Learning to Suggest Phrases

Intelligent keyboards can support writing by suggesting content. Certain types of phrases, when offered as suggestions, may be systematically chosen more often than their frequency in a corpus of text would predict. In order to generate those types of suggestions, we collected a dataset of how human authors responded to suggestions offered to them during open-ended writing tasks. We present an offline strategy for evaluating suggestions that enables us to learn the parameters of an improved suggestion generation policy without the expense of collecting additional data under that policy. We validate the approach by simulation and on human data by demonstrating improvement in held-out suggestion acceptance rate. Our approach can be applied to other scenarios where what is typical is not necessarily what is desirable.