Conceptualizing Social Harms Arising from Bias and Discrimination in Natural Language Processing : Race, Gender & Language

Natural language processing (NLP) is a subfield of artificial intelligence (AI) and has become increasingly prominent in our everyday lives. NLP systems are now ubiquitous as they are capable of identifying offensive and abusive conversational content and hate speech detection on social media platforms, voice and speech recognition and transcription, news recommendation, dialogue systems and digital assistants, language generation, etc. Yet, the benefits of these language technologies do not accrue evenly to all of its users leading to harmful social impacts as NLP systems reproduce stereotypes or fallacious results. Most AI systems and algorithms are data driven and require natural language data upon which to be trained. Thus, data is tightly associated to the functionality of these algorithms and systems. These systems generate complex social implications i.e., displaying human-like social biases (e.g. gender bias) that induce technological marginalization and increased feelings of disenfranchisement.Throughout this dissertation, I argue that how harms arise in NLP systems and who is harmed by these biases, can only be conceptualized and understood at the intersection of NLP, justice and equity (e.g., Data Science for Social Good), and the coupled relationships between language and both social and racial hierarchies. I propose to address three questions at this intersection: (1) How can we conceptualize and quantify such aforementioned harms?; (2) How can we introduce a set of measurements to understand "bias'" in NLP systems; and (3) How can we quantitatively and qualitatively ensure "fairness" in NLP systems?.To address these pertinent question, we attempt differentiate the two consequences of predictive bias in NLP: (1) outcome disparities (i.e., racial bias) and (2) error disparities (i.e., poor system performance) to explicate the importance of modeling social factors of language by exploiting NLP tools to examine predictive biases of both binary gender-specific (male and female) and LGBTQIA2S+ representations, and on an English language variety, i.e., African American English (AAE). Language reflects society, ideology, cultural identity, and customs of communicators, as well as their values. Therefore, natural language data, culture and systems are intertwined with social norms. Nevertheless, social media and online services contain rich textual information on topics surrounding ethnicity, gender identity and sexual orientation-members of the LGBTQIA2S+ community and language (e.g., AAE). This facilitates the collection of large-scale corpora to study social biases in NLP systems in hopes of reducing stigmatization, marginalization, mischaracterization, or erasure of dialectal languages and its speakers, pushing back against potentially discriminatory practices (in many cases-discriminatory through oversight more than malice). In this dissertation, I propose several studies to minimize the gaps between gender, race and NLP systems' performance within the scope of the three aforementioned questions.