LAB42 Talk | CLS: Pia Sommerauer (VU)

LAB42, L3.36

Join us in this Computational Linguistics Seminar from Pia Sommerauer, Vrije Universiteit Amsterdam

Analyzing linguistic subtleties in language models: detecting shifts in connotation and removing biases

(Large) language models achieve impressive results on various tasks based on purely distributional data; they still rely on learning (possibly complex and sophisticated) associations between words and their contexts. It is difficult to tell to what degree language models can reflect a human-like understanding of semantics and how specific information is encoded. Can models pick up subtle differences in connotation? To what degree can contextualized language models actually reflect contextual information? And, assuming specific information is captured by the model, can we remove it without changing other information? In this talk, I present two studies that aim to examine what information language models can infer based on distributional data and how we can manipulate this information: Firstly, I will present insights from a study about detecting semantic shifts in connotation between political communities using static and contextualized embeddings. We present a small, expert-informed dataset for synchronic shift detection between political communities in Dutch and English. Our experiments show that static and contextualized models can (if applied in the right way), detect subtle shifts in politically loaded terms. Secondly, I will discuss findings from a study about model debiasing methods that rely on removing specific information from models. We propose a new method for bias removal that interferes less with the embedding space than previously proposed methods. As such, it has potential for causal probing approaches. Both studies raise questions about the representations and behavior of language models.

Computational Linguistics Seminar

The CLS is the Computational Linguistics Seminar of the University of Amsterdam. Seminars are open to all interested researchers and students of all levels from UvA and elsewhere. The CLS is organized and sponsored by the Institute for Logic, Language and Computation's NLP and Digital Humanities research unit, and sponsored by the ELLIS Unit Amsterdam.


The seminar is organised by Jelke Bloem and Alina Leidinger.

To receive notifications about upcoming talks and the Zoom details, please join the CLS mailing list.