PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Knowledge Graphs: Representation, Acquisition and Applications

Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, Philip S. Yu

arXiv:2002.00388v4cs.CLcs.AI

TL;DR

Knowledge graph research spans representation, acquisition, temporal dynamics, and applications, but its methods and tasks require an organized synthesis. This survey develops full-view taxonomies, reviews major and emerging approaches, and provides research resources and future directions.

  • Problem

    Knowledge graph research covers diverse representation, acquisition, temporal, and application tasks that require a comprehensive, organized review.

  • Method

    The survey categorizes knowledge graph research across four scopes and reviews embedding, path-inference, logical-rule, and related methods, applications, datasets, and implementations.

  • Results

    The survey provides a comprehensive literature review, full-view taxonomies, coverage of emerging advances, and a summary of promising future research directions.

  • Takeaways & Limitations

    The survey offers a structured reference for developing knowledge graph models and studying knowledge acquisition, temporal graphs, and knowledge-aware applications.

Abstract

from arXiv · show

Human knowledge provides a formal understanding of the world. Knowledge graphs that represent structural relations between entities have become an increasingly popular research direction towards cognition and human-level intelligence. In this survey, we provide a comprehensive review of knowledge graph covering overall research topics about 1) knowledge graph representation learning, 2) knowledge acquisition and completion, 3) temporal knowledge graph, and 4) knowledge-aware applications, and summarize recent breakthroughs and perspective directions to facilitate future research. We propose a full-view categorization and new taxonomies on these topics. Knowledge graph embedding is organized from four aspects of representation space, scoring function, encoding models, and auxiliary information. For knowledge acquisition, especially knowledge graph completion, embedding methods, path inference, and logical rule reasoning, are reviewed. We further explore several emerging topics, including meta relational learning, commonsense reasoning, and temporal knowledge graphs. To facilitate future research on knowledge graphs, we also provide a curated collection of datasets and open-source libraries on different tasks. In the end, we have a thorough outlook on several promising research directions.

I. INTRODUCTION

This survey reviews knowledge graphs as structured representations of human knowledge, covering representation learning, acquisition, temporal graphs, and knowledge-aware applications. It introduces comprehensive taxonomies, surveys emerging techniques, and summarizes future directions and research resources.

  • Knowledge graph foundations: Knowledge graphs represent structured facts through entities, relationships, semantic descriptions, and formal semantics that support interpretation and inference.
  • Research scope: The survey organizes knowledge graph research into representation learning, knowledge acquisition, temporal knowledge graphs, and knowledge-aware applications.
  • Research scope: Knowledge graph representation learning is categorized by representation space, scoring function, encoding models, and auxiliary information.
  • Research scope: Knowledge acquisition is reviewed through embedding-based ranking, relational path reasoning, logical rule reasoning, meta relational learning, and entity- and relation-oriented tasks.
  • Emerging advances: The review covers emerging advances including transformer-based encoding, graph neural network knowledge propagation, reinforcement-learning path reasoning, and meta relational learning.
  • Resources and outlook: The survey summarizes categories and future directions while providing datasets, open-source implementations, and training-strategy information for further research.

D. Related Surveys

The survey distinguishes its full-view taxonomy from earlier surveys by organizing knowledge representation learning across four complementary scopes and multiple representation spaces.

  • Earlier surveys mainly emphasized statistical relational learning, refinement, reasoning, knowledge graph embedding, or scoring functions.
  • Knowledge representation learning is also called knowledge graph embedding, multi-relation learning, or statistical relational learning.
  • The survey organizes KRL by representation space, scoring function, encoding models, and auxiliary information.
  • Representation spaces include real-valued point-wise, complex vector, Gaussian, and manifold spaces, subject to differentiability, computability, and definable scoring functions.
  • Translational and semantic-matching models use related but distinct combinations of vector, matrix, and tensor representations.

4) Manifold and Group:.

Manifold, hyperbolic, Lie-group, torus, and dihedral representations extend point-wise embeddings with alternative geometric or algebraic structures for relational modeling.

  • Manifold and Group: ManifoldE replaces point-wise embeddings with sphere and hyperplane manifolds to obtain more expressive geometric representations.The survey describes point-wise modeling as an ill-posed algebraic system with overly restrictive geometry.
  • Manifold and Group: Hyperbolic embeddings use negative-curvature spaces to capture hierarchical information, while relation-specific curvature improves flexibility.
  • Manifold and Group: TorusE embeds entities and relations in a compact Lie group while retaining translational scoring through [h] + [r] ≈ [t].
  • Manifold and Group: Knowledge representation spaces include real-valued, complex, Gaussian, and manifold alternatives illustrated across the survey’s space taxonomy.
  • Scoring Function: Distance-based scoring measures relational projection distances, with TransE learning translations from heads toward tails under L1 or L2 constraints.
  • Scoring Function: TransH, TransR, TransD, TransA, TransF, ITransF, TransAt, and TransMS extend translational models through projections, adaptive metrics, attention, or nonlinear transformations.
  • Scoring Function: KG2E and ManifoldE apply translational distance-based scoring in Gaussian and manifold spaces, respectively.

2) Semantic Matching:

Semantic-matching models evaluate relational plausibility through interactions among entity and relation representations, using linear, bilinear, group-based, and compositional formulations.

  • Semantic Matching: SME matches separate combinations of (h, r) and (r, t) through linear and bilinear matching blocks.
  • Semantic Matching: DistMult simplifies bilinear relational representation by restricting relation matrices to diagonal form.
  • Semantic Matching: HolE uses circular correlation as a compressed tensor product, while HolEx interpolates between HolE and the full tensor product.
  • Semantic Matching: ANALOGY constrains relation matrices to normal matrices and can recover DistMult, ComplEx, and HolE under suitable restrictions.
  • Semantic Matching: Crossover interactions combine relation-specific interaction matrices with matching against tail embeddings.
  • Semantic Matching: The semantic-matching principle can also be encoded using neural networks.
  • Semantic Matching: TorusE and DihEdral also follow semantic matching through group-based relation representations and component-wise scoring.

C. Encoding Models

Encoding models represent relational interactions through linear or bilinear mappings, factorization, and neural architectures that range from triplet-level matching to recurrent relational-path modeling.

  • Linear/Bilinear Models: Linear and bilinear models encode entity–relation interactions through projection or bilinear transformation operations.
  • Linear/Bilinear Models: TransE with L2 regularization can be expanded into a linear-transformation form involving head, relation, and tail embeddings.
  • Linear/Bilinear Models: SimplE introduces inverse relations and averages canonical polyadic scores, while RESCAL, DistMult, HolE, and ComplEx are related through constraints.
  • Factorization Models: Factorization methods formulate knowledge representation as three-way tensor decomposition, including RESCAL, latent factor models, TuckER, and LowFER.
  • Neural Networks: Neural encoders use nonlinear architectures such as MLP, NTN, NAM, and CNN-based models to compute semantic matching scores.

4) Convolutional Neural Networks:

The survey reviews convolutional and related neural encoders for modeling entity–relation interactions, then situates them alongside recurrent, transformer, graph, and multimodal approaches.

  • Convolutional Neural Networks: ConvE reshapes head and relation embeddings into 2D matrices and applies convolution plus nonlinear layers to model their interactions.Its score uses convolutional filters, vectorization, and a projection matrix.
  • Convolutional Neural Networks: ConvKB convolves the concatenated entity and relation embeddings without reshaping, preserving transitional characteristics through concatenated feature maps.The survey reports better experimental performance than ConvE, which captures local relationships.
  • Recurrent Networks: Recurrent models learn long-term relational dependencies by encoding relation paths, with RSN adding recurrent skip mechanisms for semantic representation learning.Earlier RNN models differ in whether they incorporate entity information.
  • Transformer Models: Transformer models such as CoKE and KG-BERT encode contextual information from knowledge-graph edges, paths, entities, and relations.KG-BERT adapts BERT-style language-model pretraining to entities and relations.
  • Graph Neural Networks: GNN encoders model graph connectivity through relation-specific transformations, normalization, weighted neighborhoods, attention, or entity–relation composition.R-GCN treats neighboring entities equally, whereas SACN adds weighted graph convolution using node structure, attributes, and relation types.
  • Multimodal Embedding: Multimodal embedding incorporates text descriptions, entity types, relational paths, images, and other auxiliary information to enrich knowledge representation.These methods align structured knowledge with unstructured or visual information in shared or coordinated representations.

3) Visual Information:

The survey organizes knowledge acquisition around completion, entity discovery, and relation-path or rule-based reasoning, emphasizing how structured and textual information support these tasks.

  • Knowledge Acquisition: Knowledge acquisition constructs graphs from unstructured, structured, or semi-structured sources, completes existing graphs, and discovers entities and relations.The reviewed categories are knowledge graph completion, entity discovery, and relation extraction.
  • Knowledge Graph Completion: Knowledge graph completion adds missing triples through embedding-based ranking, relation-path inference, and logical rule reasoning.Typical subtasks include link prediction, entity prediction, and relation prediction.
  • Embedding-based Ranking: Embedding-based ranking scores every candidate entity after replacing a missing head or tail and returns the top k candidates.The approach learns entity and relation embeddings from existing triples.
  • Relation Path Reasoning: Relation-path reasoning addresses complex multi-step relationships through random walks, sequential decision-making, and reinforcement-learning policies.DeepPath, MINERVA, and CPL differ in rewards, state or text inputs, and policy-network architectures.
  • Rule-based Reasoning: Logical rule learning represents rules as a head implied by a body and mines them for link prediction.AMIE extracts rules, while RLvLR emphasizes scalable search and pruning.

5) Meta Relational Learning:

The survey presents meta relational learning as a response to long-tail and dynamic knowledge graphs, while also reviewing entity-discovery tasks and knowledge-graph-assisted recognition.

  • Meta Relational Learning: Meta relational learning targets unseen relational facts when relations are long-tailed, dynamic, or represented by only a few samples.This setting is also called few-shot relational learning.
  • Meta Relational Learning: GMatching combines entity embeddings, local graph structure, R-GCN, and LSTM-guided multi-step matching for few-shot prediction.Its metric-based design uses structural entity embeddings to compute similarity scores.
  • Entity Discovery: Entity discovery covers recognition, disambiguation, typing, and alignment under different entity-related settings.These tasks acquire entity knowledge from text or fuse knowledge across graphs.
  • Entity Recognition: Knowledge-graph-enhanced pretrained language models such as ERNIE and K-BERT have been applied to named entity recognition with improved performance.The passage identifies these models as combining pretrained language modeling with knowledge graphs.
  • Entity Alignment: Entity alignment finds equivalent entities across two knowledge graphs, often using embedding similarity, transformations, additional attributes, descriptions, or names.BootEA uses incremental training and editing to check newly labeled alignments.

C. Relation Extraction

The survey reviews neural relation extraction under weak supervision, covering convolutional, attention, graph, adversarial, reinforcement-learning, and few-shot approaches, while noting unresolved pipeline-versus-joint-learning questions.

  • Relation Extraction: Relation extraction builds knowledge graphs by extracting unknown relational facts from plain text, often using distant supervision to create training data.Distant supervision assumes sentences containing the same entity mentions may express a database relation.
  • Neural Relation Extraction: Neural relation extraction uses CNNs, multi-instance learning, attention, graph convolution, adversarial training, and reinforcement learning to model text and reduce noise.These methods operate on sentence features, sentence bags, dependency structures, or selected instances.
  • Reinforcement Learning: Reinforcement-learning relation extraction trains policy-based instance selectors to redistribute false positives into negative samples and mitigate noisy data.Its reward can use changes in F1 score, and the extractor is model-agnostic.
  • Few-shot Relation Extraction: Few-shot relation classification addresses low-frequency relations with unseen classes or only a few instances through prototypical and Bayesian meta-learning methods.The reviewed approaches compute relation prototypes or model relationships among relations with a global relation graph.
  • Joint Entity and Relation Extraction: Joint extraction can outperform conventional pipelines, but training–inference distribution discrepancy creates exposure bias and motivates further comparison of both paradigms.A one-stage token-pair-linking formulation is proposed to mitigate error propagation and exposure bias.

V. TEMPORAL KNOWLEDGE GRAPH

Temporal knowledge graphs extend static knowledge graphs with time so representations and completion methods can model when facts hold and how entities and relations evolve. The survey reviews temporal embeddings, temporal dependencies, contextual methods, and logical reasoning over temporal graphs.

  • Motivation: Static knowledge graphs dominate current research, although many facts hold only during specific periods.Temporal information matters because fact validity is constrained by time and fact evolution follows a sequence.
  • Temporal Information Embedding: Temporal-aware embeddings extend triples into quadruples (h, r, t, τ), where τ records when a fact held.TTransE extends TransE for temporal scope prediction over time-annotated triples.
  • Temporal Information Embedding: Temporally scoped models represent valid intervals or timestamp-specific projections, including ConT’s timestamp embeddings and HyTE’s time hyperplanes.HyTE projects entity and relation representations onto a timestamp-specific hyperplane before scoring.
  • Temporal Dynamics: Other approaches model temporal relations through sequence encoders, contextual consistency, tensor completion, entity dynamics, and temporal point processes.Know-Evolve uses a multivariate temporal point process and recurrent network to learn nonlinear temporal evolution.
  • Temporal Relational Dependency: Temporal relational dependency models encode ordering between relations and impose disjointness, ordering, and span constraints.An asymmetric matrix represents temporal ordering between relation pairs, with consistency enforced through integer linear programming.
  • Temporal Reasoning: Logical approaches address uncertain temporal graphs and streaming temporal reasoning through probabilistic logic and temporal close-path rules.These methods learn or apply rule structures for reasoning over temporal knowledge graph data.

B. Question Answering

Knowledge-aware applications use structured knowledge in question answering, recommendation, language understanding, and other downstream tasks. The survey also identifies unresolved challenges in complex reasoning, interpretability, scalability, automatic construction, temporal dynamics, and unified modeling.

  • Question Answering: KG question answering retrieves facts from knowledge graphs, using distributed semantic representations and sometimes symbolic knowledge injection for commonsense reasoning.Simple factoid QA addresses questions involving a single knowledge graph fact.
  • Question Answering: Complex multi-hop QA requires dedicated reasoning designs, with structured knowledge providing commonsense observations and relational inductive biases.Reported approaches include multi-hop bidirectional attention and pointer-generator decoding.
  • Recommendation: Knowledge-graph side information can support recommendation by adding entities, relations, and attributes, with potential to address sparsity and cold-start problems.CKE jointly trains knowledge graph embeddings with textual and visual item information, while DKN uses knowledge-aware convolutional processing.
  • Open Problems: Complex logical reasoning remains an open direction because embedding methods have limitations, motivating relational paths, message passing, reinforcement learning, and logic-embedding combinations.The survey highlights probabilistic inference that captures uncertainty and domain knowledge while retaining efficient embeddings.
  • Open Problems: A unified understanding across knowledge representation, reasoning, knowledge acquisition, and text remains underexplored despite equivalences among several embedding models.Joint learning has been proposed to share information between knowledge graphs and text.
  • Knowledge Aggregation: Neural knowledge aggregation commonly uses attention mechanisms and GNNs, while large-scale language-model pretraining offers another route for injecting factual knowledge.The survey calls for efficient and interpretable knowledge aggregation.
  • Automatic Construction and Dynamics: Automatic construction remains difficult because current knowledge graphs rely heavily on costly manual construction amid multimodal, heterogeneous, and large-scale data.The survey also identifies temporal modeling as a response to the limitations of predominantly static knowledge representation.

APPENDIX A A BRIEF HISTORY OF KNOWLEDGE BASES

The appendix traces knowledge-base history and summarizes foundations used in knowledge representation learning. It covers representation spaces, scoring and composition operations, encoding families, assumptions, objectives, and optimization choices.

  • A Brief History of Knowledge Bases: Knowledge-base history is presented as a brief historical timeline.Figure 10 is identified as illustrating this history.
  • Representation Space: Complex and quaternion spaces extend representation learning beyond ordinary Euclidean vectors through Hermitian and Hamilton products.Quaternion representations use four-dimensional hypercomplex space.
  • Scoring Functions: Hadamard products and circular correlation provide semantic-matching operations, with circular correlation offering efficient computation.These operations are used in semantic matching-based methods.
  • Model Taxonomy: Knowledge representation learning models are organized by representation space, scoring function, encoding model, and auxiliary information.The survey distinguishes point-wise Euclidean, manifold, complex, and Gaussian-distribution spaces.
  • Scoring Functions: Distance-based and semantic-matching scores measure plausibility, spanning translational methods and compositional operators such as bilinear, tensor, circular-correlation, and quaternion products.Examples include TransE, DistMult, HolE, ANALOGY, CrossE, and QuatE.
  • Encoding Models: Encoding models generally fall into bilinear and neural families, with factorization models treating knowledge graphs as three-way tensors.DistMult models only symmetric relations, while extensions address richer relational patterns.
  • Training: Training choices include open- versus closed-world assumptions, negative sampling, margin-based or logistic losses, binary cross-entropy, and regularization.The open-world assumption allows unobserved facts to be missing or false, reflecting graph incompleteness.

C. Negative Sampling

Negative sampling strategies generate corrupted triples to train knowledge graph models, but uniform replacement can produce false negatives. The survey reviews relation-aware and adversarial alternatives and notes a trade-off between accuracy and training time.

  • Uniform sampling corrupts triples by replacing entities uniformly, but may sample false-negative labels.
  • Bernoulli sampling uses relation mapping properties, while domain sampling selects entities from relation-specific domains with a probability.
  • Adversarial methods generate negative triples using probability-based or self-scoring distributions, with self-adversarial sampling controlled by a temperature parameter.
  • The survey summarizes negative sampling strategies in Table VI.
  • The number of negative samples creates a trade-off between predictive accuracy and training time.

C. Medicine and Biology

The survey catalogs knowledge graph applications, datasets, and open-source resources across general, domain-specific, task-specific, and temporal settings. It highlights biomedical use cases while noting that differing ontologies complicate direct dataset comparison.

  • Medicine and Biology: Knowledge-aware models support medicine and biology applications such as protein interaction and drug-target interaction prediction.
  • Applications: Knowledge-driven applications also include academic search, zero-shot image classification, language modeling, question answering, and dialogue systems.
  • Resources: The survey introduces general, domain-specific, task-specific, and temporal datasets, alongside open-source libraries and toolkits for knowledge graph research.
  • General Datasets: General datasets include WordNet, Cyc, DBpedia, YAGO, Freebase, NELL, and Wikidata.
  • General Datasets: Different ontologies make direct comparison of general datasets difficult, so Table VII provides only an informal comparison.
  • Domain-Specific Datasets: Domain-specific knowledge graphs cover life science, health care, scientific research, cultural heritage, and clinical terminology.
  • Task-Specific Datasets: Task-specific datasets are often sampled from larger general datasets, but WN18 and FB15k suffer from test-set leakage.
Loading 2002.00388v4…