We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: PInKS: Preconditioned Commonsense Inference with Minimal Supervision

Abstract: Reasoning with preconditions such as "glass can be used for drinking water unless the glass is shattered" remains an open problem for language models. The main challenge lies in the scarcity of preconditions data and the model's lack of support for such reasoning. We present PInKS, Preconditioned Commonsense Inference with WeaK Supervision, an improved model for reasoning with preconditions through minimum supervision. We show, both empirically and theoretically, that PInKS improves the results on benchmarks focused on reasoning with the preconditions of commonsense knowledge (up to 40% Macro-F1 scores). We further investigate PInKS through PAC-Bayesian informativeness analysis, precision measures, and ablation study.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2206.07920 [cs.CL]
  (or arXiv:2206.07920v1 [cs.CL] for this version)

Submission history

From: Ehsan Qasemi [view email]
[v1] Thu, 16 Jun 2022 04:50:00 GMT (850kb,D)

Link back to: arXiv, form interface, contact.