BiasFinder: Metamorphic Test Generation to Uncover Bias for Sentiment Analysis Systems

Artificial Intelligence (AI) software systems, such as Sentiment Analysis (SA) typically learn from large amounts of data that may reflect human biases. Consequently, the machine learning model in systems exhibit unintended demographic bias based on specific characteristics (e.g., gender, occupation, country-of-origin, etc.). Such biases manifest an SA system when it predicts a different sentiment for similar texts differ only characteristic individuals described. Existing studies revealing rely production sentences small set short, predefined templates. To address this limitation, we present BisaFinder, approach to discover biased predictions via metamorphic testing. A key feature BisaFinder is automatic curation suitable templates pieces text corpus, using various Natural Language Processing (NLP) techniques identify words describe characteristics. Next, instantiates new these by filling placeholders with associated class gender-specific female names, she, her). These are used tease out system. identifies bias-uncovering test case detects exhibits pair texts, i.e., male vs. female) target gender). Our empirical evaluation showed can effectively create number realistic and diverse cases uncover high true positive rate up 95.8\%.