Critical Thinking for Language Models

  • 2020-09-15 15:49:19
  • Gregor Betz
  • 0


This paper takes a first step towards a critical thinking curriculum forneural auto-regressive language models. We introduce a synthetic text corpus ofdeductively valid arguments, and use this artificial argument corpus to trainand evaluate GPT-2. Significant transfer learning effects can be observed:Training a model on a few simple core schemes allows it to accurately completeconclusions of different, and more complex types of arguments, too. Thelanguage models seem to connect and generalize the core argument schemes in acorrect way. Moreover, we obtain consistent and promising results for the GLUEand SNLI benchmarks. The findings suggest that there might exist arepresentative sample of paradigmatic instances of good reasoning that willsuffice to acquire general reasoning skills and that might form the core of acritical thinking curriculum for language models.


Quick Read (beta)

loading the full paper ...