Open Knowledge Graphs Canonicalization using Variational Autoencoders

Abstract

Noun phrases and Relation phrases in open knowledge graphs are not canonicalized, leading to an explosion of redundant and ambiguous subject-relation-object triples. Existing approaches to solve this problem take a two-step approach. First, they generate embedding representations for both noun and relation phrases, then a clustering algorithm is used to group them using the embeddings as features. In this work, we propose Canonicalizing Using Variational Autoencoders (CUVA) 1 , a joint model to learn both embeddings and cluster assignments in an end-to-end approach, which leads to a better vector representation for the noun and relation phrases. Our evaluation over multiple benchmarks shows that CUVA outperforms the existing state-of-the-art approaches. Moreover, we introduce CANONICNELL, a novel dataset to evaluate entity canonicalization systems.

Vaccine Coverage and Effectiveness in a School-Based Varicella Outbreak in Jinan Prefecture, Shandong Province

“…Unlike state-of-the-art approaches that focus on equating the OKB canonicalization problem to the clustering problem of NPs (RPs), ignoring the generation of more accurate representations of NPs (RPs), CUVA [12] applies variational autoencoders (VAE) to the process of learning NPs (RPs) representations. Nevertheless, the VAE generation process is not equidimensional, leading to distortion and warping of features.…”

“…In addition to the clustering assignment objective with diffusion model and the KGE learning objective, we also adopt the side information objective Lside by following [12] and [16]. The main motivation is to harness the contextual information to further facilitate the overall learning process.…”

“…• Canonicnell [12] The data in this dataset was collected from [34] generated data artefacts using an automated construction strategy, from which the cumulative knowledge was extracted as a list of (topic, relation, object) triples to form this dataset. This is a newly proposed dataset in which the co-referenced entities of the triples are labelled, pairs with scores not greater than 0.25 are filtered, and the filtered remaining pairs are considered as undirected edges in the graph.…”

“…[16] is recommended for interested authors. Notably, following [12], we do not report the results on relation canonicalization, as benchmarks before do not contain ground-truth annotations for relations.…”

“…Some fuse the two types of information sources in pursuit of a more accurate clustering result [11]. Unlike most of existing solutions that use the hard clustering strategy to assign each NP/RP to a single cluster, a recent work adopts variational autoencoder (VAE), a generative method, to tackle OKB canonicalization [12], which learns the embeddings and cluster assignments in an end-to-end fashion, resulting in better vector representations for NPs/RPs, and also state-of-the-art canonicalization results.…”