Source-linked AI summary
Large Language Models and Knowledge Graphs: Opportunities and Challenges
Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russa Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, Damien Graux
TL;DR
The paper examines how explicit knowledge in Knowledge Graphs and parametric knowledge in LLMs can be combined amid unresolved reasoning and reliability concerns. It synthesizes opportunities and challenges across both directions, reporting that KG grounding improves generation and question answering while long-tail knowledge remains difficult for LLMs.
Problem
The paper addresses how to combine explicit Knowledge Graph knowledge with parametric LLM knowledge despite unresolved reasoning, hallucination, and knowledge-coverage concerns.
Method
The paper synthesizes debates, research directions, opportunities, and challenges for using LLMs with Knowledge Graphs in both directions.
Results
KG grounding improves performance on generation and question-answering tasks, while LLMs struggle to robustly recall long-tail entities and facts.
Takeaways & Limitations
The paper supports hybrid LLM–KG research while identifying knowledge recall, extraction, prompting, and ontology engineering as continuing research challenges.
Takeaways & Limitations
Entity disambiguation remains challenging when extracting facts from LLMs because token-level operation complicates resolving polysemy, homonymy, and coreference.
Abstract
from arXiv · showhide
Large Language Models (LLMs) have taken Knowledge Representation -- and the world -- by storm. This inflection point marks a shift from explicit knowledge representation to a renewed focus on the hybrid representation of both explicit knowledge and parametric knowledge. In this position paper, we will discuss some of the common debate points within the community on LLMs (parametric knowledge) and Knowledge Graphs (explicit knowledge) and speculate on opportunities and visions that the renewed focus brings, as well as related research topics and challenges.
1 Introduction
LLMs have shifted Knowledge Representation toward hybrid representations that combine explicit knowledge with knowledge encoded in model parameters. This paper surveys the resulting integration of LLMs and knowledge graphs, identifies community debates, and presents opportunities and challenges.
- LLMs have demonstrated human-level performance across many natural-language tasks, including tasks requiring human knowledge, prompting acceptance of parametric knowledge.
- Knowledge Representation historically emphasized explicit knowledge in texts, databases, and knowledge graphs.
- The arrival of LLMs marks a paradigm shift from explicit knowledge representation toward combining explicit and parametric knowledge.Knowledge graphs are a major explicit representation being investigated alongside Transformer-based language models.
- The paper surveys community debates, state-of-the-art integrations of KGs and LLMs, and the resulting opportunities and challenges.Its scope includes relational and ontology-based KGs, tabular data, and numerical values.
2 Common Debate Points within the Community
The community debates how parametric knowledge in LLMs compares with explicit knowledge in KGs across reasoning, accuracy, numerical handling, long-tail knowledge, bias, and interpretability. KGs offer structured, explainable knowledge, while LLMs provide broad generalization but face limitations in factual precision, numerical reasoning, memorization, bias, and transparency.
- Knowledge Representation and Reasoning: LLMs generalize from large-scale text corpora and provide broad language understanding, whereas KGs explicitly represent relationships to support reasoning and inference.Critics question whether LLM statistical patterns constitute true understanding and reasoning.
- High Precision Methods: KGs are valued for high-precision factual information, with YAGO reporting over 95% accuracy and Knowledge Vault failing to reach Google’s desired 99% benchmark.The passage links KG success to accurate entity information and operational requirements.
- Numerical Values: LLMs struggle with numerical values: none of the evaluated models accurately predicted even one Wikidata birth or death year.The passage also notes that numerical shortcomings extend to KG completion tasks and straightforward arithmetic.
- Long-tail Knowledge: LLM performance deteriorates on random Wikidata facts about long-tail entities because entity frequency in pre-training affects memorization, while KGs provide an inherent advantage.This difficulty persists even for sizable LLMs and is evidenced in PopQA and other datasets.
- Bias, Fairness and Beyond: Critics argue that LLMs can perpetuate or amplify training-data biases, while proponents attribute bias to societal patterns in data and emphasize mitigation techniques.The debate centers on whether bias is inherent to LLMs or reflects stereotypes and discrimination in their training data.
- Explainability and Interpretability: KGs are preferred when explainability matters because they explicitly represent entity relationships, whereas LLMs are criticized for lacking transparency about their answers or recommendations.LLM proponents acknowledge explainability as a challenge.
3 Opportunities and Visions
The paper identifies opportunities from integrating LLMs’ parametric knowledge with explicit knowledge, including rapid access to large text corpora and advances across knowledge-engineering tasks. It envisions complementary Explicit-Knowledge-First and Parametric-Knowledge-First use cases, with knowledge graphs scaling engineering and improving LLM reliability.
- Opportunities: LLMs provide rapid access to extremely large text corpora, increasingly even on consumer hardware.This opportunity follows from the prevalence of textual knowledge and LLM processing speed.
- Opportunities: LLMs advance dependency and structured parsing, entity recognition, and relation extraction, enabling knowledge-graph construction at unprecedented scale and quality.These improvements can arise out of the box, through few-shot prompting, or with fine-tuning on a few examples.
- Visions: In Explicit-Knowledge-First use cases, LLMs can enable and simplify knowledge-engineering steps toward knowledge graphs with unprecedented scale, quality, and utility.This vision treats LLMs as mechanisms for advancing crucial parts of the knowledge-engineering pipeline.
- Visions: In Parametric-Knowledge-First use cases, knowledge graphs can improve, ground, and verify LLM generations to increase reliability and trust.The paper presents this as the complementary vision to using LLMs for explicit-knowledge-first engineering.
- Visions: The two use-case classes will coexist rather than form alternatives, with each offering opportunities for synergies between explicit and parametric knowledge.Some applications favor scrutable Explicit-Knowledge-First approaches, while others favor blackbox Parametric-Knowledge-First approaches.
4 Key Research Topics and Related Challenges
This section categorizes and summarizes recent developments in using LLMs and KGs across four themes, building on the availability of both parametric and explicit knowledge.
- 4 Key Research Topics and Related Challenges: The section organizes recent developments in using LLMs and KGs under four themes, motivated by the availability of parametric and explicit knowledge.It presents these themes as related opportunities and visions.
4.1 LLMs for KGs: Knowledge Extraction and Canonicalisation
LLMs support KG construction by extracting and canonicalising knowledge from heterogeneous tables, text, and multilingual sources, while reducing annotation burdens through synthetic data and labeling assistance. Key challenges include representing non-textual tables, processing very long documents, and achieving high recall alongside precision.
- Entity Resolution and Linking: LLM-assisted entity alignment combines general linking, embeddings, and constructed rules or labels to improve knowledge integration, reasoning performance, interpretability, and explainability.LLMs can provide labeled KG samples and support rule and labeled-data construction for entity resolution and linking.
- Tabular Knowledge Extraction: LLM-based table annotation maps columns, cells, and inter-column relationships to KG classes, entities, and properties for direct KG construction and population.Transformer-based table representations, including BERT-based methods, support prediction tasks when table semantics are unclear.
- Tabular Knowledge Extraction: Table processing requires scenario-specific sequence transformations and representations that accommodate numbers, dates, and other non-textual content.Different transformations are needed for fine-tuning, prompted inference, and instruction tuning.
- Text Knowledge Extraction: LLMs extract entities, relations, and events from text using zero-shot and few-shot learning, improving versatility across domains and supporting low-resource, cross-lingual knowledge extraction.These capabilities reduce reliance on large amounts of explicit domain-specific training while extending extraction beyond English.
- Text Knowledge Extraction: LLMs can generate synthetic data for fine-tuning smaller task-specific models, addressing limited training data and potentially improving task performance.Prompting and instruction tuning introduce additional paradigms for adapting models to extraction tasks.
- Challenges and Future Directions: Future extraction systems must handle very long documents, model long-range dependencies, support corpus-level extraction, and improve recall beyond precision-focused pipelines.The paper identifies high-coverage extraction and efficient processing of long documents as open directions, alongside substantial computational requirements.
4.2 LLMs for KGs: Knowledge Graph Construction
LLMs support KG construction through link prediction, inductive link prediction, text-enhanced embeddings, direct KG completion, and extraction of relational knowledge from model parameters. However, these approaches face efficiency, coverage, bias, disambiguation, precision, and provenance challenges.
- Approaches: LLM-based KG construction spans link prediction, inductive prediction for unseen entities, text-enhanced embeddings, direct KG completion, and triple extraction from parametric knowledge.Methods incorporate entity descriptions, structural information, or LLM-generated predictions to construct and complete KGs.
- Link Prediction: Link prediction evaluation is computationally difficult because ranking thousands of candidates requires many expensive predictions from very large LLMs.Common metrics include Hit@k and Mean Reciprocal Rank, but candidate-wise evaluation becomes inefficient at scale.
- Parametric Knowledge Retrieval: LLMs capture relational knowledge, but struggle with low-frequency phenomena, robust fact recall, broad zero-shot knowledge expression, and knowledge access comparable to symbolic knowledge bases.Some apparent accuracy may reflect memorization of training data rather than inference, and probing studies find limited knowledge-access capabilities.
- Triple Extraction: Triple extraction is affected by prompt-structure and majority-viewpoint biases, while manually designed prompts reveal only part of an LLM’s encoded knowledge.These issues complicate the interpretation, reliability, and generalisability of extracted knowledge.
- Challenges: KG construction from LLMs remains challenged by polysemy, homonymy, long-tail entities, numerical values, high-precision requirements, and missing provenance.Without provenance, verifying accuracy and detecting training-data bias become difficult; retrieval augmentation and calibration remain unresolved strategies.
4.3 LLMs for KGs: Ontological Schema Construction
LLMs offer opportunities to construct ontological rules and constraints from contextual information, long documents, and KG facts, while also supporting ontology refinement and alignment. However, practical use remains challenging because generated rules must use correct vocabularies and syntax, and LLM-based ontology systems need better tooling and evaluation under incomplete or inconsistent semantics.
- Ontological Schema Construction: Existing KG extraction pipelines are brittle and error-prone, while manually specifying the rules and constraints needed for quality validation is prohibitively difficult and expensive.Because KGs are incomplete under the open-world assumption, practical applications require semi-automatic data-quality assessment and validation.
- Ontological Schema Construction: LLMs can support automatic rule construction by extracting contextual information, reconciling synonyms and polysemy, generating rules, and producing explanations or validating examples.They may also filter candidate constraints through simple common-sense reasoning and express rules in formal forms.
- Ontological Schema Construction: Parsing procedures, regulations, and specifications against KG entities and predicates could expose constraints, but LLMs must use the correct vocabulary and rule syntax and eventually reason directly over KG facts.The latter direction would require both inductive and abductive reasoning, potentially treating rule generation as a summarization-like task.
- Ontology Refinement: Ontology refinement spans knowledge completion, erroneous-knowledge detection and repair, and knowledge canonicalization, but LLM-based tools still need substantial development before practical deployment.DeepOnto already supports several LLM-based refinement and alignment tasks, while broader accessibility and support for generative models remain challenges.
- Ontology Alignment: Ontology alignment is important for integrating incomplete, cross-domain ontologies, yet LLM-based systems require new evaluation protocols and metrics because ground-truth mappings may be incomplete and textual semantics may conflict with formal ontology semantics.Mappings may concern equivalent, subsumption, or membership relationships between concepts, individuals, or properties.
4.4 KGs for LLMs: Training and Accessing LLMs
Knowledge graphs support LLMs as training data, prompt-construction resources, and external knowledge for retrieval augmentation. This integration injects structured and evolving knowledge, improves generation and question answering, and raises challenges around prompting, hallucination detection, and scalable knowledge updating.
- 4.4 KGs for LLMs: Training and Accessing LLMs: KG factual knowledge integrated during pretraining injects structured and evolving world knowledge and transfers high-resource language knowledge to low-resource language models.KG integration with efficient prompts supports knowledge injection and adaptation to new information.
- 4.4 KGs for LLMs: Training and Accessing LLMs: Grounding KGs in language-model pretraining improves generation and question-answering performance, while vectorized entity and relation memories support knowledge-intensive QA.The neuro-symbolic KB augments language models during pretraining and fine-tuning with representations from an existing KB.
- 4.4 KGs for LLMs: Training and Accessing LLMs: KG-based prompting is proposed to extract relevant knowledge, connect prompts, adapt interactively to user feedback, enhance reasoning, and support reliable information for hallucination detection.The stated research questions also target relevance, non-bias, privacy preservation, ethical constraints, and reasoning capabilities.
- 4.4 KGs for LLMs: Training and Accessing LLMs: A pressing research direction is to unify KG-based knowledge editing with retrieval augmentation through explicit memories and semi-parametric LLMs.The cited SERAC approach stores edits in explicit memory for knowledge editing over LLMs.
4.5 Applications
Applications increasingly combine LLMs’ text-processing and language capabilities with KGs’ explicit, accurate knowledge and symbolic reasoning. This integration spans healthcare, compliance checking, commonsense knowledge, multimodal understanding, contract review, retrieval, and question answering, while raising concerns about accuracy, privacy, incompleteness, cultural variation, and bias.
- General Applications: LLM–KG integration combines accurate explicit knowledge with LLM-based text processing to support healthcare assistants, question answering, chatbots, and sustainability applications.LLMs can help address KG construction challenges from unstructured text, while KGs can mitigate LLM factuality problems.
- Commonsense Knowledge: Commonsense applications use language models and existing resources to extract or complete knowledge, but open-endedness, cultural variation, and bias complicate deciding what to materialize.Systems may either precompute relevant knowledge or invoke an LLM at query time, depending on the context.
- Build Environment: In build-environment compliance, LLMs interpret regulations and enhance rule KGs, supporting automated code checking and graph-based interrogation of 3D models.This approach addresses ambiguous regulatory text, extensive interpretation expertise, and specialized query-language requirements.
- Digital Healthcare: Healthcare applications include clinical documentation, patient-history synthesis, and clinical-trial candidate identification, but adoption must address diagnostic errors and privacy risks.Third-party LLM use can expose protected health information, while open-source deployment improves control but may be less capable and harder to maintain.
- Multimodal Applications: Applications also increasingly seek symbolic representations of image content alongside multimodal image–text embedding methods.This reflects growing interest in combining neural multimodal retrieval with more explicit representations of visual information.
- Documents and Retrieval: Combining domain-specific KGs with LLM understanding and reasoning can improve contract review and support advanced retrieval and question-answering over long-form documents.The proposed direction joins graph-based symbolic extraction, long-text understanding, and domain-specific code synthesis.
5 Outlook
The advances of LLMs mark an inflection point for KG research, creating opportunities for hybrid approaches while leaving open questions about combining their strengths. The outlook argues that structured knowledge remains essential for reliability- or safety-critical applications, even as LLMs reshape and sometimes obsolete parts of KG and ontology construction.
- 5 Outlook: LLM advances create exciting opportunities for KG research, but how to combine their respective strengths remains an open question.The community is adapting through forums such as the KBC-LM workshop and LM-KBC challenge, alongside a shift toward hybrid approaches.
- 5 Outlook: Structured knowledge should be retained because it remains indispensable for reliability- and safety-critical applications.The outlook describes multiple ways KGs and LLMs can mutually reinforce one another rather than treating KGs as obsolete.
- 5 Outlook: LLMs have substantially advanced KG and ontology construction, making some pipeline tasks obsolete and requiring critical reassessment of established workflows.