Home ➤ Projects ➤

Common Sense Reasoning

The Common Sense Reasoning project focuses on advancing culturally grounded evaluation and knowledge representation for language-based artificial intelligence systems. Commonsense reasoning is a critical capability for intelligent systems, yet existing benchmarks and resources often overlook cultural and linguistic diversity. This project addresses these limitations by developing datasets, evaluation frameworks, and knowledge resources that capture everyday reasoning across languages and cultural contexts.
A major component of the project contributes to the development and analysis of a large multilingual benchmark designed to evaluate commonsense reasoning across more than one hundred languages. Constructed through a global collaborative effort involving hundreds of researchers, the benchmark includes culturally specific scenarios referencing local customs, foods, and traditions. Experimental findings reveal that while modern language models perform well overall, their performance drops significantly for lower-resource languages, highlighting the need for more inclusive evaluation resources.
Complementing this effort, the project also introduces a culturally grounded commonsense reasoning dataset tailored to the Sinhala language, incorporating locally relevant examples reflecting Sri Lankan contexts. In addition, earlier work explores methods for automatically generating commonsense knowledge bases by extracting semantic relationships from large text corpora using rule-based and learning-based approaches. Together, these contributions aim to strengthen culturally aware commonsense reasoning capabilities in intelligent systems and improve evaluation and resource availability for underrepresented languages.

Objectives:

• Develop culturally grounded commonsense reasoning resources by creating datasets and benchmarks that capture everyday reasoning across diverse languages and cultural contexts.
• Evaluate the commonsense reasoning capabilities of language models across multilingual and culturally specific scenarios, with particular attention to performance gaps in lower-resource languages.
• Construct and expand commonsense knowledge representations by extracting semantic relationships from large text corpora using rule-based and learning-based approaches.
• Support underrepresented languages in commonsense reasoning research by developing language-specific datasets and resources that reflect local cultural knowledge and real-world contexts.

Keywords: Natural Language Processing | Machine Learning / Deep Learning | Ontologies | Sinhala | Common Sense Reasoning |

Publications

Team

External Collaborators: | Nirmal Fernando | Danaja Maldeniya | Chamilka Wijeratne | Shehan Perera | Ben Goertzel | Tyler A Chang | Catherine Arnett |