Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Omitted Variable Bias in Language Models Under Distribution Shift

Victoria Lin, Louis-Philippe Morency, Eli Ben-Michael

TL;DR

This paper introduces a framework that maps the strength of the omitted variables to bounds on the worst-case generalization performance of language models under distribution shift and shows that using these bounds directly in language model evaluation and optimization provides more principled measures of out-of-distribution performance.

Abstract

Despite their impressive performance on a wide variety of tasks, modern language models remain susceptible to distribution shifts, exhibiting brittle behavior when evaluated on data that differs in distribution from their training data. In this paper, we describe how distribution shifts in language models can be separated into observable and unobservable components, and we discuss how established approaches for dealing with distribution shift address only the former. Importantly, we identify that the resulting omitted variable bias from unobserved variables can compromise both evaluation and optimization in language models. To address this challenge, we introduce a framework that maps the strength of the omitted variables to bounds on the worst-case generalization performance of language models under distribution shift. In empirical experiments, we show that using these bounds directly in language model evaluation and optimization provides more principled measures of out-of-distribution performance, improves true out-of-distribution performance relative to standard distribution shift adjustment methods, and further enables inference about the strength of the omitted variables when target distribution labels are available.

Omitted Variable Bias in Language Models Under Distribution Shift

TL;DR

This paper introduces a framework that maps the strength of the omitted variables to bounds on the worst-case generalization performance of language models under distribution shift and shows that using these bounds directly in language model evaluation and optimization provides more principled measures of out-of-distribution performance.

Abstract

Despite their impressive performance on a wide variety of tasks, modern language models remain susceptible to distribution shifts, exhibiting brittle behavior when evaluated on data that differs in distribution from their training data. In this paper, we describe how distribution shifts in language models can be separated into observable and unobservable components, and we discuss how established approaches for dealing with distribution shift address only the former. Importantly, we identify that the resulting omitted variable bias from unobserved variables can compromise both evaluation and optimization in language models. To address this challenge, we introduce a framework that maps the strength of the omitted variables to bounds on the worst-case generalization performance of language models under distribution shift. In empirical experiments, we show that using these bounds directly in language model evaluation and optimization provides more principled measures of out-of-distribution performance, improves true out-of-distribution performance relative to standard distribution shift adjustment methods, and further enables inference about the strength of the omitted variables when target distribution labels are available.
Paper Structure (45 sections, 33 equations, 6 figures, 2 tables)

This paper contains 45 sections, 33 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Domain adaptation and distribution shift.
  4. Omitted variable bias.
  5. Proposed Approach
  6. A General Form for Doubly Robust Losses
  7. Identification.
  8. Estimation.
  9. Worst-Case Generalization Performance
  10. OVB definition.
  11. Worst-case objective.
  12. GLM Log-Likelihood Losses
  13. A Worst-Case Objective for GLM Losses
  14. Task-specific parameters.
  15. From Long to Short
  16. ...and 30 more sections

Figures (6)

  • Figure 1: Evaluations of true and distribution shift-adjusted performance on MATH-Perturb.
  • Figure 2: Strength of omitted variables under distribution shift in the Math Reasoning dataset.
  • Figure 3: Amazon dataset. On the left, we plot the test performance difference (higher is better) between our approach and a model trained with a standard DR objective against the true strength of omitted variables. On the right, we plot the test performance of our approach and both baselines.
  • Figure 4: Test performance difference (higher is better) between our approach and a model trained with a standard DR objective against assumed strength of omitted variables.
  • Figure 5: OVB metrics under distribution shift in the Math Reasoning dataset.
  • ...and 1 more figures