The State of Profanity Obfuscation in Natural Language Processing Scientific Publications

Debora Nozza, Dirk HovyACL 2023 · DOI

Introduction: The Researcher's Dilemma

NLP researchers studying hate speech face a "damned if you do, damned if you don't" dilemma: include raw profanity and risk harming readers, or hide it behind asterisks and sacrifice scientific clarity. This paper surveys the current "wild west" of profanity obfuscation in ACL publications and introduces PROF, a multilingual Python tool and resource designed to standardize how we mask offensive language without losing the reader.

The Problem: Inconsistency and Accessibility

In the field of Natural Language Processing (NLP), researchers must often report on "unsavory" subjects like hate speech and online abuse. To do this scientifically, they need to show examples. However, there is no consensus on how to present these examples.

Some authors leave profanities in the clear, which can be distressing for readers or inadvertently help hate speech proliferate online. Others use "obfuscation"—replacing letters with symbols (e.g., f*ck). The problem is that this is currently done inconsistently. One paper might use f*ck, another f***, and another ****.

Examples of inconsistent profanity obfuscation found in 2021 NLP literature.

This creates a major accessibility barrier, particularly for non-native English speakers who may not have the cultural context to "fill in the blanks" for slang or slurs they weren't taught in school. If a researcher can't decipher an example, they can't fully evaluate the science.

Methodology and the PROF Tool

The authors approached this problem in two stages:

  1. The Survey: They analyzed 150 papers from major 2021 NLP conferences (ACL, EACL, EMNLP, and the WOAH workshop) specifically tagged for offensive language or hate speech.
  2. The Tool (PROF): To solve the inconsistency, they built PROF (Profanity ObFuscation). This is a multilingual resource that maps offensive terms to their semantic concepts in BabelNet.

The PROF resource contains a dictionary of 203 profanities across five languages: English (50), French (44), German (19), Italian (42), and Spanish (48). These were validated by native speakers to ensure cross-lingual consistency. The accompanying Python module allows researchers to automatically obfuscate text using a standardized "first-vowel" rule (e.g., f*ck instead of f***) and even "reveal" the original word if a reader needs to know the exact term for replication purposes.

Key Insights and Results

The survey revealed several critical trends in the field:

  • Low Disclaimer Adoption: Less than 20% of the surveyed NLP papers included a disclaimer or content warning about offensive language.
  • The Asterisk Rules: The * symbol is the overwhelming favorite for masking, but the pattern of masking varies wildly.
  • Standardization: The authors propose a "first-vowel" obfuscation standard. By only masking the first vowel, the word remains recognizable to researchers while still providing a "buffer" against the raw offensive term.
  • Common Targets: The survey identified the most frequently obfuscated terms across the 2021 proceedings, with terms like f*ck and n*gga appearing most often.
Frequency of specific obfuscated words across the surveyed 2021 NLP publications.

The authors also noted a significant English-Centricity: Profanities in languages other than English are much less likely to be obfuscated, suggesting a bias in how "harm" is perceived across different cultures.

Limitations and Future Work

The primary limitation is language coverage. While PROF covers five major Western languages, it currently lacks data for many other languages where hate speech research is active.

Additionally, the tool relies on a static dictionary. Language—especially slang and hate speech—evolves rapidly. New slurs or creative "leetspeak" workarounds might bypass a simple dictionary-based obfuscator. The authors hope to evolve PROF into a community-driven resource where researchers can contribute new terms and mappings as they encounter them in the wild.

Summary updated

BibTeX
@inproceedings{nozza2023state,
  author = {Debora Nozza and Dirk Hovy},
  title = {{The State of Profanity Obfuscation in Natural Language Processing Scientific Publications}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.findings-acl.240},
}
Figure