Deep linguistic lensing

Abstract

Language models and semantic word embeddings have become ubiquitous as sources for machine learning features in a wide range of predictive tasks and real-world applications. We argue that language models trained on a corpus of text can learn the linguistic biases implicit in that corpus. We discuss linguistic biases, or differences in identity and perspective that account for the variation in language use from one speaker to another. We then describe methods to intentionally capture "linguistic lenses": computational representations of these perspectives. We show how the captured lenses can be used to guide machine learning models during training. We define a number of lenses for author-to-author similarity and word-to-word interchangeability. We demonstrate how lenses can be used during training time to imbue language models with perspectives about writing style, or to create lensed language models that learn less linguistic gender bias than their un-lensed counterparts.