A Sparse Plus Low-Rank Exponential Language Model for Limited Resource Scenarios

This paper describes a new exponential language model that decomposes the model parameters into one or more low-rank matrices that learn regularities in the training data and one or more sparse matrices that learn exceptions (e.g., keywords). The low-rank matrices induce continuous-space representations of words and histories. The sparse matrices learn multi-word lexical items and topic/domain idiosyncrasies. This model generalizes the standard ℓ1-regularized exponential language model, and has an efficient accelerated first-order training algorithm. Language modeling experiments show that the approach is useful in scenarios with limited training data, including low resource languages and domain adaptation.