Abstract
Numeracy is the ability to understand and work with numbers. It is anecessary skill for composing and understanding documents in clinical,scientific, and other technical domains. In this paper, we explore differentstrategies for modelling numerals with language models, such as memorisationand digit-by-digit composition, and propose a novel neural architecture thatuses a continuous probability density function to model numerals from an openvocabulary. Our evaluation on clinical and scientific datasets shows that usinghierarchical models to distinguish numerals from words improves a perplexitymetric on the subset of numerals by 2 and 4 orders of magnitude, respectively,over non-hierarchical models. A combination of strategies can further improveperplexity. Our continuous probability density function model reduces meanabsolute percentage errors by 18% and 54% in comparison to the second beststrategy for each dataset, respectively.