Gender Bias in Word Embeddings: A Theological Perspective

How do we teach a computer the meaning of words? Word embeddings, a breakthrough of natural language processing, render the semantics of words computationally tractable by representing them as vectors in a high-dimensional space—exemplifying, though they emerged from statistics and computer science, two key ideas in the philosophy of language: Frege’s principle that words have meaning only in the context of a sentence and Wittgenstein’s dictum that the meaning of a word is its use in the language. But embeddings also absorb latent forms of bias from the corpora on which they are trained. In this presentation, I introduce word embeddings from a theological perspective, examine techniques for identifying and rectifying gender bias in word embeddings, and ask what tools developed to measure demographic biases might reveal about biases related to religious identity. In an accompanying hands-on session, we explore latent biases in a large corpus of contemporary news.

Jun 5, 2025 12:00 AM — Jun 7, 2025 1:00 AM

Theologies of the Digital III: Biases and Debiasing Theology, Garrett Evangelical Theological Seminary