Timezone: »

 
Poster
Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models
Maribeth Rauh · John Mellor · Jonathan Uesato · Po-Sen Huang · Johannes Welbl · Laura Weidinger · Sumanth Dathathri · Amelia Glaese · Geoffrey Irving · Iason Gabriel · William Isaac · Lisa Anne Hendricks

Wed Nov 30 09:00 AM -- 11:00 AM (PST) @ Hall J #1014
in Poster Session 3 »

Large language models produce human-like text that drive a growing number of applications. However, recent literature and, increasingly, real world observations, have demonstrated that these models can generate language that is toxic, biased, untruthful or otherwise harmful. Though work to evaluate language model harms is under way, translating foresight about which harms may arise into rigorous benchmarks is not straightforward. To facilitate this translation, we outline six ways of characterizing harmful text which merit explicit consideration when designing new benchmarks. We then use these characteristics as a lens to identify trends and gaps in existing benchmarks. Finally, we apply them in a case study of the Perspective API, a toxicity classifier that is widely used in harm benchmarks. Our characteristics provide one piece of the bridge that translates between foresight and effective evaluation.

Keywords: language models · Evaluation · Fairness · NLP · benchmark · toxicity · ethics
[ Paper]  [ Poster]  [ OpenReview] 

Author Information

Maribeth Rauh (DeepMind)
John Mellor (DeepMind)
Jonathan Uesato (Google DeepMind)
Po-Sen Huang (DeepMind)
Johannes Welbl (Google)
Laura Weidinger (DeepMind)
Sumanth Dathathri (DeepMind)
Amelia Glaese (DeepMind)
Geoffrey Irving (DeepMind)
Iason Gabriel (DeepMind)
William Isaac (DeepMind)
Lisa Anne Hendricks (DeepMind)

More from the Same Authors