Source-linked AI summary
Knowledge Graphs: Opportunities and Challenges
Ciyuan Peng, Feng Xia, Mehdi Naseriparsa, Francesco Osborne
TL;DR
The paper addresses how to understand knowledge graphs as AI and big-data systems that organize complex real-world knowledge. It surveys their opportunities in AI systems and application fields, then synthesizes technical challenges across five knowledge-graph technologies. The survey concludes that knowledge graphs support intelligent services and applications while current technologies retain serious limitations requiring future research.
Problem
AI systems need organized knowledge to handle complex realistic tasks, while knowledge graphs still face limitations in acquiring and integrating knowledge from multiple sources.
Method
The paper conducts a systematic survey of knowledge-graph opportunities in AI systems and application fields and challenges in five representative technologies.
Results
The survey identifies knowledge graphs as supporting recommender systems, question answering, information retrieval, and applications across multiple fields.
Takeaways & Limitations
Knowledge graphs provide a basis for intelligent services and applications while motivating further research and development on their technologies.
Takeaways & Limitations
Existing knowledge-graph technologies remain limited by low-accuracy knowledge acquisition and difficulty using additional information effectively in embeddings.
Abstract
from arXiv · showhide
With the explosive growth of artificial intelligence (AI) and big data, it has become vitally important to organize and represent the enormous volume of knowledge appropriately. As graph data, knowledge graphs accumulate and convey knowledge of the real world. It has been well-recognized that knowledge graphs effectively represent complex information; hence, they rapidly gain the attention of academia and industry in recent years. Thus to develop a deeper understanding of knowledge graphs, this paper presents a systematic overview of this field. Specifically, we focus on the opportunities and challenges of knowledge graphs. We first review the opportunities of knowledge graphs in terms of two aspects: (1) AI systems built upon knowledge graphs; (2) potential application fields of knowledge graphs. Then, we thoroughly discuss severe technical challenges in this field, such as knowledge graph embeddings, knowledge acquisition, knowledge graph completion, knowledge fusion, and knowledge reasoning. We expect that this survey will shed new light on future research and the development of knowledge graphs.
1 Introduction
Knowledge graphs organize real-world knowledge in machine-readable form and support AI systems and applications across society. This survey reviews those opportunities alongside technical challenges in acquiring, integrating, completing, and reasoning over knowledge graphs.
- Opportunities and challenges: Knowledge graphs represent entities and their real-world relations using formal semantics that computers can process efficiently and unambiguously.They are defined as graphs of data that accumulate and convey knowledge of the real world.
- Opportunities and challenges: Knowledge graphs support AI systems such as recommender systems, question answering, and information retrieval, while also serving fields including education and medical care.Their use is associated with processing heterogeneous information in machine-readable contexts.
- Opportunities and challenges: Knowledge graphs offer opportunities by improving AI-system quality and reaching diverse application areas, but their development still faces significant technical challenges.A highlighted example is the difficulty of acquiring and integrating knowledge from multiple sources into a typical graph.
- Survey scope: The survey examines opportunities through knowledge-graph-based AI systems and application fields, then analyzes limitations of knowledge-graph technologies.The paper frames its analysis around systems whose performance is improved by knowledge graphs and fields that benefit from them.
- Survey scope: It specifically reviews recommender systems, question-answering systems, and information retrieval, plus applications in education, scientific research, social media, and medical care.The survey presents these areas as representative opportunities and five technologies as representative challenges.
2 Overview
Knowledge graphs represent entities and semantic relations as directed, machine-readable graph structures, commonly organized around subject–predicate–object triples. Research spans AI systems, application fields, and five core technologies whose limitations define major challenges.
- 2.1 What are Knowledge Graphs?: Knowledge-graph construction commonly includes ontology construction, which defines properties of a specific domain and how they are related.The paper distinguishes a domain schema as an ontology and identifies ontology construction as an essential construction stage.
- 2.1 What are Knowledge Graphs?: A knowledge graph represents real-world facts and semantic relations as a directed graph whose nodes are entities and whose edges convey relations.Its fundamental unit is a triple such as (Bill Gates, founderOf, Microsoft).
- 2.2 Current Research on Knowledge Graphs: The survey categorizes current research into AI systems, application fields, and five technologies: embedding, acquisition, completion, fusion, and reasoning.AI systems and application fields represent opportunities, while the five technologies correspond to major technical challenges.
- 2.2 Current Research on Knowledge Graphs: Knowledge graph embedding maps entities and relations into a low-dimensional vector space to capture graph semantics and structure for machine-learning models.The survey identifies embedding as a central research issue.
- 2.2 Current Research on Knowledge Graphs: Knowledge acquisition constructs graphs from structured and unstructured sources, while knowledge graph completion adds plausible new triplets through link and entity prediction.Completion models assess candidate triplets and add those with high plausibility.
- 2.2 Current Research on Knowledge Graphs: Knowledge fusion integrates knowledge from different sources, and knowledge reasoning infers new facts or identifies erroneous knowledge.The survey presents entity alignment as a primary fusion method and describes reasoning as forming new triplets or detecting false facts.
- 2.2 Current Research on Knowledge Graphs: Knowledge graphs underpin AI services such as recommenders, question answering, and information retrieval, and they are applied in education, scientific research, social media, and medical care.The survey links these uses to the richness of information in knowledge graphs and to diverse application needs.
3 Knowledge Graphs for AI Systems
Knowledge graphs support AI systems by representing and linking entities, enabling recommendation, question answering, and information retrieval. The survey describes improved representation, efficiency, accuracy, explainability, and multi-hop answering, while noting persistent limitations in traditional recommender systems.
- Knowledge graphs are used as input data across recommender systems, question-answering systems, and information retrieval tools.
- Knowledge Graph-based Recommender Systems: Knowledge graph-based recommenders connect users, films, directors, actors, and genres to infer latent user–movie relations.
- Knowledge Graph-based Recommender Systems: Knowledge graph-based recommenders alleviate data sparsity and cold-start issues while making recommendation reasoning easier to explain.New-item information can be obtained through entity relations, such as linking a new film to its genre.
- Question-answering Systems: Knowledge graph-based question answering reduces search space by focusing on entities with relevant properties and supports answers that combine multiple facts through multi-hop reasoning.
- Information Retrieval: Knowledge graph-based information retrieval uses formal, interlinked item representations and entity relations to improve search efficiency and retrieval accuracy.Semantic similarity, reasoning, and query expansion also make retrieved items more interpretable.
4 Applications and Potentials
Knowledge graphs are applied across education, scientific research, social networks, and health/medical care, supporting data processing, knowledge exploration, recommendations, and misinformation detection. The survey presents these domains as major application opportunities while noting that other areas such as finance remain less explored.
- The survey focuses on knowledge graph applications in education, scientific research, social networks, and health/medical care.It notes that knowledge graph-based services in other domains, including finance, remain relatively obscure.
- Education: In education, knowledge graphs structure complex data and support both offline teaching and online learning.Applications process educational data, disseminate knowledge, and can filter social-media content for formal learning.
- Scientific Research: Scientific knowledge graphs interlink publications, researchers, organizations, topics, technologies, projects, and funding to support research exploration.Examples include large heterogeneous graphs such as MAG, AMiner Graph, OAG, and AIDA KG.
- Social Networks: Social knowledge graphs analyze user activities and posts for recommendation, social relationship extraction, and fake news detection.These applications address personalized content or friend matching, relationship discovery, and misinformation analysis.
- Health/Medical Care: Medical knowledge graphs support medical recommendation, health misinformation detection, and drug discovery.They integrate rapidly growing medical information and can incorporate clinicians’ expertise when building medical knowledge graphs.
5 Technical Challenges
Knowledge graphs offer broad opportunities, but their development is constrained by limitations in existing technologies. The survey organizes these challenges around five topical areas, including knowledge graph embeddings.
- Existing knowledge graph technologies have limitations that create challenges for advancing knowledge graphs.The survey frames these limitations as the key challenges for further development.
5.1 Knowledge Graph Embeddings
Knowledge graph embeddings map entities and relations into low-dimensional vectors while preserving graph semantics, using tensor factorization, translation-based, and neural network-based approaches. Existing methods remain limited by their treatment of additional information and complex relation paths.
- Embedding objective: Knowledge graph embeddings represent entities and relations in a low-dimensional vector space while preserving semantic and structural information.Embedding methods define scores for facts and optimize their plausibility to learn entity and relation vectors.
- Tensor factorization: Tensor factorization methods transform knowledge graph triplets into a 3D tensor whose relation-specific slices encode whether each fact exists.Tensor entries are 1 for existing triplets and 0 otherwise.
- Translation-based methods: Translation-based methods connect head and tail entity vectors through relation vectors, with TransE preserving triplet semantics in a continuous vector space.TransH and TransR extend TransE to improve translation-based embedding performance.
- Neural network-based methods: Neural network-based approaches include SME for semantic matching, ConvKB for convolutional feature extraction, and R-GCN for relation-specific graph transformations.ConvKB scores triplets after applying convolution to a three-row triplet matrix, while R-GCN uses relation-specific transformations during propagation.
- Limitations of existing methods: Effectively incorporating rich multivariate information and representing complex indirect relations remain open challenges for knowledge graph embeddings.The paper specifically identifies hierarchical relation descriptions and combinations of entity types with textual descriptions as promising feature additions.
5.2 Knowledge Acquisition
Knowledge acquisition combines structured and unstructured data to build knowledge graphs, but current approaches face reliability, domain adaptation, multilingual, and multimodal challenges.
- Knowledge acquisition methods: Knowledge acquisition extracts relations, entities, and attributes from structured and unstructured data to generate knowledge graphs.Attribute extraction can be regarded as a special case of entity extraction.
- Reliability and domain specificity: Low acquisition accuracy can produce incomplete or noisy knowledge graphs and hinder downstream tasks.The section identifies tool reliability and evaluation as a critical issue.
- Reliability and domain specificity: Efficient acquisition of domain-specific knowledge graphs remains difficult because domain schemas are knowledge-oriented while constructed schemas cover raw-data features.The paper highlights generating domain-specific graphs as an essential issue.
- Multilingual acquisition: Most existing methods focus on one language, making cross-lingual extraction and multilingual knowledge graph generation important but difficult.Challenges include limited non-English training data, inaccurate translation systems, and the need to retrain cross-lingual extraction models.
- Multimodal acquisition: Multi-modal knowledge graph construction must integrate entities represented in different modalities rather than relying only on pure symbols.The passage motivates multimodal graphs by noting that symbolic representations can limit machine understanding of the real world.
5.3 Knowledge Graph Completion
Knowledge graph completion expands incomplete graphs by predicting plausible missing triplets, but current methods are constrained by closed-world assumptions, low open-world accuracy, and static-graph assumptions.
- Completion motivation: Knowledge graphs are often incomplete, with missing relevant triplets and entities.In Freebase, more than half of person entities lack birthplace and parent information.
- Completion methods: Knowledge graph completion trains models to score candidate triplets and adds those with high plausibility using link and entity prediction.For an incomplete triplet, the method predicts a missing element and evaluates candidate completions.
- Open-world completion: Closed-world completion methods can generate new triplets only when their entities or relations already exist in the knowledge graph.They therefore cannot add new entities and relations.
- Open-world completion: Open-world completion techniques retrieve potential objects outside existing knowledge bases, but they still suffer from low accuracy.ConMask is cited as a model for predicting unseen entities.
- Temporal completion: Static completion methods fail to capture knowledge graph evolution, motivating temporal methods that integrate timestamps into learning.Temporal completion explores time-sensitive facts and improves the handling of knowledge validity over time.
5.4 Knowledge Fusion
Knowledge fusion integrates knowledge from different sources, primarily through entity or ontology alignment, but cross-language, disambiguation, and multimodal settings remain challenging.
- Fusion foundations: Knowledge fusion combines knowledge from different data sources and is often necessary for generating knowledge graphs.Entity or ontology alignment is described as the primary fusion method.
- Cross-language fusion: Cross-language knowledge fusion integrates information across languages and can support cross-lingual recommender systems.Graph-matching neural networks are cited for cross-language entity alignment.
- Entity disambiguation: Entity disambiguation is required before alignment because polysemy gives the same entity different expressions across knowledge graphs.Existing methods mainly distinguish ambiguous entities using contextual information from texts.
- Multimodal fusion: Single-modality alignment performs poorly in multimodal scenes because it cannot fully reflect real-world entity relationships.Recent work therefore investigates multimodal knowledge fusion.
5.5 Knowledge Reasoning
Knowledge reasoning infers new triplets from existing graph data through single-hop prediction or multi-hop reasoning, while verification remains necessary because inferred knowledge can be uncertain or conflicting.
- Reasoning task: Knowledge reasoning infers implicit relations between entities and forms new triplets from existing graph data.For unconnected entities h and t, reasoning predicts a potential relation r.
- Reasoning methods: Neural network-based reasoning methods infer new triplets from the body of knowledge in a graph.The passage identifies neural networks as one approach to knowledge reasoning.
- Reasoning tasks: Single-hop prediction fills one triplet element, whereas multi-hop reasoning predicts elements across multi-hop logical queries.Multi-hop answers may require predicting and imputing multiple edges and nodes.
- Verification: Verification must assess inferred triplets because new knowledge can be uncertain and may conflict with existing knowledge.The section also points to multi-source reasoning as one approach for addressing these problems.
6 Conclusion
The survey presents knowledge graphs as instrumental to intelligent services and applications across fields while organizing their opportunities and technical challenges. It aims to provide perspectives that support future research and development.
- Knowledge graphs have supported many intelligent services and applications across various fields.
- The survey introduces knowledge-graph definitions and research directions before examining AI systems and representative applications.
- The survey analyzes limitations of current knowledge-graph technologies as severe technical challenges.
- The authors expect the survey to spark ideas and perspectives for future knowledge-graph research and development.