On the Gap between Adoption and Understanding in NLP

By Federico Bianchi, Dirk Hovy. 2021

25 citations recorded by the source corpus.

There are some issues with current research trends in NLP that can hamper the free development of scientiﬁc research. We identify ﬁve of particular concern: 1) the early adoption of methods without sufﬁcient understanding or analysis; 2) the preference for computational methods regardless of risks associated with their limitations; 3) the resulting bias in the papers we publish; 4) the impossibility of re-running some experiments due to their cost; 5) the dangers of unexplainable methods. If these issues are not addressed, we risk a loss of reproducibility, reputability, and subsequently public trust in our ﬁeld. In this position paper, we outline each of these points and suggest ways forward.

Subfields: Artificial Intelligence, Natural Language Processing, Evaluation & Benchmarks

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 Evaluation & Benchmarks

Citations in the corpus

Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling (2021, background): “This is important, as a general audience may have different expectations and a different understanding of the limitations of systems (Bianchi & Hovy, 2021).”
Guiding the Release of Safer E2E Conversational AI through Value Sensitive Design (2022, neutral): “derstanding of the limitations of systems (Bianchi and Hovy, 2021).”
Taxonomy of Risks posed by Language Models (2022, background): “Better understanding, interpretation and explanation of LMs are also essential for unlocking mitigations to address risks of harm [20].”
Pipelines for Social Bias Testing of Large Language Models (2022, background): “However, these models are quickly adopted without complete understanding their possible limitations (Bianchi and Hovy, 2021).”
Measuring Harmful Sentence Completion in Language Models for LGBTQIA+ Individuals (2022, background): “…Hovy, 2021), such as stereotypically biased and polarized embeddings (e.g., Bolukbasi et al., 2016; Gonen and Goldberg, 2019; Lauscher et al., 2020; Bianchi et al., 2021), and unfairly biased task-specific models, e.g., for natural language inference (e.g., Dev et al., 2020), natural language…”
Five sources of bias in natural language processing (2021, background): “Finally, not analysing the behaviour of models sufficiently, or not fully disclosing it can be harmful (Bianchi & Hovy, 2021).”