Sweet Apple, company? or food? Adjective-centric commonsense knowledge acquisition with taxonomy-guided induction

Abstract

How to enable intelligent machines to possess human commonsense knowledge is one of the central concerns of artificial intelligence. Consequently, a series of commonsense knowledge bases have been designed and constructed by scholars. However, as an important kind of commonsense knowledge, adjective-centric commonsense knowledge (a crucial type of adjective-concept pair such as (blue, sky), (eatable, food)) is far from satisfactory due to the lack of existing knowledge bases and limited acquisition methods. In this paper, we concentrate on automatically constructing large-scale adjective-centric commonsense knowledge bases and propose an effective framework to achieve the goal. The framework mainly contains a filtering module to remove unreasonable inputs, a clustering module, and a conceptualization module to obtain adjective-concept pairs and an evaluation module to assess their plausibility. Extensive automatic and human evaluation results demonstrate the effectiveness of our method, and we finally harvest over 200 k adjective-centric commonsense knowledge, where 81.56% of the implicit commonsense knowledge is not covered by WebChild.11WebChild is a large-scale commonsense knowledge base, which contains triples that connect nouns with adjectives via fine-grained relations. Moreover, we also show that our mined adjective-centric commonsense knowledge can benefit the downstream query conceptualization task.