Common Sense Knowledge Graphs (CSKGs)

February 8 or 9

Held virtually in conjunction with AAAI’21

Commonsense knowledge graphs (CSKGs) are sources of background knowledge that are expected to contribute to downstream tasks like question answering, robot manipulation, and planning. The knowledge covered in CSKGs varies greatly, spanning procedural, conceptual, and syntactic knowledge, among others. CSKGs come in a wider variety of forms compared to traditional knowledge graphs, ranging from (semi-)structured knowledge graphs, such as ConceptNet, ATOMIC, and FrameNet, to the recent idea to use language models as knowledge graphs. As a consequence, traditional methods of integration and usage of knowledge graphs might need to be expanded when dealing with CSKGs. Understanding how to best integrate and represent CSKGs, leverage them on a downstream task, and tailor their knowledge to the particularities of the task, are open challenges today. The workshop on CSKGs addresses these challenges, by focusing on the creation of commonsense knowledge graphs and their usage on downstream commonsense reasoning tasks.

Key Dates

September 9: First call for papers
November 9: Workshop submissions due
November 30: Notifications sent to authors
January 15: Release of final workshop schedule
February 8 or 9: Workshop (Full day)

Confirmed Keynote Speakers and Panelists

  • Yejin Choi
    University of Washington & AI2
  • Joshua Tenenbaum
  • David Ferrucci
    Elemental Cognition
  • Shih-Fu Chang
    Colombia University


Topics of interest include, but are not limited to:

  • Creation/extraction of new CSKGs
  • Integration of existing CSKGs
  • Exploration of CSKGs
  • Impact of CSKGs on downstream tasks
  • Methods of including CSKG knowledge in downstream tasks
  • Probing for knowledge needs in downstream tasks
  • Evaluation data/metrics relevant for CSKGs
  • Identifying and/or filling gaps in CSKGs


We welcome submissions of long (max. 8 pages), short (max. 4 pages), and position (max. 4 pages) papers describing new, previously unpublished research in this field. The page limits are including the references. Submissions must be formatted in the AAAI submission format. All submissions should be done electronically via EasyChair:

Panel: Are language models enough?

(Deep) Language models are so popular these days that it’s becoming harder to find a scientific publication in the field of Natural Language Processing (NLP) that doesn’t use any. Popularity is well-deserved: BERT and its descendants (“BERTology”), GPT-3, T5, etc. have been quite phenomenal in improving the state of the art in a wide range of NLP tasks, from semantic parsing to question answering and text-generation. But “are language models enough to understand meaning as broadly and deeply as humans do?” Are the trivial errors that language models (still profusely) make just an indication that we need larger and better curated corpora, and proportionally higher-capacity models? Or are these errors clues that language models - no matter how many billion words have been used to train them - can only capture surface-level meanings, i.e. semantic features that are exhibited by training data? We invited top experts in the field to help us shed some light on these questions.


  • Filip Ilievski
  • Alessandro Oltramari
    Bosch Research and Technology Center (Pittsburgh)
  • Deborah McGuinness
    Rensselaer Polytechnic Institute
  • Pedro Szekely

Program Committee

  • Chandra Bhagavatula (Allen Institute for AI, Washington, USA)
  • Michele Catasta (Stanford University, California, USA)
  • Marjorie Friedman (USC Information Sciences Institute, Massachusetts, USA)
  • Aldo Gangemi (University of Bologna and ISTC, National Research Council, Italy)
  • Henry Lieberman (MIT, Massachusetts, USA)
  • Roberto Navigli (Sapienza University of Rome, Italy)
  • Valentina Presutti (ISTC, National Research Council, Italy)
  • Simon Razniewski (Max Planck Institute for Informatics, Germany)
  • German Rigau (University of the Basque Country, Spain)
  • Daniel Schwabe (Pontificia Universidade Católica (PUC), Rio de Janeiro, Brazil)
  • Niket Tandon (Allen Institute for AI, Washington, USA)
  • Piek Vossen (VU Amsterdam, The Netherlands)