Taxonomies are valuable resources for many applications, but the limited coverage due to the expensive manual curation process hinders their general applicability. Prior works attempt to automatically expand existing taxonomies to improve their coverage by learning concept embeddings in Euclidean space, while taxonomies, inherently hierarchical, more naturally align with the geometric properties of a hyperbolic space. In this paper, we present HyperExpan, a taxonomy expansion algorithm that seeks to preserve the structure of a taxonomy in a more expressive hyperbolic embedding space and learn to represent concepts and their relations with a Hyperbolic Graph Neural Network (HGNN). Speciﬁcally, HyperExpan leverages position embeddings to exploit the structure of the existing taxonomies, and characterizes the concept proﬁle information to support the inference on unseen concepts during training. Experiments show that our proposed HyperExpan outperforms baseline models with representation learning in a Euclidean feature space and achieves state-of-the-art performance on the taxonomy expansion benchmarks.
Is hyperbolic space better than the Euclidean space for taxonomy concept representation? Our #EMNLP2021 Findings paper (w/ @muhao_chen, @eggnome0411, @VioletNPeng) HyperExpan presents a taxonomy expansion method using hyperbolic representation learning https://t.co/KPsT3UbxHr 1/3 pic.twitter.com/0X7LfbRKSl— Mingyu Derek MA (@mingyu_ma) November 7, 2021
As the taxonomy level increases, the number of concepts grows exponentially. We show that by learning concept representations in a more expressive negative-curved space using a hyperbolic #GNN, HyperExpan performs better than its Euclidean counterparts. 2/3 pic.twitter.com/0eevdllBqV— Mingyu Derek MA (@mingyu_ma) November 7, 2021
Please refer to our paper for more details! https://t.co/9VNE6NN69v Our talk and slides are available at https://t.co/8uwjvTLAHa @uclanlp🐻 @USC_ISI🐎 3/3— Mingyu Derek MA (@mingyu_ma) November 7, 2021
(6/6) Hyperbolic RepL: HyperExpan captures the structure and concept profiles of a taxonomy in a more expressive hyperbolic space w/ trainable curvature, allowing effective taxonomy expansion for unseen concepts (w/ @mingyu_ma, Te-Lin, @VioletNPeng) https://t.co/sQ6Aau4768— 🌴Muhao Chen🌴 (@muhao_chen) November 6, 2021
.@mingyu_ma investigated whether hyperbolic space is better than the Euclidean space for taxonomy concept representation.https://t.co/vMsNiwkrug— uclanlp (@uclanlp) November 8, 2021