Research Stream: Social Technologies
Author: Manokamna Singh, PhD Researcher at Maynooth University’s Department of Psychology, Research Funded through the Science Foundation of Ireland (SFI) Centre for Research Training in Advanced Networks for Sustainable Societies (ADVANCE CRT)
Decision-making is a crucial part of human life. Human beings take decisions and make choices on a daily basis to drive their social life. The decisions made produce an impact on the family and shape the future of society. The reasoning behind the decisions can be implicit and explicit attitudes. Implicit attitudes are evaluations that occur without deliberate or unconscious thought. The primary factors which shape the implicit attitudes are past experiences, cultural norms, and social influences. Explicit attitudes are conscious evaluations that everyone can articulate and are aware of holding. This attitude is a reflection of the beliefs and ideals that people consciously support. Unconscious bias is also commonly referred to as implicit bias, as noted by Lopez (2018). The term was first coined in 1995 by Mazarin Banaji and Anthony Greenwald in their article on implicit social cognition. The two psychologists argued that social behavior was significantly affected by unconscious associations and judgments.
The Implicit Association Test (IAT), as proposed by Greenwald, McGhee, and Schwartz in 1998, assesses individual variations in implicit social cognition. The IAT is a psychological tool designed to measure the strength of an individual’s automatic associations between mental concepts, such as groups (e.g., racial categories, gender) and evaluations (e.g., good, bad) or stereotypes (e.g., career, family). In this classification task, participants classify words or images into categories. The categories are paired in various ways and reaction times are recorded. Faster response times are thought to reflect stronger implicit associations between the paired concepts. The IAT uses recorded responses to compute a D-score, which reflects the strength of implicit associations. It has been observed after data analysis that faster reactions for congruent pairings (e.g., “male-career”) suggest a stronger implicit association while slower reactions for incongruent pairings (e.g., “male-family”) suggest weaker or opposite associations.
Text data and human learning are deeply interconnected, as text is a primary medium through which humans acquire, share, and process knowledge. This relationship plays a crucial role in both traditional education and modern applications of artificial intelligence. The differences between human and machine learning—when it comes to language (as well as other domains)—are stark. Whereas LLMs are introduced to and trained with trillions of words of text, human language training happens at a much slower rate. To illustrate, a human infant or child hears—from parents, teachers, siblings, friends, and their surroundings—an average of roughly 20,000 words a day (e.g., Hart and Risley 2003, Gilkerson et al. 2017). So, in its first five years, a child might be exposed to—or trained with—some 36.5 million words. By comparison, LLMs are trained with trillions of tokens within a short time interval of weeks or months.
To assess such unconscious attitudes, researchers use paradigms like the IAT that do not rely on explicit responding to determine the level of bias a person holds towards a particular target concept (e.g., race, gender, age). However, attitudes can evolve and change over time depending on our everyday experience. An important source of attitudinal information comes from the language we are exposed to, and people are sensitive to the statistical patterns that exist between words (i.e., whether a concept is generally associated with positive or negative attributes). Such linguistic distributional knowledge can provide important information about where attitudes come from and how they change over time. It has been shown that large-scale corpora and advancements in natural language processing can play a role in improving our ability to model human cognition and behaviorand to make progress on understanding human attitudes.
The Linguistic Distribution Model (LDM) can be interpreted as a framework or concept in linguistics and computational linguistics that addresses how language elements are distributed across various contexts, populations, or systems. We examined three families of LDM: count vector models, n-gram models, and predict models. In my research study, we examine and investigate a range of linguistic distribution models (varying by corpus, semantic distance measure, and other parameters) in their ability to capture people’s group-level Implicit Biases, established from behavioural experiments with the IAT. We provide a systematic evaluation of a wide range of machine learning models to measure unconscious biases (IAT-D score) for a wide range of topics of social significance. The topics include religion (e.g., Jews–Muslims); political issues (e.g., gun rights, abortion), and other characteristics (e.g., tall people–short people, Whites–Asians). The model uses different sets of publicly available corpora such as British Web (ukWaC) and British National Corpus (BNC) to predict D score. The study compares and benchmarks the correlation between IAT-D score of linguistic models and behavioral experiments to establish the usefulness of linguistic distribution model. The correlation score will establish the contribution of linguistic distribution model to cognition such as self-reflective tool for unconsciousness biasness.
