PaperlayerView exact revision on arXiv

Source-linked AI summary

Unifying Large Language Models and Knowledge Graphs: A Roadmap

Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, Xindong Wu

arXiv:2306.08302v3cs.CLcs.AI

TL;DR

LLMs lack reliable factual access and interpretability, while KGs face construction and evolving-knowledge challenges. This article reviews a roadmap spanning three integration frameworks, synthesizes existing research, and identifies future directions, while noting computational and interpretability limitations in synergized reasoning and generation-based KG completion.

  • Problem

    LLMs can hallucinate and hide knowledge in black-box parameters, while KGs struggle with incomplete, dynamically changing knowledge and unseen entities.

  • Method

    The article proposes and reviews three frameworks: KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLMs + KGs.

  • Results

    The roadmap organizes recent integration methods across LLM enhancement, KG applications, and mutual LLM-KG reasoning, and summarizes emerging research directions.

  • Takeaways & Limitations

    The roadmap presents unification as a research direction in which LLMs and KGs can mutually enhance knowledge representation and reasoning.

  • Takeaways & Limitations

    Synergized reasoning may add parameters and computational costs while lacking interpretability, and generated KG entities may not belong to the graph.

Abstract

from arXiv · show

Large language models (LLMs), such as ChatGPT and GPT4, are making new waves in the field of natural language processing and artificial intelligence, due to their emergent ability and generalizability. However, LLMs are black-box models, which often fall short of capturing and accessing factual knowledge. In contrast, Knowledge Graphs (KGs), Wikipedia and Huapu for example, are structured knowledge models that explicitly store rich factual knowledge. KGs can enhance LLMs by providing external knowledge for inference and interpretability. Meanwhile, KGs are difficult to construct and evolving by nature, which challenges the existing methods in KGs to generate new facts and represent unseen knowledge. Therefore, it is complementary to unify LLMs and KGs together and simultaneously leverage their advantages. In this article, we present a forward-looking roadmap for the unification of LLMs and KGs. Our roadmap consists of three general frameworks, namely, 1) KG-enhanced LLMs, which incorporate KGs during the pre-training and inference phases of LLMs, or for the purpose of enhancing understanding of the knowledge learned by LLMs; 2) LLM-augmented KGs, that leverage LLMs for different KG tasks such as embedding, completion, construction, graph-to-text generation, and question answering; and 3) Synergized LLMs + KGs, in which LLMs and KGs play equal roles and work in a mutually beneficial way to enhance both LLMs and KGs for bidirectional reasoning driven by both data and knowledge. We review and summarize existing efforts within these three frameworks in our roadmap and pinpoint their future research directions.

1 INTRODUCTION

LLMs provide broad language capabilities but retain factuality and interpretability limitations, while KGs provide explicit, structured knowledge yet face construction and coverage challenges. The paper proposes a roadmap that categorizes complementary ways to unify them.

  • LLMs show strong generalization and language-processing abilities, but their knowledge is implicit, potentially hallucinated, and difficult to interpret.
  • KGs store explicit facts in triples and support accurate, decisive, interpretable, and evolving knowledge representation.
  • Existing KGs remain difficult to construct and inadequately represent incomplete, dynamically changing knowledge, unseen entities, and new facts.
  • The article reviews integration strategies and tasks, covers emerging LLM and KG advances, and identifies challenges and future research directions.
  • The roadmap identifies KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLMs + KGs as three frameworks for integration.

2 BACKGROUND

The background introduces LLM architectures and prompting alongside KG representations and categories. It also presents representative applications and the roadmap’s three integration frameworks.

  • Large Language Models: LLMs are commonly grouped as encoder-only, encoder-decoder, or decoder-only models based on Transformer architecture and self-attention.
  • Prompt Engineering: Prompt engineering uses task-specific natural-language inputs containing instructions, context, and input text to improve LLM performance.
  • Prompt Engineering: Prompts can incorporate knowledge graphs by linearizing graph structure or converting it into task-specific textual representations.
  • Knowledge Graphs: Knowledge graphs store triples over entities and relations and are categorized as encyclopedic, commonsense, domain-specific, or multimodal.
  • Unification Roadmap: The roadmap organizes unification into KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLMs + KGs.
  • Applications: Representative applications combine LLMs and KGs for chatbots, coding assistance, search, recommendation, external knowledge, and healthcare assistance.

3 ROADMAP & CATEGORIZATION

The paper organizes LLM–KG unification into three frameworks: KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLMs + KGs. It further categorizes research by integration strategy, task type, and mutually enhancing representation and reasoning.

  • Roadmap: The roadmap identifies KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLMs + KGs as three unification frameworks.The first two enhance LLMs and KGs separately, while the third mutually enhances both.
  • Synergized LLMs + KGs: The unified framework contains Data, Synergized Model, Technique, and Application layers.It processes textual and structural data, combines model capabilities and techniques, and supports applications such as search, recommendation, and AI assistants.
  • KG-enhanced LLMs: KG-enhanced LLM research covers pre-training, inference, and interpretability.These categories respectively inject knowledge during training, provide access during inference, and analyze learned knowledge or reasoning.
  • LLM-augmented KGs: LLM-augmented KG research covers embedding, completion, construction, graph-to-text generation, and other KG-related tasks.LLMs encode textual descriptions, generate or complete facts, support entity and relation extraction, and help convert graph content into text.
  • Synergized LLMs + KGs: The Synergized LLMs + KGs framework integrates both technologies to mutually enhance knowledge representation and reasoning.The review examines synergized approaches from the perspectives of representation and reasoning.

4 KG-ENHANCED LLMS

KG-enhanced LLM methods inject structured knowledge during training or inference and use KGs to improve interpretability. The section compares their knowledge coverage, updateability, training requirements, and retrieval or prompting mechanisms.

  • Pre-training: KG-enhanced LLM pre-training incorporates KGs into training objectives, inputs, or instruction-tuning datasets.The surveyed methods expose entities, align words with graph entities, concatenate graph information, or train models to follow KG-based instructions.
  • Limitations: KG instruction-tuning and pre-training approaches require retraining to incorporate knowledge changes.Instruction-tuning is resource-intensive, and pre-training methods cannot update incorporated knowledge without retraining.
  • Inference: Inference methods retrieve or prompt with external KG information without incorporating it permanently into model parameters.Retrieval-augmented fusion supplies relevant documents as generation context, while KG prompting formats graph structure for reasoning.
  • Inference: RAG outperforms parametric-only and non-parametric-only baselines in open-domain QA.The passage also reports more specific, diverse, and factual text than parameter-only baselines.
  • Comparison: Pre-training can align knowledge with linguistic context, whereas inference supports knowledge updates but may yield sub-optimal performance.Inference methods can maintain updates by changing inputs, while the underlying LLM may not be trained to fully leverage the supplied knowledge.
  • Interpretability: KG-based interpretability research uses KGs for language-model probing and analysis.The goal is to understand hidden knowledge and reasoning processes while addressing hallucination-related reliability concerns.

5 LLM-AUGMENTED KGS

LLM-augmented KG research addresses incompleteness and limited textual processing by applying LLMs to KG-related tasks. The section frames LLMs as tools for enriching graph representations and downstream applications.

  • Motivation: LLMs augment KGs by incorporating textual information and improving downstream task performance.This direction responds to incomplete conventional KGs and methods that insufficiently process text corpora.
  • Categorization: LLM-augmented KG research is organized around tasks including embedding, completion, construction, generation, and question answering.The supplied figure identifies LLMs as text encoders for KG embedding, while the broader roadmap abstract lists the principal task families.

5.1 LLM-augmented KG Embedding

LLM-augmented KG embedding methods combine textual descriptions with graph structure to produce richer entity and relation representations. They either use LLM encodings within KGE models or jointly model graph-text inputs.

  • Background: KGE maps entities and relations into low-dimensional vectors containing semantic and structural information.These embeddings support question answering, reasoning, and recommendation.
  • LLM-enhanced KGE: Pretrain-KGE encodes entity and relation descriptions with an LLM before a KGE model generates final embeddings.The KGE model learns structural information while retaining some knowledge from the LLM representations.
  • Joint text and KG embedding: Joint text–KG embedding methods treat entities and relations as LLM tokens and train masked-entity prediction.After training, token representations serve as entity and relation embeddings.
  • Joint text and KG embedding: Neighbor-aware methods concatenate one-hop neighboring entities with triples to better capture graph structure.LambdaKG feeds sampled neighbors and the triple into the LLM as a sentence.

5.2 LLM-augmented KG Completion

LLM-augmented KG completion uses language models either as encoders that score candidate facts or as generators that produce missing entities directly. These approaches incorporate textual information and support different trade-offs between fine-tuning, efficiency, and validity.

  • LLM-based KG completion methods encode text or generate facts, forming encoder-based PaE and generator-based PaG categories.
  • LLM as Encoders (PaE): PaE encodes triples or their components and feeds the resulting representations into prediction heads or KG scoring functions.Joint, masked-language-model, and separated-encoding variants are described.
  • LLM as Encoders (PaE): Joint encoding represents a triple as a text sequence, then uses the final [CLS] representation for binary plausibility classification.KG-BERT and related methods extend this setup with auxiliary tasks, supporting information, or jointly learned semantic and structural embeddings.
  • LLM as Encoders (PaE): Masked and separated encoding variants predict masked entities or score separately encoded (h, r) and t representations.OpenWorld KGC uses sequential entity-description and incomplete-triple prediction modules, while SimKGC applies contrastive learning to encoded representations.
  • LLM as Generators (PaG): PaG receives a query triple (h, r, ?) and generates the tail entity directly with encoder-decoder or decoder-only language models.Prompt engineering, demonstrations, entity-aware decoding, constrained decoding, and task-specific conditions are among the described techniques.
  • Comparison: PaE is easier to fine-tune and integrates readily with candidate scoring, whereas PaG avoids ranking all candidates and can generalize to unseen entities.PaG nevertheless may generate entities absent from the KG, requires longer autoregressive inference, and still faces open prompt-design questions.

5.3 LLM-augmented KG Construction

LLMs augment KG construction across entity discovery, coreference resolution, relation extraction, and newer end-to-end pipelines. They also support knowledge distillation from implicit model knowledge into newly generated KG tuples.

  • KG construction typically includes entity discovery, coreference resolution, and relation extraction, with Fig. 19 presenting LLM use across these stages.More recent approaches also explore end-to-end construction.
  • Entity Discovery: Entity discovery identifies entities in unstructured sources and includes flat, nested, and discontinuous named-entity recognition settings.GenerativeNER uses a sequence-to-sequence LLM with a pointer mechanism to address all three NER types.
  • Entity Discovery: Entity typing assigns fine-grained or ultra-grained types, while entity linking connects textual mentions to corresponding KG entities.Generative approaches can formulate linking as sequence-to-sequence generation of identifier-annotated text.
  • Coreference Resolution: Coreference resolution groups expressions referring to the same entity or event within one document or across multiple documents.LLM-based systems use span encoders, language modeling, mention scoring, and approaches designed to reduce memory demands.
  • Relation Extraction: Relation extraction identifies semantic relations between entities at sentence or document scope, with LLM-based systems modeling contextual, global-local, graph, and evidence information.Document-level methods address relations across multiple sentences and may incorporate multi-label handling and localized context.
  • Knowledge Distillation and End-to-End Construction: LLMs can construct commonsense KGs by learning from seed tuples and generating novel tuples, while prompt-based pipelines use ChatGPT and GPT-4 for construction tasks.The roadmap also presents knowledge distillation from LLMs as a construction direction.

5.4 LLM-augmented KG-to-text Generation

LLM-augmented KG-to-text generation converts graph knowledge into fluent text for applications such as storytelling and knowledge-grounded dialogue. Existing work fine-tunes LLMs on linearized graphs, while later methods add graph structure or aligned corpora.

  • KG-to-text generation aims to produce accurate and consistent text describing input KG information for realistic natural-language-generation applications.The passage identifies storytelling and knowledge-grounded dialogue as examples.
  • LLM-based KG-to-text Methods: Early approaches fine-tune LLMs such as BART and T5 using a linear traversal of the input graph.These approaches report that the simple linearized representation outperforms many existing KG-to-text systems, and continued pre-training can further improve performance.
  • LLM-based KG-to-text Methods: Linearized-graph methods cannot explicitly incorporate the rich graph semantics of KGs, motivating approaches that inject KG structure information into LLMs.JointGT is presented as an approach designed to enhance LLMs with KG structure.
  • Aligned Data Construction: Because unsupervised LLM pre-training objectives may not align with KG-to-text generation, researchers construct large-scale KG-text aligned corpora.One work builds a 1.3M unsupervised corpus and provides more than 1,000 human-annotated KG-to-text test examples.

5.5 LLM-augmented KG Question Answering

LLMs are applied to KGQA to bridge natural-language questions and structured graph facts through retrieval, path construction, and answer reasoning. These approaches use LLMs to extract relevant entities and relations, retrieve paths, and assess candidate answers against retrieved evidence.

  • KGQA answers natural-language questions using structured facts from knowledge graphs, with LLMs helping bridge language understanding and graph-based reasoning.
  • Retrieval: LLM-based entity and relation extractors identify question concepts and retrieve related graph facts, improving relation prediction over shallow neural networks.
  • Retrieval: LLM-based path retrievers select question-related relations hop by hop and construct reasoning paths whose probabilities can be calculated.
  • Answer reasoning: Answer reasoners combine questions with retrieved facts, paths, or candidate answers and use LLM representations to score or classify answer hypotheses.
  • Integrated frameworks: KGQA frameworks increasingly integrate retrieval and reasoning, including graph-neural modules, knowledge interaction layers, semantic matching, and joint message passing.

6 SYNERGIZED LLMS + KGS

Synergized LLMs + KGs aim to represent and reason over textual and structured knowledge jointly, with both technologies contributing complementary capabilities. The roadmap distinguishes joint representation, fusion reasoning, and agent-based reasoning while noting trade-offs in flexibility, interpretability, and computational cost.

  • Synergized LLMs + KGs combine LLM language understanding with KG factual knowledge to enhance downstream knowledge representation and reasoning.
  • Synergized Knowledge Representation: Synergized knowledge representation models jointly represent implicit textual knowledge and explicit structured graph knowledge for downstream tasks.
  • Synergized Knowledge Representation: Additional KG fusion modules are jointly trained with LLMs, using architectures such as dual encoders and joint graph-text representation learning.
  • Synergized Reasoning: LLM-KG fusion reasoning uses separate language and graph encoders that jointly fuse knowledge from text and KGs for reasoning.
  • Comparison and Discussion: Fusion frameworks support end-to-end training and flexible encoder choices, whereas LLM agents can reason across KGs without additional training cost and provide interpretable processes.
  • Comparison and Discussion: Fusion frameworks may add parameters and computational costs while lacking interpretability, and defining actions and policies remains challenging for LLM agents.

7 FUTURE DIRECTIONS AND MILESTONES

Future work targets persistent hallucinations, knowledge updating, black-box access, multimodal graph data, and direct understanding of graph structure. The roadmap also anticipates bidirectional reasoning in which KG search and LLM inference mutually validate one another.

  • Challenges: Hallucinations continue to threaten LLM reliability, motivating KG-based methods for more reliable generation and hallucination detection.
  • Challenges: Updating a single fact in an LLM can have ripple effects on related knowledge, while current editing methods may show poor performance or computational overhead.
  • Challenges: API-only access to state-of-the-art LLMs prevents conventional approaches that modify model structures by adding knowledge-fusion modules.
  • Future directions: Future KGs must better leverage diverse modalities by encoding and aligning entities across multimodal representations.
  • Future directions: LLMs need methods that directly understand and reason over KG structure because whole-graph linearization is infeasible and may lose underlying information.
  • Bidirectional reasoning: Bidirectional reasoning combines KG knowledge-driven search with LLM data- or text-driven inference so the two processes can mutually validate derived goals and data.

8 CONCLUSION

The article surveys the active effort to unify LLMs and KGs, organizing integration methods, KG applications, challenges, and future directions. Its roadmap envisions progression from separate enhancement and augmentation approaches to synergized systems and broader structural, multimodal, and updating capabilities.

  • The article provides a thorough overview of research integrating KGs to enhance LLMs and applying LLMs across KG tasks.
  • The roadmap organizes future development into stages covering KG-enhanced LLMs and LLM-augmented KGs, synergized systems, then graph understanding, multimodality, and knowledge updating.

APPENDIX A PROS AND CONS FOR LLMS AND KGS

LLMs offer broad knowledge, language-processing ability, and generalizability, but their implicit and probabilistic reasoning can produce hallucinations and limit interpretability. KGs provide structured, accurate, decisive, interpretable, domain-specific, and evolving knowledge, but remain incomplete, text-limited, and challenged by unseen facts.

  • Pros of LLMs: LLMs provide general knowledge, strong language processing, and broad generalizability across downstream tasks.Their capabilities support question answering, translation, text generation, few-shot learning, and multi-task applications.
  • Cons of LLMs: LLMs represent knowledge implicitly, making it difficult to interpret or validate what they know.Their black-box decision patterns and functions are not directly accessible or explainable to humans.
  • Cons of LLMs: LLMs often generate plausible but factually incorrect content, reducing their trustworthiness in real-world scenarios.Their probabilistic reasoning also produces indecisive results that are difficult to control.
  • Pros of KGs: KGs store factual knowledge structurally, with curated facts and deterministic symbolic reasoning that support accuracy, decisiveness, and interpretability.Their triple-based representation is understandable by humans and machines, while symbolic reasoning provides an interpretable process.
  • Pros of KGs: KGs can provide precise domain-specific knowledge and evolve by inserting new facts and deleting outdated ones.Experts can construct domain-focused KGs, while updates accommodate changing knowledge.
  • Cons of KGs: KGs are often incomplete, underuse textual information, and struggle to model unseen entities and new facts.These limitations constrain comprehensive knowledge coverage and affect tasks such as KG completion and KG question answering.
Loading 2306.08302v3…