Exploring sequence-to-sequence taxonomy expansion via language model probing

Taxonomy is a knowledge graph of concept hierarchy which plays a significant role in semantic entailment and is widely used in many downstream natural language processing tasks. Distinct from building a taxonomy from scratch, the task of taxonomy expansion aims at enriching an existing taxonomy by adding new concepts. However, existing methods often construct only part of semantic relationships for representing the taxonomy, which may ignore sufficient features. Meanwhile, as many recent models usually take this task in insertion only manner, they preserve limitations when the new concept is not an insertion to taxonomy. Therefore, we propose TaxoSeq, a method that converts the task of taxonomy expansion into a sequence to sequence setting, thereby effectively exploiting the entire structural features and naturally dealing with more expansion cases. Empowered by pre-trained language models such as T5, our approach is shown to achieve significant progress over other methods in SemEval's three publicly benchmark datasets.

Automated summary

Taxonomy is a knowledge graph used in semantic entailment and natural language processing tasks. Taxonomy expansion involves adding new concepts to enrich an existing taxonomy. Our method, TaxoSeq, converts taxonomy expansion into a sequence to sequence setting, effectively utilizing structural features and handling various expansion cases. It outperforms other methods on SemEval's benchmark datasets.