Distributionally Robust Language Modeling
This addresses the issue of distribution shift in language modeling for applications like targeted text generation, though it is incremental as it builds on existing DRO techniques.
The paper tackles the problem of language models performing poorly on unknown target distributions by proposing a distributionally robust optimization method, which achieves a 5.5 point perplexity reduction over standard training when tested on reviews after training on a mix of reviews and news.
Language models are generally trained on data spanning a wide range of topics (e.g., news, reviews, fiction), but they might be applied to an a priori unknown target distribution (e.g., restaurant reviews). In this paper, we first show that training on text outside the test distribution can degrade test performance when using standard maximum likelihood (MLE) training. To remedy this without the knowledge of the test distribution, we propose an approach which trains a model that performs well over a wide range of potential test distributions. In particular, we derive a new distributionally robust optimization (DRO) procedure which minimizes the loss of the model over the worst-case mixture of topics with sufficient overlap with the training distribution. Our approach, called topic conditional value at risk (topic CVaR), obtains a 5.5 point perplexity reduction over MLE when the language models are trained on a mixture of Yelp reviews and news and tested only on reviews.