On the Gap between Adoption and Understanding in NLP

Federico Bianchi, Dirk HovyACL 2021 · DOI

Introduction

The field of Natural Language Processing (NLP) is currently experiencing a period of unprecedented growth and deployment. However, this rapid advancement has birthed a significant "Gap between Adoption and Understanding" (GAU). While models like BERT and GPT-3 are integrated into global infrastructure, our scientific understanding of their internal mechanics, failure modes, and societal impacts lags dangerously behind. This position paper argues that the current "move fast and break things" culture—driven by publication bias and high computational costs—threatens the scientific integrity, equity, and public trust of the field.

The Five Pillars of the GAU

The authors identify five systemic issues that contribute to the widening gap between how we use models and how we understand them:

  1. Early Adoption: Models are frequently deployed in production environments before the community fully understands their failure modes, such as how they handle negation or latent social biases.
  2. Method-Driven Research: There is a growing preference for "what works" (pure empiricism) over "why it works" (linguistic theory). This shift prioritizes leaderboard performance over theoretical grounding.
  3. Publication Bias and the "Winner’s Curse": The academic incentive structure favors state-of-the-art (SOTA) results. This leads researchers to bury negative findings, creating a "false supremacy" where faulty methods remain the standard because the evidence against them is never published.
  4. Computational Inequality: The prohibitive cost of training and fine-tuning large-scale models has created an "unbridgeable gap" between resource-rich industry labs and the rest of academia. This makes independent verification of results impossible for most researchers.
  5. Explainability Deficits: Models that appear intelligent often make unexplainable decisions. This is particularly hazardous in sensitive domains like medicine or law, where the reasoning behind an output is as important as the output itself.

The Reproducibility Crisis

The paper highlights a stark reality regarding the current state of NLP research. A cited study by Belz et al. (2021) found that only 15.61% of 506 NLP papers checked were independently reproducible. This crisis is exacerbated by the "computational unobtainability" of modern models; the average academic lab cannot afford the dozens of training runs necessary for rigorous tuning and verification.

Furthermore, the authors argue that the quality of code is a scientific contribution. If an experiment cannot be replicated due to undocumented Jupyter notebooks or specific hardware requirements, its scientific value is significantly diminished. Writing "bad code" should be viewed as equivalent to writing a bad paper.

Proposed Solutions: A Path Toward "Slow Science"

To bridge the GAU, the authors propose several structural shifts in how NLP research is conducted and evaluated:

  • Code Statements: Researchers should move beyond sharing simple GitHub links to providing full environment specifications (e.g., CUDA versions) and documented dependencies.
  • Workshops for Negative Results: The community must encourage the publication of "failed" experiments to prevent researchers from stagnating on faulty SOTA methods.
  • Pre-registration: Adopting practices from social psychology, such as pre-registering hypotheses, can prevent "fishing expeditions" for interesting results and ensure more rigorous methodology.
  • Public Engagement: Researchers must better manage public expectations, explaining that "intelligent-looking" outputs from Large Language Models (LLMs) do not equate to human-level reasoning or understanding.

Conclusion

The responsibility for closing the Gap between Adoption and Understanding lies with senior researchers, reviewers, and conference organizers. By shifting the incentive structure away from leaderboard scores and toward scientific depth and computational equity, the field can ensure that its progress is both sustainable and scientifically sound.

Summary updated

BibTeX
@inproceedings{bianchi2021gap,
  author = {Federico Bianchi and Dirk Hovy},
  title = {{On the Gap between Adoption and Understanding in NLP}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.findings-acl.340},
}
Figure