Towards Sound Measurement of Stereotypes and Attitudes in Humans and Language Models
Dominik Bachmann

Abstract:
Much of empirical science involves measurement: compressing the near-infinite complexity of the outside world into a finite set of data (usually in the form of numbers). Measurement is potentially consequential: Based on data, decisions (e.g., diagnoses) are made, and conclusions are drawn about the world. These decisions and conclusions are valid only to the extent to which the data mean what one thinks they mean. Given the importance of good measurement in empirical research, the goal of this dissertation is to contribute to the development of high quality measurement tools for humans and language models.

Part 1 of this dissertation bridges Natural Language Processing (NLP; i.e., the subfield of artificial intelligence specialized in understanding and acting on human language) and Psychometrics (i.e., the subfield of psychology concerned with the theory and practice of measurement). Here, we argue that psychometrics provides NLP with theoretical lenses (i.e., reliability and validity, Chapter 2, and item response theory, Chapter 3) that can help address shortcomings of contemporary NLP bias measurement tools. Additionally, to support the validation of those NLP gender bias measurement tools that rely on occupational terms, we provide a large dataset of human ratings for occupational gender stereotypes (Chapter 4).

Part 2 explores theoretical assumptions that underlie an approach to measuring human stereotypes, and an approach to approximating human opinions. Specifically, in Chapter 5, we test whether, as is often assumed, evidence for gender stereotypes can be obtained with a sequential priming paradigm, when participants are unaware of what we attempt to measure. Subsequently, in Chapter 6, we discuss a pilot questionnaire study, in which we measured people’s newspaper consumption and several of their (political) opinions. This pilot study was part of a project with which we aimed to assess the extent to which Large Language Models can predict how different groups of people respond to different questionnaire items (i.e., an approach called silicon sampling).

The main argument of this dissertation is that assumptions that underlie measurement tools and one’s process of measuring deserve to be made explicit, communicated, scrutinized, and tested.