A manual experiment on commonsense knowledge acquisition from web corpora

Acquiring commonsense knowledge from text is an important but challenging problem. In this paper, we described a three-subject experiment on commonsense knowledge acquisition from Chinese sentences extracted from a web corpus, aiming to investigate how people acquire commonsensical assertions from given sentences. We analyzed the experiment results from the perspectives of agreement test, concordance test, and divergence test. An important conclusion of our experiment is that sentences are different in their suitability, i.e. difficulty grade, for commonsense knowledge acquisition. And this difficulty grade also affects the number of commonsensical assertions acquired from a sentence, as well as the difference among the acquisition performances of different human subjects. We also discussed the problem of characterizing the difficulty grade by co-occurrence frequency of words and basic level category words.