Negation: An Effective Method to Generate Hard Negatives

Reasoning commonsense knowledge is essential for Artificial Intelligence, which requires high-quality commonsense knowledge. Recently, much progress has been made in automatic commonsense knowledge generation. However, most of the works focus on obtaining positive knowledge and lack negative information. Only a few works capture the importance of negative statements, but they struggle to produce high-quality knowledge. Although some efforts have been made to generate negative statements, they fail to consider the taxonomic hierarchy between entities and are not generally applicable, leading to the generation of low-quality negative samples. To resolve the issue, we put forward Negation, a framework for effectively generating hard negative knowledge. For each entity in the commonsense knowledge base, congeners are identified with hierarchical and semantic information. Then, negative candidates are produced by replacing the entity with congeners in each triple. In order to make negative knowledge more confusing and avoid false positive examples, we design two filtering steps to remove the amount of meaningless candidates. We empirically evaluate our proposed method Negation on the downstream task, and the results demonstrate that Negation and its components effectively help generate high-quality negative knowledge.