aclanthology.org web signal

SocioProbe: PLMs encode gender and age, spread across layers

By Published Updated

TL;DR

  • SocioProbe probes single-GPU PLMs on multiple English data sets using classifier probing and information-theoretic minimum description length probing.
  • The authors find that PLMs do encode gender and age, and that this knowledge is sometimes spread across the layers of tested models.
  • Models that excel in general language understanding do not seem to own more sociodemographic knowledge, which requires large amounts of pre-training data.

Pre-trained language models do encode information about a speaker's gender and age, but the signal is diffuse across the network and does not scale with a model's general-language ability. That is the finding of SocioProbe, a paper by Anne Lauscher, Federico Bianchi, Samuel R. Bowman and Dirk Hovy presented at EMNLP 2022 in Abu Dhabi.

The authors probe several single-GPU PLMs on multiple English data sets using two techniques: traditional classifier probing and information-theoretic minimum description length probing. A multilingual analysis and a study of supplementary training sit alongside the core experiments.

On what the probes recover, the paper is direct: "PLMs do encode these sociodemographics, and that this knowledge is sometimes spread across the layers of some of the tested PLMs." The take-home for practitioners is less flattering to the leaderboard reflex. "PLMs require large amounts of pre-training data to acquire the knowledge," the authors write, "and models that excel in general language understanding do not seem to own more knowledge about these aspects." Their overall verdict is that "sociodemographic knowledge is still a major challenge for NLP."

Two researchers we track flagged the paper's link in our Who's Who directory.

Shared on Bluesky by 2 AI experts

Originally reported by aclanthology.org

Read the original article →

Original headline: SocioProbe: What, When, and Where Language Models Learn about Sociodemographics