Creating a framework for a Benchmark Religion Dataset Deepa Muralidhar1 , Ashwin Ashok2 1 Georgia State University 2 Georgia State University Abstract Development of Language Models (LM) such as OpenAI’s GPT series generating natural language text is growing at a rapid pace. The LMs take a text prompt as input and generate text as output that represent the most probable sequence of words matching the prompt’s context and pattern. Our preliminary investigations revealed bias but not much evidence for its cause. Our goal is, therefore, to build a benchmark dataset on various religions for evaluating this bias. We envision that our conceptual method of creating a dataset and developing a bias rating mechanism can serve as a fundamental tool establishing a process to measure bias. Comparing the Bias Indicator Value (BIV) for one religion against another should give us enough information to provide a holistic bias rating for the text generated. Keywords Large Language Models, religious bias, algorithmic bias, bias, metrics, Text mining, socioeconomic factors, mitigating, 1. Introduction and Motivation why LMs generate texts differently for different religions. This metric which acts as a indicator of religious bias Our research on religious bias targets two key challenges in GPT-3 is computed for the generated data. We share for AI text generators: (1) The need for a religion-based our results of these experiments with the community benchmark data set to evaluate an AI text generator for through a preliminary religion dataset that includes tex- bias. The problem with existing data sets is that there are tual prompts, the test data, the associated AI-generated no clear documentation and other data set management output text and the analysis done of the text. The metrics practices in place [1]. A well-designed benchmark data and graphs calculated are part of this benchmark dataset. set helps verify that the output data is unbiased across a diverse distribution of real-world contexts [2]. (2) To present indicators of bias,a quantitative value that can 3. Open Challenges represent the implicit bias numerically. Our key obser- vation is that it is challenging to create a bias metric in An open research question for future work is that one LMs as the bias changes depending on the context of the metric maybe insufficient to measure biases, instead look text. to develop a bias-reporting toolkit. This could include transparency measurement, an examination on the how and why of the decision-making process in an AI system, 2. Target design goals is useful in detecting systemic biases.[4] We conduct experiments to measure the sentiment of the AI generated text and test for religious bias. For every References prompt, we program GPT-3 to generate 200 tokens, ap- proximately 178 words. For six sets of 20 prompts (five [1] K. Peng, A. Mathur, A. Narayanan, Mitigating religions and one religion-neutral that acts as control dataset harms requires stewardship: Lessons from value) the text has about 11100 words (12000 tokens). 1000 papers (2021). URL: https://arxiv.org/abs/2108. Using VADER [3], a rule-based tool that measures the 02922. sentiment within the text, we measure positive and nega- [2] P. P. Liang, C. Wu, L. Morency, R. Salakhutdinov, tive values (between -1 to +1) and compare the sentiment Towards understanding and mitigating social biases value of each sentence. We use this to create a prelimi- in language models, 2021. URL: https://arxiv.org/ nary quantifiable metric,a Bias Indicator Value, to identify abs/2106.13219. stereotypical bias with respect to a religion and interpret [3] G. E. Hutto, C.J., Vader: A parsimonious rule-based model for sentiment analysis of social media text, 2015. Human in the Loop Data Curation Workshop’22: CKIM ’22, October 17–21, 2022, Atlanta, Ga [4] J. Stoyanovich, Transfat: Translating fairness, ac- Envelope-Open deepa.muralidhar@gmail.com (D. Muralidhar); aashok@gsu.edu countability and transparency into data science prac- (A. Ashok) tice, 2018. URL: http://ceur-ws.org/Vol-2417/paper1. © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). pdf. CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)