Pseudo-OOD training for robust language models
This addresses the need for robust machine learning models in industry-scale NLP applications by improving OOD detection without requiring access to OOD samples during training.
The paper tackles the problem of unreliable predictions on out-of-distribution (OOD) inputs in pre-trained language models by proposing POORE, a post hoc framework that generates pseudo-OOD samples from in-distribution data and fine-tunes models with a regularization loss, achieving state-of-the-art results in OOD detection on three real-world dialogue systems.
While pre-trained large-scale deep models have garnered attention as an important topic for many downstream natural language processing (NLP) tasks, such models often make unreliable predictions on out-of-distribution (OOD) inputs. As such, OOD detection is a key component of a reliable machine-learning model for any industry-scale application. Common approaches often assume access to additional OOD samples during the training stage, however, outlier distribution is often unknown in advance. Instead, we propose a post hoc framework called POORE - POsthoc pseudo-Ood REgularization, that generates pseudo-OOD samples using in-distribution (IND) data. The model is fine-tuned by introducing a new regularization loss that separates the embeddings of IND and OOD data, which leads to significant gains on the OOD prediction task during testing. We extensively evaluate our framework on three real-world dialogue systems, achieving new state-of-the-art in OOD detection.