deterministic real-time rational PTMs.