Theme I - Text Reasoning, Being Right for the Right Reasons 🤔

For humans, intuitive inferences are made every now and then. However, it would require reasons for humans to convince others and justify themselves of their inferences or decisions. How can machines better convince humans of their predictions? The key may lie in making the right and faithful reasons for self-justification.

What is human-like reasoning? What is the holy grail of machine cognition? It is easy to be right due to various spurious correlations, but it would require some actual reasoning skills to be right for the right and faithful reasons. More importantly, symbolic reasoning is too fragile to handle everyday reasoning. Can machine reasoning happen over natural language like humans do?

Exemplar papers in this theme include:

  • NEON (AAAI 2023): a two-phrase, unsupervised explanation generation framework for explaining why a statement is wrong;
  • E-KAR (Findings of ACL 2022): a benchmark for analogical reasoning with free-text rationales for both positive and negative candidate answers;
  • LOREN (AAAI 2022): generating faithful and accurate rationales without supervision;
  • EDUCAT (AAAI 2022): unsupervised counterfactual reasoning for imagining possible outcomes;
  • PRobr (Findings of ACL 2021): reasoning over natural language statements via an induced graphical model.
Jiangjie Chen
Jiangjie Chen
Ph.D. Candidate

His research interests mainly include natural language reasoning and large language models.