Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

D3CODE: Disentangling Disagreements in Data across Cultures on Offensiveness Detection and Evaluation

Aida Mostafazadeh Davani, Mark Díaz, Dylan Baker, Vinodkumar Prabhakaran

TL;DR

Offensiveness judgments are culturally and individually subjective, and prior NLP datasets underrepresent this diversity. D3CODE aggregates cross-cultural annotations from 4,309 participants across 21 countries and eight geo-cultural regions, enriching labels with MFQ-2 moral foundations data and drawing items from Jigsaw datasets to maximize disagreement. The study shows strong regional and moral-value influences on labeling and highlights higher disagreement for content mentioning social groups, offering pathways for annotator- and culture-aware NLP modeling and evaluation. This resource enables more pluralistic, fair, and context-sensitive language technologies and RLHF practices that respect diverse moral norms across cultures.

Abstract

While human annotations play a crucial role in language technologies, annotator subjectivity has long been overlooked in data collection. Recent studies that have critically examined this issue are often situated in the Western context, and solely document differences across age, gender, or racial groups. As a result, NLP research on subjectivity have overlooked the fact that individuals within demographic groups may hold diverse values, which can influence their perceptions beyond their group norms. To effectively incorporate these considerations into NLP pipelines, we need datasets with extensive parallel annotations from various social and cultural groups. In this paper we introduce the \dataset dataset: a large-scale cross-cultural dataset of parallel annotations for offensive language in over 4.5K sentences annotated by a pool of over 4k annotators, balanced across gender and age, from across 21 countries, representing eight geo-cultural regions. The dataset contains annotators' moral values captured along six moral foundations: care, equality, proportionality, authority, loyalty, and purity. Our analyses reveal substantial regional variations in annotators' perceptions that are shaped by individual moral values, offering crucial insights for building pluralistic, culturally sensitive NLP models.

D3CODE: Disentangling Disagreements in Data across Cultures on Offensiveness Detection and Evaluation

TL;DR

Offensiveness judgments are culturally and individually subjective, and prior NLP datasets underrepresent this diversity. D3CODE aggregates cross-cultural annotations from 4,309 participants across 21 countries and eight geo-cultural regions, enriching labels with MFQ-2 moral foundations data and drawing items from Jigsaw datasets to maximize disagreement. The study shows strong regional and moral-value influences on labeling and highlights higher disagreement for content mentioning social groups, offering pathways for annotator- and culture-aware NLP modeling and evaluation. This resource enables more pluralistic, fair, and context-sensitive language technologies and RLHF practices that respect diverse moral norms across cultures.

Abstract

While human annotations play a crucial role in language technologies, annotator subjectivity has long been overlooked in data collection. Recent studies that have critically examined this issue are often situated in the Western context, and solely document differences across age, gender, or racial groups. As a result, NLP research on subjectivity have overlooked the fact that individuals within demographic groups may hold diverse values, which can influence their perceptions beyond their group norms. To effectively incorporate these considerations into NLP pipelines, we need datasets with extensive parallel annotations from various social and cultural groups. In this paper we introduce the \dataset dataset: a large-scale cross-cultural dataset of parallel annotations for offensive language in over 4.5K sentences annotated by a pool of over 4k annotators, balanced across gender and age, from across 21 countries, representing eight geo-cultural regions. The dataset contains annotators' moral values captured along six moral foundations: care, equality, proportionality, authority, loyalty, and purity. Our analyses reveal substantial regional variations in annotators' perceptions that are shaped by individual moral values, offering crucial insights for building pluralistic, culturally sensitive NLP models.
Paper Structure (28 sections, 7 figures, 5 tables)

This paper contains 28 sections, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Disagreement-aware Modeling
  4. Annotator-aware Modeling
  5. Annotators in NLP Datasets
  6. D3CODE Dataset
  7. Recruitment
  8. Annotation items
  9. Random:
  10. Moral Sentiment:
  11. Social Group Mentions:
  12. Annotation task
  13. Moral Foundation Questionnaire
  14. Analyses
  15. Analysis of Lack Understanding
  16. ...and 13 more sections

Figures (7)

  • Figure 1: The distribution of labels provided from different countries. Annotators from China, Brazil, and Egypt provided significantly different labels.
  • Figure 2: The likelihood of an annotator not understanding the message, grouped based on their socio-demographic information. Annotators identifying as Men, or of 50 years of old or younger are generally less likely to state they did not understand a message.
  • Figure 3: Disagreement between regions on items from each category (a) and each sub-category (b). We considered the standard deviation of majority votes from different regions as the cross-regional disagreement. The plot shows that items related to social groups (christian, transgender, jewish, muslim and LGB) generally evoke more disagreement compared to random items.
  • Figure 4: Sample of MFQ-2 questions in our survey
  • Figure 5: Distribution of the different labels provided by annotators of different countries. The y-axis is sorted based on the average offensive label captured in each country.
  • ...and 2 more figures