Comparing Constraints for Taxonomic Organization

Abstract

Building a taxonomy from the ground up involves several sub-tasks: selecting terms to include, predicting semantic relations between terms, and selecting a subset of relational instances to keep, given constraints on the taxonomy graph. Methods for this final steptaxonomic organization-vary both in terms of the constraints they impose, and whether they enable discovery of synonymous terms. It is hard to isolate the impact of these factors on the quality of the resulting taxonomy because organization methods are rarely compared directly. In this paper, we present a head-to-head comparison of six taxonomic organization algorithms that vary with respect to their structural and transitivity constraints, and treatment of synonymy. We find that while transitive algorithms out-perform their non-transitive counterparts, the top-performing transitive algorithm is prohibitively slow for taxonomies with as few as 50 entities. We propose a simple modification to a non-transitive optimum branching algorithm to explicitly incorporate synonymy, resulting in a method that is substantially faster than the best transitive algorithm while giving complementary performance.

TaxoExpan: Self-supervised Taxonomy Expansion with Position-Enhanced Graph Neural Network

“…Finally, we show that TaxoExpan-FWFS can achieve the new state-of-the-art performance on this dataset. 10 We use the wiki-news-300d-1M-subword.vec.zip version on fastText official website. 11 This metric is used because the original task allows a model to decline to place new concepts in order to avoid making placements with low confidence.…”

“…To reduce the human efforts, many automatic taxonomy construction methods [31,41,60] are proposed. They first identify "is-A" relations (e.g., "iPad" is an "Electronics") using textual patterns [16,38] or distributional similarities [3,43], and then organize extracted concept pairs into a directed acyclic graph (DAG) as the output taxonomy [10,14,24]. As the web contents and human knowledge are constantly growing, people need to expand an existing taxonomy to include new emerging concepts.…”

“…In this work, we also follow this convention and attach each new concept to the top-ranked synset in the WordNet. Finally, we obtain the initial feature vectors (for both new concepts and existing words in the WordNet) using pre-trained subword-aware fasttext embeddings 10 . For each concept, we generate its definition embedding and name embedding by averaging the embedding of each token in its textual definition and name string, correspondingly.…”

Enquire One's Parent and Child Before Decision: Fully Exploit Hierarchical Structure for Self-Supervised Taxonomy Expansion

“…The first focuses on topic-based taxonomy, where each node is a cluster of several terms sharing the same topic [32,48]. The other subdivision tackles the problem of term-based taxonomy construction, in which each node represents the term itself [3,24,35]. A typical pipeline for this task is to extract "is-A" relations with a hypernymy detection model first using either a pattern-based model [1,8,11,28] or a distributional model [4,18,42,45], then integrate and prune the mined hypernym-hyponym pairs into a single directed acyclic graph (DAG) or tree [7].…”

STEAM: Self-Supervised Taxonomy Expansion with Mini-Paths

“…Focused on taxonomy induction, these methods organize hypernymy pairs into taxonomies. Graph optimization techniques [3,8,13,19] have been proposed to organize the hypernymy graph into a hierarchical structure, and Mao et al [26] utilize reinforcement learning to organize term pairs by optimizing a holistic tree metric over the training taxonomies. Very recently, Shang et al [33] design a transfer framework to use the knowledge from existing domains for generating taxonomy for a new domain.…”