Multi-Cultural Commonsense Knowledge Base Construction

From International Center for Computational Logic

Multi-Cultural Commonsense Knowledge Base Construction

Talk by Simon Razniewski
Commonsense knowledge (CSK) about concepts and their properties is useful for AI applications such as robust dialogue. Prior works like ConceptNet, TupleKB and others compiled noteworthy commonsense knowledge bases (CSKBs), but are restricted in their expressiveness to subject-predicate-object (SPO) triples with simple concepts for S and monolithic strings for P and O. Also, these projects have either prioritized precision or recall, but hardly reconcile these complementary goals. In this talk I will present several of our CSKB construction works (DICE, ASCENT, CANDLE, MANGO), where we automatically build large-scale CSKBs with advanced expressiveness and both better precision and recall than prior works.

In DICE, we introduce multi-faceted scoring, and joint reasoning for consistency and corroboration. With ASCENT, we go beyond triples by capturing composite concepts with subgroups and aspects, and by refining assertions with semantic facets. CANDLE focuses on multi-cultural CSK, while MANGO unifies the entity-centric model of ASCENT and the culture-centric model of CANDLE.

The works rely on a combination of textual information extraction and knowledge distillation from LLMs. Intrinsic evaluation shows the superior size and quality of our CSKBs, and extrinsic use cases prove their utility for retrieval-augmented generation.

Links:

Project overview: https://www.mpi-inf.mpg.de/commonsense

Ascent: https://ascent.mpi-inf.mpg.de

Candle: https://candle.mpi-inf.mpg.de

Mango: https://bit.ly/cultural-csk