Insert or Attach: Taxonomy Completion via Box Embedding

Anonymous

Abstract: Taxonomy completion, enriching existing taxonomies by inserting new concepts as parents or attaching them as children, has gained significant interest. Previous approaches embed concepts as vectors in Euclidean space, which makes it difficult to model asymmetric relations in taxonomy. In addition, they introduce pseudo-leaves to convert attachment cases into insertion cases, leading to an incorrect bias in network learning dominated by numerous pseudo-leaves. To address these issues, we propose a novel taxonomy completion framework, called \textsc{TaxBox}, which leverages the geometric properties of insertions and attachments in the box embedding space. By mapping concepts to box embeddings, \textsc{TaxBox} can capture the complex relations between them, relying on the geometric connections between boxes. We also introduce a granular box constraint loss based on the hierarchy of the taxonomy, leading to more accurate concept mapping. Moreover, we design two geometric scorers, one for insertion and the other for attachment, which take into account the distinct behaviors of these two operations in the box embedding space. To balance the scores from the two scorers, we employ a dynamic ranking loss to adaptively adjust the magnitudes of the insertion score and the attachment score. Experiments on four real-world datasets show that \textsc{TaxBox} significantly outperforms previous methods, yielding average performance improvements of 6.7\%, 34.9\%, and 51.4\% in MRR, Hit@1, and Prec@1, respectively.

Paper Type: long

Research Area: NLP Applications

Languages Studied: English

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.