PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey on Automatic Knowledge Graph Construction

Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, Xindong Wu

arXiv:2302.05019v1cs.IR

TL;DR

Automatic knowledge graph construction needs methods that move beyond factual extraction to conceptualized knowledge in increasingly sophisticated scenarios. This survey reviews more than 300 methods across acquisition, refinement, and evolution, alongside practical resources and open challenges. It concludes that conditional, dynamic, long-context, federated, and cross-lingual knowledge remain important areas for further work.

  • Problem

    Existing knowledge graph construction has expanded beyond factual extraction, creating a need to systematically organize methods for conceptualized knowledge and sophisticated scenarios.

  • Method

    The survey analyzes more than 300 knowledge graph construction methods across knowledge acquisition, refinement, and evolution, and organizes them by data environment, motivation, and architecture.

  • Results

    The survey systematically presents representative methods for fine-grained entity typing, few-shot extraction, document-level relation extraction, logic reasoning, and conditional knowledge graphs.

  • Takeaways & Limitations

    The survey provides practical knowledge graph resources and identifies challenges involving long contexts, privacy-preserving federation, evolving semantics, and cross-lingual low-resource knowledge.

  • Takeaways & Limitations

    Knowledge semantics under changing associated conditions remains unexplored, limiting understanding of how conditions affect related facts.

Abstract

from arXiv · show

Automatic knowledge graph construction aims to manufacture structured human knowledge. To this end, much effort has historically been spent extracting informative fact patterns from different data sources. However, more recently, research interest has shifted to acquiring conceptualized structured knowledge beyond informative data. In addition, researchers have also been exploring new ways of handling sophisticated construction tasks in diversified scenarios. Thus, there is a demand for a systematic review of paradigms to organize knowledge structures beyond data-level mentions. To meet this demand, we comprehensively survey more than 300 methods to summarize the latest developments in knowledge graph construction. A knowledge graph is built in three steps: knowledge acquisition, knowledge refinement, and knowledge evolution. The processes of knowledge acquisition are reviewed in detail, including obtaining entities with fine-grained types and their conceptual linkages to knowledge graphs; resolving coreferences; and extracting entity relationships in complex scenarios. The survey covers models for knowledge refinement, including knowledge graph completion, and knowledge fusion. Methods to handle knowledge evolution are also systematically presented, including condition knowledge acquisition, condition knowledge graph completion, and knowledge dynamic. We present the paradigms to compare the distinction among these methods along the axis of the data environment, motivation, and architecture. Additionally, we also provide briefs on accessible resources that can help readers to develop practical knowledge graph systems. The survey concludes with discussions on the challenges and possible directions for future exploration.

1 INTRODUCTION

Knowledge graph construction has progressed from extracting factual mentions toward organizing conceptualized knowledge across acquisition, refinement, and evolution. This survey reviews methods, resources, and challenges across diverse data environments and construction stages.

  • Knowledge Graph Construction: Knowledge graphs organize real-world knowledge as nodes and relationship edges, with knowledge tuples forming their minimum knowledge-carrying units.The survey distinguishes entities, relationships, and the tuples connecting them as the basic elements discovered during construction.
  • Knowledge Graph Construction: Early systems extracted factual tuples from semi-structured or unstructured text but struggled to distinguish entities across expressions and lacked informative background knowledge.These limitations hindered knowledge aggregation and restricted extraction beyond patterned mentions.
  • Knowledge Acquisition: A classic acquisition pipeline discovers and links entities, resolves coreferences, and extracts relationships among entities.This pipeline organizes semantic knowledge acquisition into partitioned subtasks.
  • Knowledge Acquisition and Refinement: Deep learning has improved major acquisition tasks and enabled refinement operations such as tuple completion, graph-structure-based tuple discovery, and graph merging.The reviewed tasks include named entity recognition, entity typing, entity linking, coreference resolution, and relation extraction.
  • Challenges and Environments: Recent KG construction addresses heterogeneous, complex, noisy, low-resource, temporal, interpretable, and autonomous data settings.The survey frames these challenges through HACE environments and discusses corresponding construction, refinement, and evolution methods.
  • Survey Contributions: The survey goes beyond prior overviews by organizing recent models across construction stages and HACE aspects while providing practical resources and future directions.Its contributions include formal definitions, classifications, model analyses, practical KG projects and tools, and discussions of challenges affecting KG construction.

2 BACKGROUND

This section defines knowledge graphs and construction, surveys representative KG project types, and introduces practical construction resources and tools.

  • Definitions: A knowledge graph comprises concepts, relations, factual triples, and background knowledge that constrains potential facts.
  • Definitions: Knowledge graph construction maps data sources and their background knowledge into a knowledge graph.
  • Knowledge Graph Projects: The survey organizes practical KG projects into encyclopedia, linguistic, commonsense, enterprise, domain-specific, and federated categories.
  • Knowledge Graph Projects: Encyclopedia KGs cover factual or event knowledge across domains, with examples including DBpedia, Freebase, Wikidata, YAGO, and multilingual XLore.
  • Knowledge Graph Projects: Linguistic KGs provide semantic ontologies or external features, with WordNet supplying synonymy and hyponymy relations and BabelNet extending them cross-lingually.
  • Construction Tools: Construction tools support preprocessing, knowledge acquisition, and refinement, including web extraction, information extraction, NLP, table-mining, neural, and end-to-end toolkits.

3 SEMI-STRUCTURED DATA PRE-PROCESSING

Semi-structured preprocessing removes irrelevant content and interprets tables before knowledge acquisition, producing cleaner documents and filtering non-relational structures.

  • Overview: Preprocessing handles messy multi-structure data through content extraction and structure interpretation.
  • Content Extraction: Content extraction removes webpage noise such as advertisements while retaining knowledge-bearing content.
  • Content Extraction: Wrapper-based methods capture informative content with matching rules, generated templates, or bootstrapping from seed examples.
  • Content Extraction: Statistical extractors identify informative webpage regions using text length, tag counts, text density, and DOM-tree path ratios.
  • Content Extraction: After content extraction, semi-structured pages become renewed noise-free semi-structured or unstructured documents.
  • Structure Interpretation: Relational table interpretation classifies whether tables contain informative relational structures, excluding decorative containers.

4 KNOWLEDGE ACQUISITION

Knowledge acquisition collects entities and relationships from multi-structured data through entity recognition, coreference resolution, and relation extraction.

  • Knowledge Acquisition: Knowledge acquisition discovers entity mentions, resolves referred mention pairs, and links entities through semantic relationships.

4.1 Entity Discovery

Entity discovery identifies mentions, assigns increasingly fine-grained types, and links them to KG nodes, using rule-based, statistical, neural, and embedding-based methods.

  • Entity Discovery: Entity discovery consists of named entity recognition, entity typing, and entity linking, progressing from mention detection to specific types and KG nodes.
  • Named Entity Recognition: NER tags entity positions and classifications in semi-structured or unstructured data, whose available structural hints differ.
  • Named Entity Recognition: NER methods include wrappers and rules for semi-structured data, sequential statistical models, and deep sequence-to-sequence architectures for text.
  • Named Entity Recognition: Deep NER architectures use CNNs for local features, RNNs for contextual features, attention for salient interactions, GCNs for linguistic graphs, and pretrained language models.
  • Entity Typing: Entity typing supplies fine-grained and ultra-grained categories, addressing information loss while managing infrequent and overly specific types.
  • Entity Typing: Embedding-based entity typing combines graph structure and background knowledge, including heterogeneous entity-type, text-feature, ontology, and instance graphs.
  • Entity Linking: Entity linking disambiguates mentions such as “Tesla” by connecting them to corresponding KG objects using contextual and structural hints.
  • Entity Typing: Few- and zero-shot entity typing remains challenging, with prototype, hierarchical-label, contextual, word-embedding, and memory-network approaches proposed.

4.2 Coreference Resolution

Coreference resolution identifies mentions that refer to the same entity, including aliases and pronouns, and outputs co-referred spans. Its methods range from statistical clustering and graph partitioning to deep models using contextual, recurrent, attention, and embedding-based features.

  • Task definition: Coreference resolution detects mentions referring to the same entities, including aliases and pronouns, and identifies singleton mentions.The task outputs pairs of co-referred word spans from unstructured sentences.
  • Classical models: Early approaches used statistical features of entities, mentions, and antecedents, while cluster-based methods formulated coreference as pairwise binary classification.Single-link clustering and mention-pair models were used to build coreference chains or singleton leaves.
  • Classical models: Tree-based and graph-based models convert coreference resolution into a partition task by constructing hypergraphs whose edges can link multiple mentions.Statistical edge weighting, clustering, and relaxed labeling support coreference partitions.
  • Deep learning models: Deep models learn document representations and use CNNs, RNNs, LSTMs, and attention mechanisms to capture contextual and span-level interactions.Examples include end-to-end LSTM models, biaffine attention, and mutual attention for linked-expression detection.
  • Deep learning models: Embedding-based distribution models represent antecedents, mention pairs, and tree structures to support coreference inference and feature optimization.These approaches use distributional semantics and scalable processing to improve coreference representations.

4.3 Relation Extraction

Relation extraction links entities by extracting factual relations from unstructured or semi-structured data. The survey covers open extraction, relation classification, n-ary relations, distant supervision, noisy labels, external knowledge, and low-resource settings.

  • Task definition: Relation extraction extracts relational facts that connect entities, with binary methods handling entity pairs and n-ary methods covering multiple entities.The resulting semantic links endow knowledge graphs with relations; n-ary links can be represented as super-edges over multiple nodes.
  • Open relation extraction: Open relation extraction discovers subject–predicate–object triples from free text without predefined relation types.Methods detect nominal subjects or objects and verbal predicates, while statistical, rule-based, semi-supervised, and neural approaches provide alternative designs.
  • Relation classification: Domain-specific relation classification assigns mentions relation tags from a predefined relation set using kernel methods or deep learning frameworks.CNN, LSTM, GCN, and attention architectures are among the surveyed model designs.
  • Relation classification: Semantic role labeling can decompose n-ary relations into binary relations for models that must represent relationships among multiple entities.NNF and dependency path embedding are examples of this strategy.
  • Distant supervision: Distant supervision reduces manual annotation by labeling relation instances with an external knowledge base, but its entity-pair assumption introduces noisy labels.Methods address noise through feature extraction, learning strategies, instance selection, and multi-instance multi-label hypotheses.
  • Advanced scenarios: External entity knowledge and syntactic graphs are incorporated into deep models, while few-shot and zero-shot methods target relation classification with limited examples.Knowledge graph embeddings, side information, and meta-learning support these settings.

5 KNOWLEDGE GRAPH REFINEMENT FROM STRUCTURED DATA

Knowledge graph refinement addresses sparsity, incompleteness, and corruption in graphs built from unstructured or semi-structured data. It uses background semantics or additional structured knowledge and includes completion and fusion.

  • Refinement overview: Knowledge graph refinement repairs sparse, incomplete, or corrupted triples by using background semantics or additional structured knowledge graphs.Its sub-tasks include knowledge graph completion and knowledge fusion.

5.1 Knowledge Graph Completion

Knowledge graph completion fills missing triples, derives new facts, and evaluates triple correctness. The surveyed approaches use embeddings, path reasoning, reinforcement learning, neural networks, rule mining, and logic-enhanced models, while negative-sampling limitations remain important.

  • Completion objectives: Knowledge graph completion fills incomplete triples and derives new triples from completed ones, while triple classification evaluates their correctness.Triple classification distinguishes reliable triples from abnormal or untrue triples.
  • Embedding-based completion: Embedding-based link prediction searches for missing entities or relations using distributional representations, exemplified by TransE’s h + r = t formulation.Later models extend embeddings to represent semantic hierarchies, common-sense rules, and improved negative sampling.
  • Path reasoning: Relation path reasoning infers new facts from sequences of completed triples that provide supporting evidence.Approaches include random-walk models, semantic enrichment, Markov decision processes, deep reinforcement learning, recurrent networks, and attention.
  • Rule-based reasoning: Rule-mining methods extract and prune logical rules from knowledge graphs, then use rule templates for link prediction, but unseen paths are difficult to derive.Other models combine first-order or soft logic rules with neural embeddings and relation-path reasoning.
  • Interpretability: Interpretability methods explain completion models either through transparent rule-based reasoning or post-hoc proxies for black-box embeddings and neural networks.Examples include learned rules, Bayesian networks with first-order logic, and plausible explanation paths.
  • Triple classification: Insufficient negative sampling can leave many reasonable triples uncovered, particularly for one-to-many relations.Proposed remedies add fine-grained type chains, ontological information, or memory-based embeddings to improve triple classification.

5.2 Knowledge Fusion

Knowledge fusion integrates new knowledge by aligning attributes and matching or aligning entities across knowledge graphs. Methods range from linguistic and embedding-based models to graph-based, neural, and active-learning approaches.

  • Fusion tasks: Knowledge fusion merges semantically equivalent elements and integrates new concepts or facts into existing knowledge graphs.Its sub-tasks include attribute alignment, entity matching for small incoming triple sets, and entity alignment for complete knowledge graphs.
  • Attribute alignment: Attribute alignment unifies different notations for the same property, such as “birthday” and “date of birth.”Embedding-based methods use attribute names, contextual values, or distributional representations to estimate equivalence.
  • Entity matching: Entity matching combines linguistic, numerical, textual, graph, and neural similarity signals to identify equivalent entities in small-scale knowledge bases.Examples include random-forest models, contextual neural representations, description embeddings, graph features, and fuzzy clustering.
  • Entity matching: Active learning queries users to resolve unique missing information and conflicts that autonomous communities contribute to knowledge graphs.This is particularly relevant for personalized data such as nicknames and telephone numbers.
  • Entity alignment: Large-scale entity alignment learns mappings using seed entities, augmented attributes, semantic graph structures, and cross-graph interaction.Representative approaches use adversarial negative sampling, co-training, GCNs, graph matching, heterogeneous attribute partitions, and cross-graph attention.

6 KNOWLEDGE EVOLUTION

Knowledge evolution addresses knowledge that changes with environmental conditions through conditional knowledge extraction, conditional graph completion, and dynamic modeling. The survey highlights temporal methods while identifying general multi-condition reasoning and changing-condition dynamics as open challenges.

  • 6.1 Conditional knowledge graphs: Conditional knowledge graphs represent facts established under prerequisites, with temporal knowledge graphs as a simplified timestamp-based case.A conditional tuple is formulated as (h, r, t, γ), where γ can be a prerequisite triple or temporal information.
  • 6.1 Conditional knowledge acquisition: Conditional knowledge extraction targets fact tuples, condition tuples, and links between facts and their conditions.The MIMO model separates factual and conditional tagging before completing tuples, while later work addresses overlapping triples through hierarchical parsing.
  • 6.1 Conditional knowledge acquisition: Temporal extraction models use recurrent structures, stacked attention, and event-state-duration representations to identify temporal relationships and fine-grained timelines.These methods treat temporal information as the condition attached to extracted knowledge.
  • 6.2 Condition knowledge graph completion: Condition knowledge graph completion fills missing entities, relations, or conditions, with the survey focusing mainly on temporal knowledge graph completion.Approaches include temporal embeddings, temporal-aware sequence or context models, tensor decomposition, and graph-based temporal reasoning.
  • 6.2 Condition knowledge graph completion: General conditional completion remains open when tuples contain multiple conditions, such as chemical reactions restricted to a temperature range.The survey points to causality discovery as a related direction for these complex scenarios.
  • 6.3 Knowledge Dynamic: How knowledge evolves as different conditions change remains challenging, including conditions associated with ending the COVID-19 outbreak.The survey recommends combining causal feature selection with expert knowledge and multi-source evidence.

7 KNOWLEDGE GRAPH STORAGE

Knowledge graph storage tools span relational, key/value, and graph-database paradigms for different data environments. The surveyed systems support operations ranging from CRUD and distributed storage to graph queries and parallel computation.

  • Relational storage: Relational models provided early graph storage, using RDBMS reliability and graph algorithms such as depth-first traversal and shortest-path search.PostgreSQL and Filament are representative examples, although sparse or personalized knowledge graphs can be costly for relational databases.
  • Key/value storage: Key/value databases offer lightweight, flexible, distributed storage for large knowledge graphs.Trinity targets billion-node graphs in memory, CouchDB supports replication for dynamic graphs, and MapReduce transforms data groups into key/value mappings.
  • Key/value storage: Pregel supports parallel graph computation through a superstep mechanism that shares messages between vertices.It is presented alongside MapReduce-based graph storage approaches.
  • Graph databases: Graph databases are designed around knowledge-triple structures and support embedded dynamic storage, object-oriented queries, or semantic-web data models.Examples include Neo4j, SOnes, RDF, and OWL-based systems.

8 DISCUSSION ON KNOWLEDGE GRAPH CONSTRUCTION

The discussion identifies unresolved challenges in constructing knowledge graphs from complex contexts, private and multimodal data, evolving conditions, human feedback, and multiple languages. It highlights future directions including advanced semantic interpretation, privacy-preserving federation, human–machine collaboration, and integrated acquisition-refinement models.

  • Long and Intricate Contexts: Cross-sentence and cross-paragraph contexts require reasoning over pattern recognition, coreference, logic, and commonsense to capture high-order semantics.The survey notes that document-level models must handle intricate patterns across multiple linguistic objects.
  • Long and Intricate Contexts: Commonsense-derived facts can be inaccurate, while many document-level datasets lack evidence for validating correct logic paths.Temporal and geographical conditions also need consideration for rigorous contextual interpretation.
  • Federated Learning: Federated KG learning avoids data exchange but still faces an entity-alignment bottleneck because sharing multi-source KGs can expose sensitive information.The proposed direction is an encrypted entity-alignment space that preserves privacy while federating features.
  • Advanced Semantic and Dynamic Knowledge: Knowledge semantics remain underexplored when general associated conditions evolve beyond temporal changes in knowledge graphs.The survey points to studying how conditions affect related facts, such as professional networks during pandemic phases.
  • Human-machine Synergy: HAO-based active learning assigns uncertain samples to authorities, domain users, or AI systems, supporting collaborative completion and correction of unknown facts.The approach is presented as a promising way to obtain open-world knowledge through role-based human–machine cooperation.
  • Cross-lingual Knowledge Graph: Cross-lingual KG construction remains limited by translation errors, conflicts, and insufficient resources for minority languages.The survey identifies accurate low-resource translation and cross-lingual conflict resolution as promising directions.
  • Integrated Construction: Future work should unify knowledge acquisition and refinement within end-to-end frameworks rather than treating them as separate tasks.The survey describes this integration as a formidable bottleneck despite progress in unified extraction models.

9 CONCLUSION

The survey consolidates methods, challenges, and resources for constructing, refining, and integrating knowledge graphs across data types and scenarios. It identifies future directions for knowledge graphs operating in massive, heterogeneous, autonomous, complex, and evolving data environments.

  • The survey reviews tasks, methods, challenges, and resources for constructing, refining, and integrating knowledge graphs from varied data types and scenarios.
  • It presents models for fine-grained entity typing, few-shot extraction, document-level relation extraction, logic and interpretable reasoning, and conditional knowledge graph structures.
  • The paper provides briefs on practical knowledge graph toolkits and projects.
  • Future research is directed toward empowering knowledge graphs in massive heterogeneous, autonomous, complex, and evolving data environments.
  • The survey also calls for enhanced collaboration between knowledge communities.
Loading 2302.05019v1…