PaperlayerView exact revision on arXiv

Source-linked AI summary

Domain-specific Knowledge Graphs: A survey

Bilal Abu-Salih

arXiv:2011.00235v3cs.DBcs.AI

TL;DR

Domain-specific KGs lack a consensual definition, and existing construction approaches have documented deficiencies. This survey defines the concept, reviews research across seven domains, and identifies limitations and future research opportunities. Its synthesis shows KGs being used to integrate heterogeneous data for diverse domain applications.

  • Problem

    Domain-specific KGs lack a consensual definition, while existing construction approaches face limitations requiring comprehensive review.

  • Method

    The survey defines domain-specific KGs and analyzes more than 140 construction studies across seven knowledge domains, including their resources, embeddings, evaluation, and limitations.

  • Results

    The review finds that KGs integrate heterogeneous resources for applications including healthcare question answering, biodata relation discovery, environmental prediction, financial fraud detection, and political-relationship visualization.

  • Takeaways & Limitations

    The survey highlights research gaps, shortcomings, recommendations, and future opportunities for domain-specific KG construction.

  • Takeaways & Limitations

    Evaluation is the predominant weakness among reviewed KG studies, with some using superficial or subjective assessment without concrete metrics.

Abstract

from arXiv · show

Knowledge Graphs (KGs) have made a qualitative leap and effected a real revolution in knowledge representation. This is leveraged by the underlying structure of the KG which underpins a better comprehension, reasoning and interpretation of knowledge for both human and machine. Therefore, KGs continue to be used as the main means of tackling a plethora of real-life problems in various domains. However, there is no consensus in regard to a plausible and inclusive definition of a domain-specific KG. Further, in conjunction with several limitations and deficiencies, various domain-specific KG construction approaches are far from perfect. This survey is the first to offer a comprehensive definition of a domain-specific KG. Also, the paper presents a thorough review of the state-of-the-art approaches drawn from academic works relevant to seven domains of knowledge. An examination of current approaches reveals a range of limitations and deficiencies. At the same time, uncharted territories on the research map are highlighted to tackle extant issues in the literature and point to directions for future research.

1. Introduction

Domain-specific KGs are increasingly used across real-world applications, but their quality, robustness, evaluation, and definition remain unresolved. This survey addresses these gaps by defining the concept and reviewing construction approaches across multiple domains.

  • 1. Introduction: Domain-specific KGs are widely developed because no single schema or ontology fits all real-life problems.Their construction commonly relies on domain-specific ontologies.
  • 1. Introduction: Changing contextual knowledge and inconsistent-quality data sources raise concerns about KG correctness, robustness, and evaluation.Neglecting knowledge dynamics can diminish represented-fact quality and contribute to poor data-based decisions.
  • 1. Introduction: The survey provides an inclusive definition of domain-specific KGs and reviews construction approaches in seven knowledge domains.Its review records construction resources, embedding use, evaluation measures, and reported limitations.
  • 1. Introduction: The paper distinguishes its contribution from prior work by combining an inclusive definition with a comprehensive analysis of domain-based KG construction.It also summarizes issues and suggests improvements and future research opportunities.

2. Methodology

The study systematically reviews recent domain-specific KG construction research using selected academic publications and a defined set of knowledge domains. It examines more than 140 papers and reports that research interest has increased substantially over recent years.

  • 2. Methodology: The review covers healthcare, education, ICT, science and engineering, finance, society and politics, and travel.Articles were collected from relevant conference proceedings and high-quality journals using domain-related keywords.
  • 2. Methodology: More than 140 research articles published between 2016 and 2020 were examined from high-quality computer science and information systems venues.The paper uses Figure 1 to show the number of collected papers by year.
  • 2. Methodology: The review identifies shortcomings in existing approaches and proposes solutions to address them.This distinguishes the study from work focused on generic KGs or brief treatments of domain-specific KGs.

3. Preliminaries

Knowledge graphs represent entities and their relationships as directed, semantically connected structures, while domain-specific KGs focus this representation through domain knowledge and ontologies. Their construction spans extraction, schema design, embedding, and evaluation, with quality and completeness remaining central concerns.

  • Generic Knowledge Graphs: Generic KGs represent real-world entities and relationships as directed graphs, commonly encoding facts as RDF triples.Entities form vertices, relationships form edges, and a fact can be represented as <h, r, t>.
  • Domain-specific Knowledge Graphs: Domain-specific KGs address subject-matter problems but lack a well-established, consensual, and comprehensive definition.The survey positions this definitional gap as a motivation for its proposed formulation.
  • Domain-specific Knowledge Graphs: A domain knowledge graph is an explicit conceptualisation of a high-level subject-matter domain and its subdomains using semantically interrelated entities and relations.The definition combines a domain context, formal conceptualisation, and linked graph content.
  • Knowledge Graph Construction: KG construction extracts entities and relationships through tasks including named entity recognition, disambiguation, linking, and relation extraction.Relation extraction may be local, based on mentions, or global, spanning several local relations.
  • Knowledge Graph Construction: Construction approaches differ by schema use and technical method, including schema-based, schema-free, hybrid, knowledge-based, and learning-based approaches.The survey organizes construction perspectives around ontology design and methods for mining entities and relations.
  • Knowledge Graph Embedding: Knowledge graph embeddings assign vectors to entities and relations and use score functions to capture semantic properties in low-dimensional space.The goal is to represent semantically similar entities and relations with similar vectors while separating disconnected ones.
  • Knowledge Graph Evaluation: Evaluating KG quality requires assessing completeness and correctness because incomplete or inaccurate facts affect fitness for downstream purposes.Domain-specific evaluation is difficult because collecting all true facts for a domain may be impossible.

4. Domain-specific KGs

The survey reviews domain-specific KG research across seven broad domains and examines how KGs are constructed, resourced, embedded, evaluated, and limited. In healthcare, reviewed work spans generic health data, diseases, and healthcare management.

  • Scope of the Review: The review covers healthcare, education, ICT, science and engineering, finance, society and politics, and travel.Appendix tables summarize each domain’s uses, construction algorithms, resources, embeddings, evaluation, and limitations.
  • Healthcare: Healthcare KGs are used to derive insights from voluminous and heterogeneous data, particularly amid heightened attention to the sector.The literature is organized into generic healthcare, diseases, and healthcare management.
  • Healthcare: Generic healthcare studies apply KG approaches to health-data mining and plausible reasoning over fine-grained biomedical ontologies.Examples include process-mining graphs and the SeDan framework for incomplete health data.
  • Healthcare: Disease-focused KGs capture relationships among diseases and symptoms using large electronic medical-record collections and external health knowledge graphs.One reviewed example used 273,174 electronic medical records and the Google Health Knowledge Graph.
  • Healthcare: Healthcare-management studies construct KGs for applications such as healthy-diet knowledge and chronic-disease problem solving.One approach used a domain ontology as the structure of a diet KG and enriched it with entities from healthcare websites.

4.2 Education

Educational KG research applies graph representations to teaching resources, education management, and educational technologies. The reviewed work includes curriculum-based knowledge, course scheduling, prerequisite learning, and academic networks.

  • Education: Educational KGs have expanded with learning-system applications and abundant pedagogical data.The review groups education studies into teaching and classroom resources, education management, and educational technologies.
  • Teaching and classroom resources: K12EduKG represents Chinese K-12 mathematics curriculum concepts and relations extracted using a CRF model and probabilistic methods.Its knowledge source is domain-specific educational data from curriculum standards.
  • Education management: Education-management KGs support course-allocation scheduling and management, while graph-based relational learning addresses concept-prerequisite prediction.The review notes that one course-allocation KG had inadequate evaluation.
  • Educational technologies: Educational-technology work uses KGs to depict academic networks linking research entities and topics.The reviewed literature reports inadequate incorporation of KG embeddings, especially for rich literals.

4.3 ICT

ICT applications use domain-specific KGs for cybersecurity, software development, telecommunications, and the Internet of Things. These studies exploit graph structure to organize complex entities, relations, devices, and operational knowledge.

  • Cybersecurity: Cybersecurity KGs combine domain ontologies with construction techniques and path-ranking methods to obtain new knowledge.A reviewed approach used a quintuple model for knowledge acquisition.
  • Software development: Software-engineering KGs store and manage complex entities and relations and can support functional-requirements work through test-case extraction.The survey presents test-case assistance as an example application.
  • Telecommunication: Telecommunication KGs explicitly model network structures and operator concerns using knowledge from documents, websites, and engineering reports.The reviewed model includes billing, access rights, topology, application hierarchy, and cable-television service components.
  • Internet of Things (IoT): IoT KGs map heterogeneous electronic devices into a new layer that unifies their communications.The reviewed approach uses the KG structure to bridge gaps among IoT devices.

4.4 Sciences and Engineering

Science and engineering applications use domain-specific KGs to represent, integrate, and analyze complex knowledge across chemistry, biology, geology, and engineering. The reviewed approaches span semantic representation, relation discovery, geological information mining, and ecotoxicological prediction.

  • Natural sciences: Semantic web technologies and ontologies provide formal representations that model rich and complex knowledge in the natural sciences.These technologies support semantic analytics and knowledge integration in science-focused applications.
  • Chemistry: Chemistry KGs integrate information from multiple chemical kinetic systems and support automated comprehension of mechanisms and complex semantic queries.They also improve chemical-data sharing and interoperability.
  • Biology: Biology KGs represent molecular interactions and support relational discovery through comparisons of KG embedding models.One example models the heme interactome, while another benchmarks embeddings for relation discovery in molecular biology.
  • Geology: Geology KGs are constructed from sources such as encyclopedias, local documents, dictionaries, and terminology corpora to support geological-data mining and conceptualization.Some approaches use CRF-based geological word segmentation to enrich the graph.
  • Engineering: Engineering applications use domain-specific KGs and embeddings to integrate dissimilar ecotoxicology resources and improve effect prediction.Electric-power research also constructs dispatch KGs using semi-automated labeling and BiLSTM-CRF extraction of entities and behavioral relations.

4.5 Finance

Finance research applies KGs across investment, fraud detection, and related social and political data settings. The reviewed studies use KGs and embeddings to support prediction, representation learning, fraud discovery, and political-relationship analysis.

  • Finance: Finance KGs are used across subdomains including financial investment and fraud detection.The survey presents these areas as important applications of KG technology in finance.
  • Financial investments: Investment studies combine domain-specific KGs with deep learning, stochastic optimization, genetic programming, and KG embeddings for market forecasting and return or stock-price prediction.FinKG learns representations from event-argument relations and lead-lag relations across the graph.
  • Fraud detection: Finance KGs support fraud detection through label propagation and partitioning algorithms that distinguish groups of fraudulent users.The reviewed work argues that fraudulent users tend to be positioned close together, unlike normal users.
  • Social sciences and politics: KGs provide a formal way to conceptualize relationships among people, organizations, and resources in social sciences and politics.This broader relational framing connects finance-related KG applications with social and political knowledge representation.
  • Social science: Social-media KGs address data-quality and entity-identity problems through event clustering, event-driven KG construction, and detection of aliases that conceal entities.The reviewed approaches use hashtag matrices, SKG models, topic modeling, and SVMs.
  • Politics: Political KGs incorporate ontologies, credibility information, and KG embeddings to represent and validate political entities and relationships.One approach extends the BBC politics ontology as a backbone schema.

4.7 Travel

Travel applications use KGs in tourism and transportation to support question answering, recommendations, accommodation-price selection, and traffic-congestion prediction. These graphs combine structured, web, geographic, static, and dynamic urban data.

  • Travel: Travel KG research focuses on tourism and transportation or traffic applications.These subdomains have adopted KGs to address travel-related information and intelligent-system needs.
  • Tourism: Tourism KGs support question answering and recommender systems for personalized attractions and accommodation prices.The reviewed systems target applications including eCommerce.
  • Tourism: The Tirol Tourism KG integrates entities and relations extracted from destination organizations, geographic information systems, and tourism and accommodation websites.This illustrates construction from heterogeneous tourism-related resources.
  • Tourism: A Chinese tourism KG combines encyclopedia knowledge with unstructured web pages and uses Skip-Gram for entity alignment.The approach aligns related entities across the constructed tourism graph.
  • Transportation and traffic: Urban transportation KGs combine static and dynamic raw urban data to support models for predicting traffic congestion.The reviewed approach applies CNNs to model spatio-temporal correlations.

4.8 Summary

The survey reviews KG construction approaches across seven knowledge domains and finds that KGs commonly integrate heterogeneous resources into coherent schemas for practical applications. It also identifies limited-source construction in parts of science and engineering as a continuing gap.

  • Scope: The study reviews recent KG construction research across healthcare, education, ICT, science and engineering, finance, society and politics, and travel.The review is organized around seven knowledge domains.
  • Cross-domain findings: Across domains, KGs help mitigate complexity, provide flexibility, and establish common-ground topologies for integrating information from different sources.The survey identifies information integration as a recurring capability in the selected studies.
  • Cross-domain findings: Heterogeneous resources have been combined into coherent schemas for healthcare QA and recommendation, biodata relation discovery, ecotoxicological prediction, fraud detection, and political-relationship discovery and visualization.These examples span multiple application types and domains.
  • Research gaps: Some science and engineering studies use limited data sources rather than exploiting the heterogeneous-resource integration capability of KGs.The survey presents this as a contrast to the broader cross-domain pattern.

5. Findings from the Survey

The survey identifies recurring limitations in domain-specific KG construction, evaluation, scalability, reasoning, availability, and temporal representation. It also highlights research gaps involving data quality, interoperability, automation, and broader domain coverage.

  • KG construction algorithms: Domain-specific KG construction lacks consolidated methodologies for selecting suitable NLP and data-driven algorithms for entity and relation extraction.Many studies omitted their algorithms or provided limited detail and rationale for their choices.
  • Time-aware KGs: Static KGs often omit temporal validity periods even though entity facts can change over time, potentially harming fact correctness and decision making.The survey notes exceptions such as Wikidata and YAGO, where some facts already include time information.
  • KG evaluation: Evaluation was the predominant weakness, with superficial assessments, limited concrete metrics, and no testbeds or benchmark datasets for comparison.Reviewed approaches used recall, precision, and F-measure for KG completion, while accuracy and AUC were used for KG correctness.
  • Computing performance in Big KGs: Large-scale domain-specific KGs require deeper integration of Big Data infrastructures and sophisticated statistical models for reasoning, which remains open research.Big KGs exhibit volume, variety, velocity, veracity, volatility, and availability characteristics.
  • Domain-specific KG reasoning: KG embedding techniques are not incorporated consistently, despite their reported scalability and efficiency for inferring new relations in large-scale KGs.The survey specifically discusses tensor decomposition, distance, and semantic matching models.
  • Availability of domain-specific KGs: Many domain-specific KGs are not published on the Web, limiting benchmark comparison, reusability, interlinking, and interoperability.The survey also reports shortages of annotated clinical entities, relations, repositories, and expert annotators across several domains.

6. Conclusion

Knowledge graphs are established as a distinctive form of data representation and knowledge inference used across industrial applications. The survey defines domain-specific KGs, reviews construction research across seven domains, and identifies shortcomings and future research opportunities.

  • Knowledge graphs have become a distinctive form of data representation and knowledge inference central to several industrial applications.
  • The survey presents an inclusive definition of domain-specific knowledge graphs and analyzes construction research across seven knowledge domains.
  • The review considers shortcomings and limitations in existing approaches while highlighting future research opportunities.

Appendix A. Tables

The appendix tables organize domain-specific KG studies by use case, construction resources and algorithms, embeddings, evaluation, and reported limitations. Across healthcare, education, ICT, science and engineering, finance, society and politics, and travel, the reported limitations repeatedly concern evaluation, scope, data resources, and methodological justification.

  • Healthcare: Healthcare studies cover management, diet recommendation, process mining, and disease question answering, with limitations in evaluation, representation, data scope, and contextual coverage.
  • Education: Education studies address learning assessment, classroom resources, mathematics, course scheduling, and education management, but often report narrow scope, limited resources, or insufficient evaluation.
  • ICT: ICT approaches include cybersecurity, software development, telecom, and IoT applications, with recurring gaps in benchmark comparison, construction rationale, evaluation, and data coverage.
  • Science and Engineering: Science and engineering studies report applications such as electrical engineering and oil-and-gas search, while noting limited sources, evaluation measures, structure, or application scope.
  • Finance: Finance studies include investment and stock prediction applications, with reported concerns about evaluation strategy, metrics, embeddings, benchmark rationale, and data sources.
  • Society and Politics; Travel: Society, politics, and travel studies use varied extraction, embedding, and recommendation methods, but report limitations involving data sources, metrics, generalization, usability, and named-entity extraction.
Loading 2011.00235v3…