Izzidien A, Fitz S, Romero P, Loe BS, Stillwell D. Developing a sentence level fairness metric using word embeddings.
INTERNATIONAL JOURNAL OF DIGITAL HUMANITIES 2022;
5:1-36. [PMID:
36249081 PMCID:
PMC9549858 DOI:
10.1007/s42803-022-00049-4]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [Grants] [Track Full Text] [Download PDF] [Figures] [Subscribe] [Scholar Register] [Received: 12/21/2021] [Accepted: 09/04/2022] [Indexed: 11/06/2022]
Abstract
Fairness is a principal social value that is observable in civilisations around the world. Yet, a fairness metric for digital texts that describe even a simple social interaction, e.g., 'The boy hurt the girl' has not been developed. We address this by employing word embeddings that use factors found in a new social psychology literature review on the topic. We use these factors to build fairness vectors. These vectors are used as sentence level measures, whereby each dimension reflects a fairness component. The approach is employed to approximate human perceptions of fairness. The method leverages a pro-social bias within word embeddings, for which we obtain an F1 = 79.8 on a list of sentences using the Universal Sentence Encoder (USE). A second approach, using principal component analysis (PCA) and machine learning (ML), produces an F1 = 86.2. Repeating these tests using Sentence Bidirectional Encoder Representations from Transformers (SBERT) produces an F1 = 96.9 and F1 = 100 respectively. Improvements using subspace representations are further suggested. By proposing a first-principles approach, the paper contributes to the analysis of digital texts along an ethical dimension.
Collapse