The Uli Dataset: An Exercise in Experience Led Annotation of oGBV

Arnav Arora, Maha Jinadoss, Cheshta Arora, Denny George, Brindaalakshmi, Haseena Khan, Kirti Rawat, Div, Ritash, Seema MathurNAACL 2024

TL;DR

This paper introduces the Uli Dataset, a novel collection of annotated tweets in Hindi, Tamil, and Indian English designed to capture online gender-based violence (oGBV) from the perspective of marginalized communities. By employing a participatory approach with activists and researchers, the dataset aims to address the critical lack of contextualized data for building more effective AI systems to combat oGBV, especially in non-English languages.

What Problem Does This Paper Address?

Online gender-based violence (oGBV) is a growing problem, particularly for users in the Global Majority who communicate in languages other than English. Existing AI tools for detecting hate speech and abuse often struggle with the nuances and context of these languages. Furthermore, many datasets are built without adequately considering the lived experiences of those most affected by oGBV. This gap hinders the development of effective automated systems that can truly understand and mitigate this pervasive issue, especially in regions with existing gender inequalities.

Methods

The Uli Dataset was created through a multi-stage process:

  • Semi-supervised Data Collection: A large corpus of tweets was initially collected using a semi-supervised approach. This involved crowdsourcing slurs and offensive phrases, identifying accounts frequently targeted by or perpetuating abuse, and scraping replies to influential women on Twitter. This resulted in approximately 1.3 million tweets.
  • Stratified Pooling: To ensure a higher concentration of relevant content, stratified pooling was employed. This involved using various pre-trained models (IndicBERT, XLM-T, etc.) to assign confidence scores to tweets for hate speech and offensive language detection. Tweets were then binned based on these scores, and a fixed number of posts were sampled from each bin to create a more balanced dataset for annotation.
  • Experience-Led Annotation: The core of the methodology lies in its "experience-led" annotation process. Instead of relying on general annotators, the dataset was annotated by activists and researchers who identify as women or members of the LGBTQIA+ community and have direct experience with or have responded to online abuse. This ensured that the annotations reflected a nuanced understanding of what constitutes gendered abuse.
  • Annotation Guideline Development: The annotation guideline was iteratively developed through literature review and focus group discussions with the annotators. It was initially written in English and then translated into Hindi and Tamil, with examples tailored to mirror the motivation behind the English examples.
  • Multi-label Annotation: Annotators were tasked with answering three questions for each tweet:
  • Is this post gendered abuse when not directed at a person of marginalized gender and sexuality?
  • Is the post gendered abuse when directed at a person of marginalized gender and sexuality?
  • Is this post explicit/aggressive? These questions were designed to capture different facets of oGBV, including its intersectional nature.

Datasets

The paper introduces the Uli Dataset, a new benchmark for online gender-based violence detection. It comprises tweets annotated in three languages: Hindi, Tamil, and Indian English. The dataset was created from approximately 1.3 million scraped tweets and contains around 8000 annotated posts (though the final dataset size is slightly less due to various factors). The dataset is unique in its focus on survivor-centric definitions and expert annotations.

Key Insights and Ideas

  • The Subjectivity of oGBV: The study highlights the inherent subjectivity in defining and annotating oGBV, particularly evident in the varying levels of annotator agreement across languages. English, in particular, showed lower agreement, underscoring the difficulty in capturing nuanced abuse in this language.
  • Importance of Context and Identity: The annotators' lived experiences and identities significantly shaped their interpretations of abuse. The distinction between abuse directed at marginalized genders versus others was crucial, demonstrating that context and the target's identity are paramount in identifying oGBV.
  • Participatory AI for Social Good: The project champions a participatory approach to AI development, emphasizing the inclusion of affected communities in the data creation process. This "experience-led" annotation is presented as a more ethical and effective way to build AI systems for social impact.
  • Limitations of Text-Only Analysis: While the Uli Dataset focuses on text, the focus groups revealed that a significant portion of oGBV occurs through images and videos, pointing to the need for multi-modal approaches in future research.
  • Challenges in Scaling Participatory Design: The labor-intensive nature of this annotation process raises questions about how to balance participatory goals with practical constraints like time and resources, especially when working with multiple languages.

Key Results

  • The Uli Dataset contains approximately 8000 annotated tweets across Hindi, Tamil, and Indian English.
  • The paper reports on model performance using IndicBERT and XLM-T on the dataset. XLM-T generally performed better, which is hypothesized to be due to its familiarity with Twitter data.
  • Tamil showed the highest annotator agreement, while English had the lowest, reflecting the task's subjectivity.
  • The dataset is shared under a CC BY 4.0 license on GitHub, with annotator-level labels provided to preserve the nuances of disagreement.

Limitations and Future Work

  • Text-Only Focus: The current dataset only includes text-based abuse, while a significant portion of oGBV occurs through images and videos. Future work aims to extend to multi-modal content.
  • Labor Intensity: The creation of the dataset was labor-intensive, prompting questions about managing participatory design goals with resource constraints.
  • Annotator Agreement: The varying levels of agreement across languages and labels suggest a need for further investigation into the causes of heterogeneity and for exploring approaches beyond majority voting to address disagreements.
  • Compensation for Annotators: The paper acknowledges the lack of clear guidance on compensating experts for their time and contributions in participatory AI projects, highlighting this as an area for active research.
  • Addressing Disagreements: The authors plan to explore methods that better capture and utilize the disagreements among annotators, rather than simply flattening them into a majority vote.

Summary updated

BibTeX
@inproceedings{arora2024uli,
  author = {Arnav Arora and Maha Jinadoss and Cheshta Arora and Denny George and Brindaalakshmi and Haseena Khan and Kirti Rawat and Div and Ritash and Seema Mathur},
  title = {{The Uli Dataset: An Exercise in Experience Led Annotation of oGBV}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure