PaperlayerView exact revision on arXiv

Source-linked AI summary

Relation-Aware Neighborhood Matching Model for Entity Alignment

Yao Zhu, Hongzhi Liu, Zhonghai Wu, Yingpeng Du

arXiv:2012.08128v1cs.AIcs.CL

TL;DR

Entity alignment links equivalent entities across knowledge graphs, but existing neighborhood methods can introduce noise, omit relations, and overlook interactions with relation alignment. RNM matches neighborhoods using connected relations and iteratively aligns entities and relations in a semi-supervised manner. It outperforms state-of-the-art methods on three real-world datasets.

  • Problem

    Entity alignment supports knowledge fusion, but existing neighborhood methods may add noise, omit connected relations, and pay limited attention to interactions with relation alignment.

    Source: Introduction · pp. 1–2

  • Method

    RNM uses relation-aware neighborhood matching and semi-supervised iterative alignment of entities and relations.

    Source: Introduction · p. 2

  • Results

    RNM significantly outperforms state-of-the-art methods on three real-world datasets; its Hits@1 improvement over NMN is 10.7%, 8.7%, and 3.6%.

    Source: Introduction · p. 2, Entity Alignment · p. 6

  • Takeaways & Limitations

    The reported experiments support RNM’s effectiveness for entity alignment across three cross-lingual knowledge graph datasets.

    Source: Conclusion and Future Work · p. 7

    The problem setup assumes a set of pre-aligned entity pairs between the two knowledge graphs.

    Source: Problem Definition · p. 2

Abstract

from arXiv · show

Entity alignment which aims at linking entities with the same meaning from different knowledge graphs (KGs) is a vital step for knowledge fusion. Existing research focused on learning embeddings of entities by utilizing structural information of KGs for entity alignment. These methods can aggregate information from neighboring nodes but may also bring noise from neighbors. Most recently, several researchers attempted to compare neighboring nodes in pairs to enhance the entity alignment. However, they ignored the relations between entities which are also important for neighborhood matching. In addition, existing methods paid less attention to the positive interactions between the entity alignment and the relation alignment. To deal with these issues, we propose a novel Relation-aware Neighborhood Matching model named RNM for entity alignment. Specifically, we propose to utilize the neighborhood matching to enhance the entity alignment. Besides comparing neighbor nodes when matching neighborhood, we also try to explore useful information from the connected relations. Moreover, an iterative framework is designed to leverage the positive interactions between the entity alignment and the relation alignment in a semi-supervised manner. Experimental results on three real-world datasets demonstrate that the proposed model RNM performs better than state-of-the-art methods.

Introduction

Knowledge graphs can be incomplete and complementary, making entity alignment important for knowledge fusion. RNM addresses gaps in neighborhood matching by using connected relations and iteratively aligning entities and relations.

Introduction

  • Knowledge graphs store facts as triples, and integrating complementary graphs supports knowledge fusion; entity alignment links equivalent entities across graphs.The triples are represented as (h, r, t), with relation r from head entity h to tail entity t.Source: Introduction · p. 1
  • Embedding-based methods use entity-vector distances for alignment, while GCNs incorporate neighbor information but may introduce noise.TransE was widely adopted for entity alignment.Source: Introduction · p. 1
  • Existing subgraph-matching methods compare neighboring entities but ignore connected relations and give less attention to interactions between entity and relation alignment.The Italy example illustrates how matching relations as well as neighbors can distinguish candidates.Source: Introduction · p. 1, Figure 1 · p. 2
  • RNM combines relation-aware neighborhood matching with semi-supervised iterative entity and relation alignment, and outperforms state-of-the-art methods on three real-world datasets.It uses relations’ semantic information and mapping properties during matching.Source: Introduction · p. 2

Related Work

Prior entity-alignment work developed embedding-based and graph-based representations, then explored subgraph matching. RNM extends neighborhood matching by incorporating relations and iterative entity–relation alignment.

Related Work

  • Embedding-based approaches include TransE, separate-space transformation, path-based joint embeddings, and bootstrapping to expand seed alignments.BootEA also uses constraints to reduce error accumulation during iteration.Source: Related Work · p. 2
  • GCN-based methods incorporate structural and neighborhood information into entity representations, but neighbor aggregation may introduce noise that degrades alignment.The cited approaches include GCN-Align, HGCN, RDGCN, and AliNet.Source: Related Work · p. 2
  • Subgraph-matching methods compare local neighborhoods, but the described approaches omit connected relations from the comparison.KG-alignment formulates alignment as graph matching; NMN jointly encodes neighborhood differences across candidate pairs.Source: Related Work · p. 2
  • RNM explores relations’ semantic information and mapping properties during subgraph matching, while iteratively aligning entities and relations so the tasks reinforce each other.This addresses the relation omission and limited entity–relation interaction described for prior methods.Source: Related Work · p. 2

Problem Definition

The problem is to find equivalent entity pairs and same-meaning relation pairs across two heterogeneous knowledge graphs. The setup assumes some entity pairs are already aligned, while relation alignment is unsupervised.

Problem Definition

  • Entity alignment and relation alignment are treated as related tasks for knowledge fusion.Source: Problem Definition · p. 2
  • A knowledge graph is represented as G = (E, R, T), with sets of entities, relations, and triples, and the two graphs have a set of pre-aligned entity pairs.The pre-aligned set is L = {(e1, e2)|e1 ∈ E1, e2 ∈ E2, e1 equals to e2}.Source: Problem Definition · p. 2
  • Entity alignment seeks the remaining equivalent entity pairs, while relation alignment seeks same-meaning relation pairs and is unsupervised in this paper.Source: Problem Definition · p. 3

Proposed Model

RNM jointly learns entity and relation representations, then iteratively aligns both through relation-aware neighborhood matching for entities and entity-aware matching for relations.

Embedding Learning for Entity and Relation

Relation-Aware Neighborhood Matching

  • For entity candidates, RNM compares one-hop neighbor and relation pairs, then scores matched neighbors using relation mapping probabilities.A 1-to-1 relation can provide exact alignment, while a 1-to-N relation indicates probability 1/N; matched cases require both aligned neighbors and aligned relations.Source: Relation-Aware Neighborhood Matching · p. 4

Entity-Aware Relation Matching

  • For relation candidates, RNM matches head and tail entity pairs and combines that matching score with the relation embedding distance.The method treats simultaneous head- and tail-entity alignments as evidence that two relations have the same meaning.Source: Entity-Aware Relation Matching · p. 4

Iterative Strategy and Implementation Details

  • A semi-supervised iterative framework updates entity and relation alignments and their distance matrices using the two matching modules.Iterations stop at convergence or a maximum iteration count; the resulting matrices support alignment ranking or alignment-set generation.Source: Iterative Strategy and Implementation… · pp. 4–5
  • The alignment-set procedure accepts nearest cross-KG pairs below separate entity and relation thresholds, resolving conflicts by smaller distance; reverse triples also enlarge each KG.For example, a triple (Tokyo, CapitalOf, Japan) is supplemented with (Japan, CapitalOf^-1, Tokyo).Source: Iterative Strategy and Implementation… · p. 5

Experiments

Experiments use three DBP15K cross-lingual datasets and compare RNM with entity- and relation-alignment baselines. RNM outperforms the entity-alignment baselines, and results also support the value of relation-aware matching and iterative alignment.

Experimental Setup

  • The evaluation uses three DBP15K cross-lingual datasets, each comprising two language-specific KGs and 15,000 aligned entity pairs.The datasets are drawn from English, Chinese, Japanese, and French DBpedia versions.Source: Datasets · p. 5, Table 1 · p. 5
  • Evaluation uses Hits@1, Hits@10, and MRR, with higher values indicating better performance; baselines span TransE-based and GCN-based methods.The GCN baselines include structural-information models and models initialized with entity names.Source: Evaluation Metrics and Baselines · pp. 5–6

Experimental Results

  • RNM significantly outperforms all entity-alignment baselines on three datasets, with Hits@1 above 80%, Hits@10 above 90%, and MRR above 0.85 on each.Table 2 reports performance for the compared entity-alignment methods.Source: Table 2 · p. 6, Entity Alignment · p. 6
  • Removing relation matching causes the largest ablation drop; removing alignment probability lowers Hits@1 by 2.2%, 1.5%, and 0.8% across the three datasets.RNM without iteration still outperforms all Table 2 baselines; the ablations separately remove relations, alignment probability, or the iterative strategy.Source: Table 3 · p. 6, Ablation Study · p. 6
  • RNM with 10% seed alignments outperforms NMN with 40% on all three datasets, and RNM leads NMN at every tested seed proportion.Figure 4 presents entity-alignment results at different seed-alignment proportions.Source: Figure 4 · p. 7, Analysis · p. 7
  • RNM exceeds the best relation-alignment baseline, HGCN-JR, in Hits@1 by 10.3%, 9.5%, and 7.0% across the three datasets.The authors attribute the gains to matching relations through their head and tail entities and report that entity and relation alignment reinforce each other.Source: Table 4 · p. 7, Relation Alignment · p. 7
  • Entity- and relation-alignment performance both rise as RNM's iteration count increases.Figure 3 reports both alignment results against the number of iterations.Source: Analysis · p. 7, Figure 3 · p. 7

Conclusion and Future Work

RNM combines entity and relation information in an iterative, semi-supervised approach, and experiments on three cross-lingual KG datasets demonstrate its effectiveness. Future work will explore attributes and descriptions as additional side information for entity alignment.

Conclusion and Future Work

Related papers

Loading 2012.08128v1…