Source-linked AI summary
Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond
Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, Diyi Yang
TL;DR
Causal inference research in NLP lacks unified definitions, benchmarks, and a consolidated account of challenges specific to text. This survey synthesizes the area by covering causal-effect estimation with text and causal formalisms for improving NLP models, concluding that both spaces remain nascent with many open challenges.
Problem
Research on causality in NLP remains scattered and lacks unified definitions, benchmark datasets, and clear accounts of challenges and opportunities for textual data.
Method
The survey consolidates research into estimating causal effects when text is a causal variable and using causal formalisms to improve NLP robustness and interpretability.
Results
The survey provides a unified overview of causal inference for NLP and identifies challenges and opportunities across causal estimation and causal approaches to NLP methods.
Takeaways & Limitations
Causal reasoning can help make NLP models more robust, trustworthy, transparent, and clearer about their assumptions.
Takeaways & Limitations
Text-based causal estimation lacks real-world benchmarks because individual counterfactuals and true causal effects cannot be observed.
Abstract
from arXiv · showhide
A fundamental goal of scientific research is to learn about causal relationships. However, despite its critical role in the life and social sciences, causality has not had the same importance in Natural Language Processing (NLP), which has traditionally placed more emphasis on predictive tasks. This distinction is beginning to fade, with an emerging area of interdisciplinary research at the convergence of causal inference and language processing. Still, research on causality in NLP remains scattered across domains without unified definitions, benchmark datasets and clear articulations of the challenges and opportunities in the application of causal inference to the textual domain, with its unique properties. In this survey, we consolidate research across academic areas and situate it in the broader NLP landscape. We introduce the statistical challenge of estimating causal effects with text, encompassing settings where text is used as an outcome, treatment, or to address confounding. In addition, we explore potential uses of causal inference to improve the robustness, fairness, and interpretability of NLP models. We thus provide a unified overview of causal inference for the NLP community.
1 Introduction
Causal inference is increasingly relevant to NLP as systems enter interdisciplinary and high-stakes settings where prediction alone is insufficient. This survey organizes work on estimating effects from text and using causal ideas to improve robustness and interpretability.
- Motivation: Causal questions arise in interdisciplinary research using NLP, including questions about how interventions affect outcomes.Examples include asking whether a drug changes disease progression or whether signaling gender changes post popularity.
- Motivation: Classical NLP often accepts predictive correlations, but deployed systems face distribution shifts and demands for interpretable predictions.Causal structure can suggest inductive biases for robustness and offer insight into model behavior.
- Causal inference with text: Text-based causal effects can differ from observed correlations because confounders may influence both treatment and outcome.In the forum example, post topic could correlate with both gender-icon choice and the number of likes.
- Causal inference with text: Text can serve as a treatment, outcome, or source of confounding information in causal analyses.Examples include gender signal affecting reply sentiment and writing style affecting received likes.
- Causal inference with text: Causal inference with text is difficult because text is high-dimensional, semantic factors require sophisticated modeling, and interventions must be formalized carefully.NLP representations such as topic models and contextual embeddings may help extract information needed for effect estimation.
- Survey scope: The survey separates estimating causal effects from text and applying causal formalisms to make NLP methods more robust and interpretable.It reviews existing research, challenges, opportunities, and open problems across these two lines of work.
2 Background
Causal inference defines effects through interventions and counterfactual outcomes, then identifies them under assumptions about treatment assignment and observed confounders. The background introduces ATE and CATE, identification assumptions, and causal DAGs for encoding and checking those assumptions.
- Counterfactuals: Counterfactuals specify outcomes under interventions, but no unit reveals its treated and untreated potential outcomes simultaneously.This fundamental problem makes causal inference impossible without identification assumptions.
- 2.1 Causal Estimands: ATE measures the average difference between potential outcomes, while CATE measures average effects within a predefined subgroup.CATEs quantify heterogeneity across population subgroups, such as posts on political topics.
- 2.2 Identification Assumptions for Causal Inference: Identification of the ATE relies on consistency, conditional ignorability, and positivity, allowing the final expression to use observable quantities.The assumptions ensure well-defined potential outcomes, independence after conditioning, and treatment availability across confounder values.
- 2.2 Identification Assumptions for Causal Inference: Ignorability can fail when confounders affect both treatment assignment and potential outcomes, unless those confounders are observed and conditioned on.Randomization guarantees ignorability by design, whereas author experience can confound the forum-icon example.
- 2.2 Identification Assumptions for Causal Inference: Consistency can be violated by unrecorded treatment versions or interference between units.Different icon colors could have different effects, and one unit’s treatment could affect another unit’s potential outcomes.
- 2.2 Identification Assumptions for Causal Inference: Causal DAGs encode assumptions and support d-separation analyses for selecting variables that can make conditional ignorability plausible.In the forum example, the treatment and outcome become independent only after conditioning on appropriate variables.
3 Estimating Causal Effects with Text
Estimating causal effects with text requires adapting causal methods to automatically extracted, high-dimensional representations and carefully checking their assumptions. The survey highlights challenges involving confounding, heterogeneous effects, measurement, interventions, consistency, positivity, and evaluation benchmarks.
- Core challenges: Text-based causal analyses require assumptions about automatically extracted variables, making valid causal conclusions a joint NLP and causal-estimation challenge.The paper calls for collaboration to identify requisite assumptions and address settings where text is a confounder, outcome, or treatment.
- Textual confounders: NLP methods can extract textual confounders for adjustment, but their violations of causal assumptions remain open questions.Prior approaches use dimensionality reduction, topic models, embeddings, auto-encoders, and supervised language models to infer confounding information.
- Causal assumptions: Conditional ignorability requires measuring all confounding aspects in text, an untestable requirement that should be justified with domain expertise.The assumption is especially strong when treatment assignment is not randomized.
- Causal assumptions: High-dimensional text makes positivity difficult when representations nearly encode treatment choices, limiting plausible changes while holding textual confounders fixed.The paper illustrates this with gendered posts whose wording may nearly reveal the author’s chosen gender icon.
- Causal assumptions: Automatically learned outcomes can violate consistency when models trained across units depend on treatment assignments affecting other units’ text.In the suspension example, different randomized assignments could produce different inferred outcomes and undermine the causal estimate.
- Future work: Future work includes modeling heterogeneous effects, learning causal representations, generating controlled texts, and building benchmarks that test generalization beyond synthetic data.The survey notes that no real-world text-based causal estimation benchmarks currently exist, while semi-synthetic and controlled synthetic datasets provide partial evaluation.
4 Robust and Explainable Predictions from Causality
This section presents causal approaches for making NLP predictions more robust, fair, and interpretable by addressing spurious correlations and modeling counterfactual invariance. It surveys counterfactual augmentation, distributional criteria, and open challenges in specifying causal structure from text.
- Motivation: Correlational NLP predictors can fail under distribution shift, perform differently across user groups, and remain difficult to interpret in high-stakes settings.These failures motivate causal criteria that go beyond predictive correlations.
- Spurious correlations: Spurious correlations arise when training features predict a confounder whose relationship with the label changes across deployment domains.In the hospital example, writing style reflects hospital identity even when it is not diagnostically useful within a hospital.
- Evaluation: Invariance tests evaluate whether predictions remain stable under causally irrelevant perturbations, while sensitivity tests seek minimal changes that flip predictions.These evaluations operationalize the goal of being “right for the right reasons.”
- Data augmentation: Counterfactual data augmentation adds manually edited, keyword-replaced, or automatically rewritten examples to train predictors against confounding and label perturbations.Confounder perturbations can penalize prediction disagreement, while label counterfactuals can improve out-of-distribution generalization and reduce noise sensitivity.
- Data augmentation: Counterfactual augmentation is limited by the difficulty of producing meaningful interventions, specifying their intended scope, and avoiding newly introduced spurious correlations.These problems make resulting robustness guarantees difficult to ascertain.
- Distributional criteria: Counterfactual invariance can be translated into observed-data independence conditions and incorporated into training as regularization informed by the causal structure.For hospital-style confounding, the desired predictor satisfies f(X) ⊥⊥ Z | Y; other causal structures can imply f(X) ⊥⊥ Z.
- Open challenges: Domain generalization may require data from many environments to disentangle the true causal structure, and current methods largely use top-down rather than bottom-up causal models of text.Hybrid models connecting document metadata with sentence- or paragraph-level spans remain a future direction.
- Open challenges: When the target variable is unobserved and only a noisy or coarsened proxy is available, causal invariance remains an emerging and largely unexplored NLP problem.Dialect invariance using geography as an approximate proxy illustrates this limitation.
5 Conclusion
The survey unifies research connecting causal inference and NLP across causal-effect estimation and more reliable language technologies. It concludes that both directions are nascent, with many open challenges, while causal methodology can make assumptions explicit and improve understanding of language and NLP models.
- Scope: The survey organizes causality and NLP into estimating causal effects and applying causal formalisms to traditional NLP tasks.It brings these scattered touchpoints into one framework.
- Contributions: NLP modeling can support causal conclusions from text, while causal ideas can make NLP models more robust, trustworthy, and transparent.The survey also relates implicitly causal approaches to causal inference.
- Implications: Causal methodology forces practitioners to state assumptions explicitly, which may improve scientific standards and understanding of language and the models built to process it.The conclusion frames explicit assumptions as a particular advantage of causal methods.