Distributionally Robust Language Modeling

By Yonatan Oren, Shiori Sagawa, Tatsunori B. Hashimoto, Percy Liang. 2019

210 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

Detoxifying Language Models Risks Marginalizing Minority Voices (2021, background): “As an alternative or in addition to higher quality data, there is growing interest in training accurate models in the presence of biased data (Oren et al., 2019; Clark et al., 2019).”
Measuring Robustness to Natural Distribution Shifts in Image Classification (2020, methodology): “[62] Oren, Y.”
Mind the Gap: Assessing Temporal Generalization in Neural Language Models (2021, background): “…et al., 2006; Daumé III, 2007; Axelrod et al., 2011), and has recently been addressed for neural LMs and transfer learning (Fried et al., 2019; Oren et al., 2019; Gururangan et al., 2020), where pretrained LMs lead to substantial improvements and increased robustness (Hendrycks et al., 2020).”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “Another potential approach is using baselining to account for different groups having different intrinsic levels of difficulty (Oren et al., 2019).”
DEMix Layers: Disentangling Domains for Modular Language Modeling (2021, background): “Even if training data is sourced from many domains, dense training can in practice emphasize subsets of the data in proportion to their ease of access (Oren et al., 2019; Fan et al., 2020), limiting generalization to less prevalent domains.”
An Empirical Study on Robustness to Spurious Correlations using Pre-trained Language Models (2020, background): “, 2019), and group shift (Oren et al., 2019; Sagawa et al., 2020) in the literature.”
Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization (2019, methodology): “Sinha et al. (2018) study neural networks but with a more conservative Wasserstein uncertainty set that leads to non-vanishing training loss; and Oren et al. (2019) study neural networks but for generative modeling where loss tradeoffs arise naturally.”
Turning Tables: Generating Examples from Semi-structured Tables for Endowing Language Models with Reasoning Skills (2021, background): “and distributionally robust optimization (Oren et al., 2019; Sagawa et al., 2020), where the goal is to perform well over a family of distributions.”
Utility Is in the Eye of the User: A Critique of NLP Leaderboard Design (2020, background): “The perplexity of language models rises when given out-of-domain text (Oren et al., 2019).”
An Investigation of Why Overparameterization Exacerbates Spurious Correlations (2020, background): “To mitigate this problem, common approaches reduce the worst-group training loss, e.g., through distributionally robust optimization (DRO) or simply upweighting the minority groups.”
Distributionally Robust Neural Networks (2020, methodology): “Extending existing minimax algorithms for DRO (Namkoong & Duchi, 2016; Oren et al., 2019), we interleave gradient-based updates on θ and q.”
Robustness to Spurious Correlations via Human Annotations (2020, background): “…as subpopulations with size at least α ∗ , Prior work has shown that such shifts over groups can be used to capture a wide range of test time distributions including label shifts (Hu et al., 2018), topics within a corpus (Oren et al., 2019) , and demographic groups (Hashimoto et al., 2018).”
Just Train Twice: Improving Group Robustness without Training Group Information (2021, methodology): “Many approaches for this setting fall under the DRO framework, where models are trained to minimize the worst-case loss across all distributions in a ball around the empirical distribution (Ben-Tal et al., 2013; Lam & Zhou, 2015; Duchi et al., 2016; Namkoong & Duchi, 2017; Oren et al., 2019).”
BREEDS: Benchmarks for Subpopulation Shift (2020, unknown): “[Ore+19] Yonatan Oren et al. “Distributionally Robust Language Modeling”. In: Empirical Methods in Natural Language Processing (EMNLP). 2019. [Pen+19] Xingchao Peng et al. “Moment matching for multi-source domain adaptation”. In: International Conference on Computer Vision (ICCV). 2019. [Pon+06] Jea”
DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining (2023, methodology): “Within DRO methods for deep learning [4; 47; 35; 43], we target a restricted form of shift called group shifts [14; 35; 43], where the test distribution can be an unknown mixture of groups (domains).”
Neural Unsupervised Domain Adaptation in NLP—A Survey (2020, background): “A recent solution is distributionally robust optimization (Oren et al., 2019), i.e., optimizing for worst-case performance without the knowledge of the test distribution.”
DORO: Distributional and Outlier Robust Optimization (2021, unknown): “Several recent papers (Hashimoto et al., 2018; Oren et al., 2019; Xu et al., 2020) proposed using DRO to deal with subpopulation shift, but it was also observed that DRO was prone to overﬁt in practice (Sagawa et al., 2020a;b).”
Challenges and Strategies in Cross-Cultural NLP (2022, methodology): “(2021) propose using Group Distributionally Robust Optimisation (group DRO; Oren et al., 2019) to optimise worst-case performance across languages.”
Adaptive Risk Minimization: A Meta-Learning Approach for Tackling Group Shift (2020, background): “This setting of structural DRO allows for more tractable optimization while still permitting a wide range of realistic distribution shifts [63, 31, 50, 55].”