Analyzing Bias in LLM-Augmented Knowledge Graph Systems: Taxonomy, Interaction Mechanisms, and Evaluation

Abstract

Large Language Models (LLMs) are increasingly integrated into Knowledge Graph (KG) construction and augmentation pipelines to reduce manual effort and enable scalable knowledge extraction, completion, and reasoning. While this integration offers substantial benefits, it also introduces new forms of bias and unreliability that extend beyond those observed in standalone LLMs or traditional knowledge graphs. In particular, biases originating from language models, such as social and representational bias, hallucination, prompt sensitivity, and domain coverage limitations that interact with structural and content biases inherent to knowledge graphs, result in compounded distortions that propagate across the pipeline. This paper provides a structured and comprehensive analysis of bias in LLM-augmented knowledge graph systems. We first review bias mechanisms in LLMs and standalone KGs, and then examine how these biases interact and amplify during key stages of LLM-based entity extraction, relation generation, graph completion, and reasoning. Based on this analysis, we introduce a unified taxonomy that characterizes bias as a pipeline-level phenomenon rather than an isolated model. We further consolidate recent evaluation metrics adapted for LLM-generated graphs, including semantic and soft lexical measures. Additionally, we survey representative datasets and benchmarks used to study bias in LLMs, KGs, and hybrid LLM–KG systems and identify open research gaps in developing pipeline-aware evaluation frameworks. This work aims to support the design of more reliable, accurate, and fair LLM-augmented knowledge graphs for engineering and domain-specific applications.

1. Introduction

Product development increasingly relies on the effective use of data-driven methods and intelligent tools. Designing complex products requires not only subject-matter expertise but also the ability to efficiently capture and reuse prior knowledge. Knowledge Graphs (KGs) can support this process by organizing information, enabling case retrieval, recommendations, and semantic reasoning. However, their adoption faces three key challenges: (1) collecting sufficient domain-specific training data is difficult, and manually labeling it is time-consuming; (2) building and training KG models requires coding and machine learning skills, limiting accessibility for most engineers; and (3) current KGs have limited transferability, as knowledge learned in one domain often cannot be applied effectively to another [1].

In response to these challenges, large language models (LLMs) offer advantages in the construction, completion, and enrichment of Knowledge Graphs (KGs). By leveraging their pretraining on large text corpora, LLMs can automatically extract entities, relationships, and patterns from unstructured text, reducing reliance on manual annotation or domain experts. They improve the usability of knowledge graphs by converting natural language into a structured query language for knowledge retrieval and by inferring missing entities, relations, and knowledge in the graph using their pretrained knowledge and contextual reasoning. These enhancements improve the effectiveness of knowledge graphs for reasoning, knowledge management, and decision-making.

LLMs can also convert KG triples into natural language descriptions, enabling applications such as KG-to-text generation and question answering [2]. One of the benefits of leveraging LLMs for constructing knowledge graphs is in healthcare domains. LLMs have been used to extract relationships among diseases, treatments, and symptoms from research papers, clinical records, and other textual sources to build structured KGs [3,4]. Large Language Models (LLMs) can also process multimodal data during knowledge graph construction, yielding semantically richer representations that support more effective requirement analysis. This capability extends to the integration of textual, visual, and symbolic information, enabling advanced applications in robotics and embodied AI [5,6]. The main contributions of this paper are summarized as follows:

• We introduce a structured taxonomy of bias in LLM-augmented knowledge graph pipelines, distinguishing between model-level phenomena (i.e., social, representational, hallucination, prompt-sensitivity, and domain coverage bias) and graph-level distortions such as (structural and content bias).
• We provide a unified conceptual framework that links biases across three layers: (i) Large Language Models, (ii) standalone Knowledge Graphs, and (iii) their interaction within LLM-augmented pipelines, clarifying how these biases differ in origin, manifestation, and impact on entity extraction, relation generation, graph completion, and reasoning.
• We conduct a comprehensive comparative analysis of bias types and their implications for knowledge graph construction.
• We consolidate and categorize existing datasets and benchmarks for bias evaluation in LLMs, KGs, and hybrid LLM-KG systems, organizing them across textual, structural, multimodal, and algorithmic settings and highlighting their evaluation focus and suitability for studying different bias dimensions.

Several surveys have examined bias in large language models and the structural properties of knowledge graphs independently; the interaction between the two components in hybrid LLM-KG pipelines still needs further studies. Existing work focuses either on fairness and social bias in LLM outputs [7] or structural characteristics of knowledge graphs and their embeddings [8]. More recent surveys investigate the integration of LLMs with knowledge graphs for tasks such as entity extraction and graph augmentation [2]. However, these studies do not provide a comprehensive analysis of how biases propagate through the full LLM–KG pipeline.

Table 1. Positioning of this work relative to existing surveys on bias in LLMs, Knowledge Graphs, and LLM–KG systems. Note: ✔ indicates full coverage; — indicates not covered.

Survey	Primary Focus	LLM	KG	Pipeline	Evaluation Perspective
[7]	Fairness and bias in Large Language Models	✔	—	—	Focus on linguistic bias detection and mitigation within LLM outputs
[8]	Knowledge Graph structure and embedding behaviour	—	✔	—	Focus on structural bias in KG topology and link prediction models
[2]	LLM–KG integration frameworks and benchmarks	Partial	Partial	Limited	Evaluation of KG construction and augmentation tasks
Our Work	Pipeline-level bias propagation in LLM-augmented KG systems	✔	✔	✔	Unified taxonomy and evaluation synthesis addressing linguistic and structural bias interactions

positions our work relative to these prior surveys.

Illustrative Example of Bias Propagation: To illustrate how biases may propagate across the stages of an LLM-augmented knowledge graph pipeline, consider a simplified biomedical example. Suppose the input text states: “Drug X significantly reduces symptoms of Disease Y in clinical trials.” During entity extraction, the model correctly identifies the entities Drug X and Disease Y, and may generate the relation treats. However, due to hallucination bias, the LLM may produce an unsupported triple such as (Drug X, cures, Disease Y), despite the source text only indicating symptom reduction rather than complete recovery.

If such a hallucinated triple is incorporated during knowledge graph construction, it introduces an incorrect semantic relation that overstates the underlying evidence. When accumulated across multiple extractions, these errors can distort the graph structure, for example, by creating relation imbalances or incorrect connectivity patterns between entities.

These distortions can subsequently affect downstream applications that rely on the constructed knowledge graph. For instance, a question answering system querying the graph with “Does Drug X cure Disease Y?” may incorrectly return an affirmative response due to the presence of the hallucinated relation. Such bias propagation can be identified using quantitative evaluation metrics. At the generation level, the Unsupported Triple Ratio (UTR) measures the proportion of generated triples that cannot be verified against source evidence or external knowledge bases. At the graph level, structural indicators such as the Edge Imbalance Ratio (EIR) can reveal abnormal growth of specific relations resulting from hallucinated outputs. For clarity, the list of abbreviations used throughout this paper can be found in

Table 2. List of abbreviations used in this paper.

Abbreviation	Definition
LLM	Large Language Model
KG	Knowledge Graph
RAG	Retrieval-Augmented Generation
CoT	Chain-of-Thought
KGQA	Knowledge Graph Question Answering
CQ	Competency Question
EA	Entity Alignment
GAL-KARS	Graph Augmentation with LLMs for Knowledge-Aware Recommender Systems
DCR	Domain Coverage Ratio
UTR	Unsupported Triple Ratio
EIR	Edge Imbalance Ratio
ASS	Alignment Stability Score
PBS	Position Bias Score
PopBS	Popularity Bias Score
G-BS	G-BERTScore
G-BL	G-BLEU
G-RO	G-ROUGE
RD	Representation Disparity

.

The structure of this paper is organized as follows: Section 2 provides a background on bias in Large Language Models and Knowledge Graphs. Section 3 analyzes the integration of LLMs into knowledge graph pipelines and examines how biases propagate and interact across extraction, completion, and reasoning stages. Section 4 highlights the work done in the literature. Section 5 explains the evaluation metrics, and Section 6 reviews representative datasets and benchmarks used for bias analysis in LLMs. Section 7 addresses the research gaps. Finally, Section 8 is for the discussion and conclusion.

2. Background: Bias in Large Language Models and Knowledge Graphs

Bias has become a central concern in the use of Large Language Models and Knowledge Graphs, as these systems increasingly influence knowledge access, reasoning, and decision-making. Such bias can arise from data imbalance, modeling choices, or deployment context and may affect both factual reliability and graph fairness. In this section, we summarize the main categories of bias identified in prior work to provide background for the analysis of LLM-augmented knowledge graph systems.

Table 3. Comparative overview of bias types in Large Language Models, Knowledge Graphs, and LLM-augmented Knowledge Graph pipelines.

Bias Type	System	Primary Source	Manifestation	Implications for LLM-KG Pipelines	Observable Signal/Mitigation
Biases in Large Language Models
Social and Representational	LLM	Imbalanced training data and annotation practices	Stereotypical associations, ideological skew, and unequal performance across demographic groups [9,10,11]	Biased entity descriptions, skewed relation generation, and distorted population-level simulations	Demographic performance disparity in generated entities; mitigation via dataset balancing or bias-aware prompting
Hallucination and Factual	LLM	Domain gaps, weak grounding, and confidence miscalibration	Fluent but factually incorrect or unverifiable content [12,13]	Injection of false entities or relations that corrupt graph structure and reasoning	Unsupported or unverifiable triples in generated graphs; mitigation via retrieval grounding or fact verification
Prompt-sensitivity	LLM	Prompt formulation, formatting, and example ordering	Output instability under semantically equivalent prompts [14,15]	Inconsistent entity extraction, relation phrasing variation, and reduced reproducibility	High output variance across prompt paraphrases; mitigation via prompt standardization or ensemble prompting
Domain Coverage	LLM	Uneven representation of domains or languages in training corpora	Asymmetric accuracy across high- and low-resource domains [16,17]	Sparse or unreliable graph augmentation in specialized or technical domains	Domain-specific recall imbalance in extracted entities; mitigation via domain-adaptive fine-tuning or expert validation
Biases in Knowledge Graphs
Structural	KG	Graph construction, sampling strategies, and data availability	Degree imbalance, popularity bias, and long-tail underrepresentation [18,19]	Completion and retrieval biased toward dense or well-connected subgraphs	Skewed node degree distribution or popularity bias; mitigation via debiased sampling or re-weighted training
Content	KG	Source-dependent documentation and ontology design	Coverage gaps and culturally contingent representations [20]	Systematic omission or under-specification of minority entities and attributes	Missing entities or attribute imbalance across groups; mitigation via ontology refinement or curated knowledge sources
Biases Emerging in LLM-Augmented Knowledge Graphs
Pipeline-level Interaction	LLM-KG	Cross-component bias amplification	Reinforcement of linguistic and structural biases across extraction and completion stages [21]	Persistent, compounded bias that is difficult to detect using standalone evaluation	Bias amplification across pipeline stages; mitigation via cross-stage validation and pipeline auditing

summarizes the main categories of bias discussed in this section, highlighting their sources, manifestations, and implications across Large Language Models, Knowledge Graphs, and LLM-augmented knowledge graph pipelines.

2.1. Bias in Large Language Models

Large Language Models inherit and amplify biases present in their training corpora, model architectures, and optimization objectives. Prior work has shown that LLMs exhibit systematic social biases, domain coverage imbalance, and hallucination tendencies, which can affect the factual reliability, completeness, and fairness of the resulting knowledge graph [22]. Existing research on bias in LLMs is commonly categorized into social bias, hallucination and factual bias, prompt-sensitivity bias, and domain coverage bias.

2.1.1. Social and Representational Bias

Social bias in Large Language Models refers to systematic differences in how social groups, identities, or viewpoints are reflected in model outputs. These biases originate from imbalances in training data, annotation practices, and modeling choices, and they can manifest as stereotypical associations, ideological skew, or unequal performance across demographic groups. Research has consistently shown that LLMs not only reproduce linguistic patterns but also encode the social structures and power relations present in the data they are trained on [9]. Recent studies provide empirical evidence of such biases across multiple dimensions. For instance, prior work has identified a pronounced left-libertarian ideological tendency in responses generated by widely deployed conversational models, revealing a systematic political skew in ideologically oriented outputs [10]. Other studies have demonstrated that LLMs encode gender stereotypes, producing responses that align with socially entrenched assumptions about roles and attributes associated with different genders [11]. Beyond output content, it has also been argued that the use of LLMs in experimental and survey-based settings may hinder the goal of obtaining representative and inclusive samples, particularly when model behavior varies across populations that are unevenly represented in training data. A line of research has examined representational bias through the lens of political simulation tasks [23]. In this context, LLMs are used to simulate public opinion or voting behavior, which places strong demands on representational accuracy and demographic balance. These studies focus on vote choice and opinion formation because of their central role in political science and their growing relevance in agent-based simulations for social science research. Representation bias in such simulations is commonly defined as a systematic difference in simulation performance across groups [24], where certain populations are modeled more accurately or consistently than others.

Representation Disparity (RD): To quantify representational or social bias in the generated knowledge graph, we measure the disparity between the proportion of entities belonging to a demographic or social group in the generated graph and a reference distribution. The Representation Disparity (RD) is defined as:

$$RD={\displaystyle \frac{|p_g-p_{ref}|}{p_{ref}}}$$

(1)

where $p_g$ represents the proportion of entities or relations associated with group g in the generated knowledge graph, and $p_{ref}$ denotes the corresponding proportion in a reference dataset or corpus. Higher RD values indicate stronger representational imbalance.

2.1.2. Hallucination and Factual Bias in LLMs

Hallucination and factual bias refer to the tendency of large language models to generate content that is syntactically fluent and semantically plausible but factually incorrect, unverifiable, or unsupported by external knowledge. While social bias impacts fairness and representation in model outputs, hallucination and factual bias primarily affect the accuracy and reliability of generated content, particularly in knowledge-intensive or high-stakes domains such as medicine, law, and scientific reasoning. Prior work has shown that hallucinations are not random errors but exhibit systematic patterns that correlate with domain coverage gaps, prompt formulation, and model confidence calibration [12]. This form of bias becomes critical in settings where LLM outputs are used to populate, augment, or reason over structured knowledge representations. As a result, hallucination and factual bias play a central role in LLM-augmented knowledge graph pipelines, where generations may directly corrupt the graph structure or propagate false relations [13]. A detailed analysis of the specific mechanisms through which these hallucinations manifest as structural distortions, such as entity omission and relational sparsity, is provided in Section 3.

Prior work has shown that hallucinations are not random errors but follow systematic patterns correlated with domain coverage gaps, prompt formulation, and model confidence calibration [12]. This bias is particularly critical when LLM outputs are used to populate, augment, or reason over structured knowledge representations. In such settings, hallucination and factual errors can directly distort the knowledge graph, introducing false relations or omitting important entities [13]. Section 3 provides a detailed analysis of the specific mechanisms through which these hallucinations manifest as structural distortions, including entity omission and relational sparsity.

2.1.3. Prompt-Sensitivity Bias

Prompt sensitivity bias refers to systematic variations in model output that arise from prompts that preserve the original intent but differ in surface realization. These differences can include changes in phrasing, formatting, the ordering of constraints, or the placement of few-shots [14]. This phenomenon is relevant in bias analysis because it can lead to inconsistent decisions across different demographic groups or topics, even when the underlying query remains the same. Such instability complicates reproducibility and makes observed model behavior highly sensitive to specific researcher choices rather than the model’s inherent knowledge [15]. In LLM-augmented knowledge graph pipelines, prompt sensitivity can lead to significant changes in extracted entities, relation phrasing, and triple selection. This dibasis of prompt sensitivity, the specific mechanisms and pipeline-level implications for graph construction, including their effects on specific mechanisms and pipeline-level implications for graph construction, including how this sensitivity affects relation extraction and schema alignment, are discussed in detail in Section 3.

2.1.4. Domain Coverage Bias

Domain coverage bias in LLMs refers to systematic performance disparities that arise when certain knowledge domains, languages, or topical areas are unevenly represented in training corpora. Unlike prompt sensitivity or social bias, which mainly affect how information is expressed, domain coverage bias influences what information is reliably accessible to the model [16]. Research has shown that LLMs exhibit strong asymmetries in factual accuracy and reasoning depth across domains, with markedly better performance in high-resource, well-documented areas such as general encyclopedic knowledge, and weaker reliability in specialized, low-resource, or technical domains [17]. These gaps are further amplified in scientific, medical, and legal settings, where sparse supervision and rapidly evolving knowledge create conditions that lead models to omit relevant facts or generate unsupported content.

2.2. Bias in Knowledge Graphs

Bias in knowledge graphs arises from the data sources used for construction, ontology design choices, and graph completion mechanisms. Unlike LLM bias, KG bias is often encoded structurally and persists across downstream applications.

Structural and Content Bias

Structural and content bias in Knowledge Graphs refers to systematic distortions in the graph structure and the organization of its internal facts. Structural bias emerges when connectivity patterns, degree distributions, or relation frequencies unevenly represent specific entities or groups [18]. This often privileges popular or well-documented nodes while leaving long-tail entities unconnected. Such distortions directly impact completion and retrieval tasks, as many graph algorithms rely heavily on neighborhood evidence and local topology to function. Recent work on bias detection in graph embeddings suggests that these structural biases can be quantified using numerical measures and identified without manual auditing of specific relations [19,25]. Furthermore, analyses of link prediction have shown that choices made during dataset construction and sampling can introduce selection effects that alter performance metrics, disproportionately benefiting certain regions of the graph or specific relation types [19]. In entity alignment tasks, evidence also indicates that embedding-based methods may exhibit a bias against smaller or less connected subgraphs, confirming that graph topology acts as an implicit advantage signal for learning algorithms [26]. In contrast, content bias focuses on the factual content of the graph, including omissions, skewed coverage of attributes, and culturally or historically contingent representations. Because many large-scale knowledge bases are built from web-derived or community-edited sources, the content often reflects uneven documentation practices across different regions, languages, and communities. These skews manifest as coverage gaps and the overrepresentation of dominant perspectives, leading to the under-specification of minority entities [20].

2.3. From Standalone Biases to LLM-Augmented Knowledge Graphs

While biases in LLMs and KGs have been studied in isolation, recent research highlights that their interaction introduces new and compounded bias dynamics. In LLM-augmented knowledge graph pipelines, biases originating from language models and structured graphs interact, amplify, and propagate through multiple stages of knowledge extraction, completion, and reasoning. Consequently, errors or skews present in one component may be reinforced by the other, leading to biased graph structures that are difficult to diagnose using standalone evaluation methods [21]. The bias in LLM-augmented knowledge graphs should be treated as a pipeline-level phenomenon rather than an isolated property of either component. Hybrid systems introduce unique sources of bias, including positional bias during graph construction and prompt-induced variation in extracted triples. These observations motivate the need for evaluation and mitigation strategies that jointly consider linguistic generation, structured representation, and their interaction. Building on this perspective, the following section analyzes the specific mechanisms underlying these biases and proposes a unified framework for mitigating them within LLM-augmented pipelines.

3. Bias in LLM-Knowledge Graph Integration

Figure 1 illustrates a conceptual framework for understanding bias propagation in LLM-augmented knowledge graph pipelines. The process starts with input knowledge sources, which are processed by the LLM layer, where model-level biases such as social bias, representational bias, hallucination/factual bias, prompt sensitivity, and domain coverage bias may arise.

The extracted information then flows through the pipeline interaction layer, encompassing tasks such as entity extraction, relation generation, graph completion, and reasoning. Biases introduced at the model level can propagate through these stages, influencing the selection and inference of entities and relations. These effects manifest in the knowledge graph layer as distortions, including structural bias, content bias, edge imbalance, and skewed node distributions. Finally, the resulting knowledge graph is used in downstream applications such as question answering, decision-making, and recommendation, where accumulated biases can affect the accuracy, fairness, and reliability of system outputs.

Illustrative Example of Bias Propagation: To clarify how biases propagate in LLM–KG pipelines, consider a simplified biomedical example:

“Drug X significantly reduces symptoms of Disease Y in clinical trials.”

During entity extraction, the LLM identifies the entities Drug X and Disease Y and generates the relation treats. However, due to hallucination/factual bias, the model may generate an additional unsupported triple: "(Drug X, cures, Disease Y)" even when the source text indicates only symptom reduction.

If such a hallucinated triple is incorporated during knowledge graph construction, it introduces an incorrect semantic relation that overstates the underlying evidence. Over time, multiple hallucinated relations can distort the graph structure, leading to relation imbalances or incorrect connectivity between entities. The impact becomes visible in downstream tasks. For example, a question-answering system querying the KG with "Does Drug X cure Disease Y?" may incorrectly return a positive answer due to the hallucinated relation stored in the graph. In general, quantitative metrics can detect these distortions:

• Unsupported Triple Ratio (UTR): Fraction of generated triples that cannot be verified against source evidence or external knowledge bases.
• Edge Imbalance Ratio (EIR): Measures disproportionate growth of specific relations in the graph caused by hallucinated outputs.

  $$\mathrm{EIR}\left(r\right)={\displaystyle \frac{\frac{E_r^{\mathrm{generated}}}{{\left|E\right|}^{\mathrm{generated}}}}{\frac{E_r^{\mathrm{baseline}}}{{\left|E\right|}^{\mathrm{baseline}}}}}$$

  (2)

  where $E_r$ denotes the number of edges of relation r, and $\left|E\right|$ denotes the total number of edges in the graph. The superscripts generated and baseline refer to the generated graph and the reference graph, respectively.
  Interpretation: $\mathrm{EIR}\left(r\right)>1$ indicates overrepresentation of relation r (potential hallucination), $\mathrm{EIR}\left(r\right)\approx 1$ indicates balanced distribution, and $\mathrm{EIR}\left(r\right)<1$ indicates underrepresentation.

Large Language Models (LLMs) can enhance contextual reasoning in knowledge graph (KG) construction; however, they also introduce systematic biases that may affect the completeness and reliability of generated graphs. These biases can arise from prompt design, token ordering, domain-knowledge imbalance, or training-data distributions. Among these, positional bias and prompt variation bias are particularly relevant when LLMs are used for entity extraction and relation generation. Prior studies have shown that transformer-based models may prioritize information at the beginning or end of sequences (the lost-in-the-middle phenomenon) [27,28,29] and exhibit sensitivity to prompt phrasing, leading to inconsistent outputs and biased representations in KGs. The following subsections discuss these biases and present quantitative metrics for evaluating their impact on KG construction. Figure 2 illustrates how biases introduced during LLM generation can propagate through successive KG construction stages and become structurally amplified in the resulting graph.

3.1. Hallucination and Factual Bias

Hallucination refers to the tendency of LLMs to generate syntactically fluent but factually incorrect or unverifiable information. In knowledge graph construction, this can lead to false relations, unsupported triples, or omitted entities, thereby directly compromising graph reliability. The impact of hallucinations can be quantified using metrics such as the Unsupported Triple Ratio (UTR), which measures the fraction of triples in the KG that cannot be verified against external knowledge sources [12,13].

Unsupported Triple Ratio (UTR): To quantify hallucination bias in LLM-augmented knowledge graphs, we measure the fraction of triples that cannot be verified against reliable external knowledge sources. The Unsupported Triple Ratio is defined as:

$$UTR={\displaystyle \frac{T_{\mathrm{unsupported}}}{T_{\mathrm{total}}}}$$

(3)

3.2. Positional Bias

In the context of knowledge graph (KG) construction, positional bias can influence which entities or triples are extracted from textual inputs. Because LLMs tend to assign uneven attention across input positions, entities appearing earlier or later in the sequence may be more likely to be selected, while those in intermediate positions may be underrepresented. This can lead to incomplete or imbalanced graph structures, where certain nodes or relationships are disproportionately represented. Because this bias is closely related to token ordering and attention allocation, it can be evaluated using several quantitative metrics [30,31].

Attention-Weight Distributions: Transformer-based LLMs assign attention scores to each token in the input. By visualizing or averaging these attention weights across multiple inputs, one can identify whether tokens at the beginning or end consistently receive more attention than those in the middle. For example, if early and late tokens have systematically higher average attention scores, this confirms positional bias.

Selection Frequency by Position: In tasks such as KG construction, we can count how often entities or triples are selected based on their positions in the input sequence. A higher selection frequency for early or late items indicates positional bias. This method is task-specific and directly relates the bias to KG outputs [29].

Position Bias Score (PBS): PBS quantifies positional bias in knowledge graph (KG) construction by weighting each entity’s contribution according to its position in the input sequence:

$$PBS={\displaystyle \frac{1}{N}}\sum _{i=1}^{N}w\left(i\right)·s_i$$

(4)

$$w\left(i\right)=\left\{\begin{array}{cc}\alpha ,\hfill & i\le k\hfill \\ \beta ,\hfill & k<i<N-k\hfill \\ \alpha ,\hfill & i\ge N-k\hfill \end{array}\right.,\mathrm{with}\alpha >\beta$$

(5)

Here, $s_i$ represents the importance or selection score of the i-th entity, and $w\left(i\right)$ assigns higher weights to entities at the beginning and end of the sequence.

Example in KG construction; consider the text: “Aspirin treats inflammation. Ibuprofen reduces pain. Naproxen reduces inflammation.” An LLM might preferentially extract triples from the first and last sentences:

• (Aspirin–treats–inflammation)
• (Naproxen–reduces–inflammation)

while the middle entity (Ibuprofen–reduces–pain) could be overlooked. PBS provides a quantitative measure of how skewed the KG is toward entities in certain positions, highlighting positional bias in the extracted triples.

3.3. Prompt Variation Bias

Prompt sensitivity refers to the tendency of LLMs to generate inconsistent outputs based on minor variations in prompt phrasing. Such variations can cause overrepresentation of certain entities and omission of others, potentially leading to an incomplete or biased KG [15,32]. Figure 3 illustrates an example of bias propagation in LLM-based knowledge graph construction.

Measuring Prompt Variation Bias

Quantifying prompt variation bias is more challenging than positional bias because even semantically equivalent prompts can yield substantially different outputs, affecting the consistency, reliability, and fairness of KG construction [32]. Key metrics include:

• Output Consistency Metrics: Evaluate similarity across multiple semantically equivalent prompts using measures such as BLEU, ROUGE, or embedding-based scores like BERTScore [33,34]. Large discrepancies indicate strong prompt sensitivity.
• Entity Coverage Variance: Measures the variability in which entities or triples are extracted across different prompts. High variance suggests inconsistent KG representation and potential underrepresentation of critical concepts [32].
• Alignment Stability Score (ASS): Provides a single measure of output stability across prompt variants:

  $$ASS=1-{\displaystyle \frac{1}{N}}\sum _{i=1}^N{\displaystyle \frac{|O_i▵O_{ref}|}{|O_{ref}|}}$$

  (6)
  Here, $O_i$ is the output set from the i-th prompt variant, $O_{ref}$ is a reference output set, N is the number of prompt variants, and $▵$ denotes the symmetric difference between sets. Higher ASS values indicate more stable outputs, reflecting reduced prompt variation bias [32]. Example in KG construction: Take two semantically equivalent prompts:

  −

  Prompt A output: {(Aspirin–treats–inflammation), (Ibuprofen–reduces–pain)}

  −

  Prompt B output: {(Aspirin–treats–inflammation), (Naproxen–reduces–inflammation)}

  The low overlap results in a lower ASS, indicating that KG extraction is sensitive to prompt formulation. Strategies such as prompt ensembling or multi-prompt aggregation can mitigate this instability.

By leveraging these metrics and examples, one can evaluate the impact of positional and prompt biases on KG outputs, identify weaknesses in model behavior, and implement methods to improve robustness, consistency, and reliability in LLM-augmented knowledge graphs.

Bias Metrics Clarification: The Position Bias Score (PBS) and Alignment Stability Score (ASS) used in this work are adapted conceptual metrics tailored to LLM-augmented knowledge graph pipelines. While positional and prompt-related biases have been studied in prior work, PBS and ASS are proposed here to quantify these effects in the specific context of KG extraction and completion. PBS captures the influence of input position on extracted triples (parameters $\alpha ,\beta ,k$ reflect observed attention patterns), and ASS measures output stability across semantically equivalent prompts. These formulations are interpretable and suitable for the LLM–KG pipeline, with parameter settings detailed in later sections.

3.4. Domain Knowledge Bias

Domain knowledge bias arises because Large Language Models (LLMs) are pretrained on extensive, general-purpose corpora that emphasize common, widely occurring concepts. While this training provides wide coverage across everyday language and general knowledge, it also leads models to underrepresent or misinterpret specialized, technical, or niche information [2,35,36]. This limitation manifests because the internal representations LLMs learn reflect the statistical distributions of their training data, which often lack sufficient examples of domain-specific entities, relationships, or reasoning patterns.

In KG construction, domain knowledge bias can result in knowledge graphs that omit critical domain details, fail to capture fine-grained relational structures, or substitute closely related but semantically incorrect concepts for the true domain entities because it arises from LLMs’ sensitivity to The precision and applicability of the KG in expert-driven or specialized contexts, such as biomedical, legal, or technical knowledge domains, where understanding is essential [2,36]. This bias is not unique to KG construction: surveys and empirical analyses evaluating the robustness of LLM-augmented KG systems and developing specific tasks often produce inaccurate, incomplete, or outdated outputs when the task requires specialized knowledge not well represented in their training data [36]. Addressing domain knowledge bias often requires domain-adaptation techniques, such as fine-tuning on specialized corpora, integrating structured domain knowledge sources, or leveraging domain knowledge graphs to ground and constrain the model’s outputs [36,37]. However, these methods can be resource-intensive, depend on the availability of high-quality domain datasets, and are themselves subject to challenges such as overfitting or limited generalization across subdomains.

Domain knowledge bias refers to the systematic underrepresentation or misinterpretation of specialized domain entities, relations, or concepts due to imbalanced coverage of domain-specific knowledge in the training data of large language models.

Domain Coverage Ratio (DCR): To quantify domain knowledge bias in LLM-augmented knowledge graph construction, we measure the proportion of generated triples that belong to the target domain relative to the expected domain coverage. The Domain Coverage Ratio is defined as:

$$DCR={\displaystyle \frac{T_{domain}}{T_{total}}}$$

(7)

where $T_{domain}$ denotes the number of generated triples related to the target domain and $T_{total}$ represents the total number of generated triples. Lower DCR values indicate insufficient domain representation and greater domain-knowledge bias in the constructed knowledge graph.

3.5. Social Bias

Social bias in Large Language Models (LLMs) stems from the societal, cultural, and historical imbalances embedded in their training corpora, which often reflect stereotypes, power asymmetries, and unequal representation across demographic groups. When LLMs are used for knowledge graph (KG) construction or augmentation, these biases can propagate into the resulting graphs, overrepresenting dominant social groups and marginalizing or mischaracterizing minority populations. This may manifest as skewed entity prominence, biased attribute assignments, or distorted relational structures that reinforce existing societal inequalities. Consequently, LLM-augmented KGs risk encoding discriminatory patterns as factual knowledge, undermining fairness, trustworthiness, and downstream decision-making in sensitive applications such as healthcare, hiring, and policy analysis [38,39,40].

Social bias refers to systematic differences in how demographic or social groups are represented in the entities, attributes, or relations generated by an LLM. In the context of knowledge graph construction, social bias may lead to disproportionate representation of certain groups or the reinforcement of stereotypical associations within the graph structure.

Representation Disparity (RD): To quantify social bias in LLM-augmented knowledge graphs, we measure disparities between the distribution of entities associated with demographic groups in the generated graph and a reference distribution. The Representation Disparity metric is defined as:

$$RD={\displaystyle \frac{|p_g-p_{ref}|}{p_{ref}}}$$

(8)

where $p_g$ denotes the proportion of entities or relations associated with demographic group g in the generated knowledge graph, and $p_{ref}$ represents the corresponding proportion in a reference dataset or corpus. Higher RD values indicate stronger representational imbalance across social groups.

3.6. Practical Guidance for Bias Mitigation in LLM-Augmented KG Pipelines

Based on the proposed taxonomy of biases,

Table 4. Actionable guidance for mitigating biases in LLM-augmented knowledge graph pipelines in real-world applications.

Bias Type	Mitigation Strategies	Feasibility/Considerations
Positional Bias	Prompt ensembling, multi-order input sequences, attention calibration	Low-cost to moderate; requires multiple LLM calls and careful prompt design
Prompt Variation Bias	Multi-prompt aggregation, alignment stability monitoring, post-generation validation	Moderate; increases computational cost; may require manual selection of reference outputs
Social/Representational Bias	Demographic-aware prompts, human-in-the-loop auditing, diversity-aware KG curation	High cost and data sensitivity; feasible for high-stakes domains (healthcare, policy)
Hallucination/Factual Bias	Retrieval-Augmented Generation (RAG), external KG grounding, fact-checking pipelines, human verification	Requires integration with external knowledge sources; higher latency and resource requirements
Domain Coverage Bias	Domain-specific fine-tuning, KG grounding, inclusion of specialized corpora	Dependent on availability of high-quality domain data; may require significant preprocessing
Structural/Content Bias	Schema constraints, graph validation rules, post-processing for node/edge balance	Generally feasible; can be automated, but may require domain expertise for rule design

maps each bias type to actionable strategies for practitioners, along with feasibility considerations for real-world deployments.

These strategies provide practitioners with concrete approaches to detect, mitigate, or correct bias at different stages of the LLM-KG pipeline. While some interventions (e.g., schema constraints or post-processing) are low-cost and broadly applicable, others (e.g., human-in-the-loop auditing or RAG grounding) require additional resources, access to high-quality external knowledge, and appropriate governance frameworks to ensure reliability, fairness, and reproducibility in real-world deployments.

3.7. Comparative Analysis of Bias Types

Comparing the biases introduced earlier, positional bias, prompt variation bias, and domain knowledge bias emerge at different stages of the LLM-augmented knowledge graph (KG) pipeline, yet collectively shape the quality, reliability, and fairness of the resulting graphs. Positional bias, driven by input order and attention distribution, determines where the model focuses within the input. In this type of bias, entities appearing at the start or end of a sequence may still be overrepresented, leading to a skewed graph. However, it is easier to identify than the other biases and can be mitigated by techniques such as input reordering, windowing, or aggregating predictions across multiple input permutations. On the other hand, prompt variation bias governs how instructions are interpreted. This flexibility allows the model to adapt to a range of instructions, which is useful in exploratory or interactive settings, but the same sensitivity can lead to inconsistent outputs, reducing reproducibility and reliability [7,32]. Domain knowledge bias constrains what the model can generate. Since LLMs are trained primarily on general-domain corpora, they tend to favor common concepts while underrepresenting rare or domain-specific knowledge, often resulting in incomplete or distorted KGs [35,36].

Compared with positional and prompt variation biases, which affect how information is selected and generated, domain knowledge bias affects the content the model can represent. While process-related biases can often be mitigated through pipeline-level strategies such as multiple prompt formulations, multi-round predictions, input shuffling, and output aggregation, content-related biases pose a deeper challenge, requiring domain-specific corpora, expert knowledge, or retrieval-based augmentation to ensure completeness. Addressing these biases jointly is therefore essential for constructing robust and trustworthy LLM-augmented knowledge graphs.

To complement the proposed taxonomy,

Table 5. Summary of main bias types in LLM-augmented knowledge graph pipelines, their locations, and example quantitative metrics.

Bias Type	Pipeline Layer	Metric
Positional Bias	LLM Layer	Position Bias Score (PBS), selection frequency by input position
Prompt Variation Bias	Pipeline Interaction Layer	Alignment Stability Score (ASS), BLEU, ROUGE
Social Bias	LLM Layer	Representation Disparity metrics
Hallucination/Factual Bias	LLM Layer	Unsupported Triple Ratio (UTR)
Domain Coverage Bias	LLM Layer	Domain Coverage Ratio (DCR)
Structural Bias	Knowledge Graph Layer	Edge Imbalance Ratio (EIR), Node Distribution Skew (NDS)
Content Bias	Knowledge Graph Layer	Semantic similarity deviation, triple uniqueness

summarizes the main bias types, their corresponding locations in the LLM–KG pipeline, and example quantitative metrics that can be used.

4. Literature Review

To ensure a structured and comprehensive review, we conducted a targeted search across major academic databases, including IEEE Xplore, ACM Digital Library, and Google Scholar. The search focused on publications from 2019 to 2024 using keywords such as “large language models”, “knowledge graphs”, “bias in LLMs”, “LLM–KG integration”, “entity extraction”, and “knowledge graph completion”. Papers were screened based on their relevance to bias analysis in LLMs, knowledge graph construction, or hybrid LLM–KG systems. Selected studies were then categorized according to the proposed taxonomy, mapping each work to specific bias types and stages of the LLM–KG pipeline.

In LLM-augmented Knowledge Graphs, bias originates from the nature of LLMs, which generate knowledge based on patterns embedded in their training corpora. These corpora might contain social, cultural, and domain-specific biases that are then leveraged in model outputs. Furthermore, LLMs lack source attribution and fact verification, making them prone to generating hallucinations that will later be integrated into the graph. This hallucinated or biased information can compromise the objectivity and trustworthiness of the resulting knowledge graph, thereby reducing its reliability for reasoning and decision-making tasks.

To address this problem, integrating the Chain-of-Thought (CoT) approach during KG creation or enrichment helps reduce bias by ensuring that extracted knowledge is derived through explicit and transparent reasoning steps rather than opaque model outputs. In addition, axiom- and logic-based quality control mechanisms apply domain-specific rules to detect contradictions and eliminate spurious triples, further mitigating hallucinations [41].

Building on reasoning-based approaches such as Chain-of-Thought (CoT), other frameworks have explored structured reasoning to enhance interpretability and reduce hallucinations in LLM-driven KG systems. One such framework, MemQ, as shown in Figure 4, improves LLM-based knowledge graph question answering (KGQA) by separating reasoning from tool invocation through a memory-augmented design. The method first decomposes complex queries into smaller, semantically meaningful statements. Each statement is paired with a natural-language description generated by the LLM, and these statement–description pairs are stored in a query memory, forming a reusable, human-interpretable knowledge base for future reasoning. Instead of directly generating SPARQL queries, MemQ prompts the LLM to produce clear, step-by-step reasoning traces in natural language. Each reasoning step is then aligned with relevant examples stored in the query memory by measuring semantic similarity using Sentence-BERT, allowing the model to retrieve the most contextually appropriate statements for query reconstruction. These reasoning steps are matched to relevant memory items based on semantic similarity using Sentence-BERT, and the retrieved statements are then combined to reconstruct the final SPARQL query for answer retrieval. By leveraging query memory and emphasizing transparent reasoning, MemQ reduces hallucinated tool calls, improves interpretability, and delivers more reliable KGQA performance [42].

While MemQ focuses on enhancing reasoning and transparency in question answering, other studies have explored how LLMs can support the structural organization of knowledge graphs through semi-automatic ontology construction. In one such work, Large Language Models (LLMs) are applied to facilitate ontology creation for knowledge graphs. The pipeline begins with competency questions (CQs) Figure 5, followed by ontology development and KG construction. LLMs extract concepts, entities, and relationships from scholarly publications, while a judge LLM evaluates the outputs against ground truth to ensure quality. Ontology construction in LLM-augmented KGs is guided by prompt-based instructions, where designed prompts and exemplar outputs direct the model in generating entities, relations, and attributes. To ensure that the ontology supports domain-specific reasoning, competency questions are employed. Answers to these CQs are obtained through a Retrieval-Augmented Generation (RAG) process, which combines retrieved knowledge and triples with generative inference.

Apart from ontology-focused methods, other approaches have explored LLM-based augmentation to improve the quality and coverage of knowledge graphs. For instance, the GAL-KARS framework guides LLMs with task-specific prompts to infer missing item attributes and user preferences from product descriptions and user interactions, improving their completeness and reliability in recommender systems. This enrichment produces a more comprehensive and structured knowledge base, enabling the model to better capture item–user relationships and reducing the risk of missing critical information. Moreover, the LLM-based augmentation process is guided by carefully crafted prompts that constrain generation within the contextual boundaries of the existing data, thereby minimizing hallucinations and ensuring that only plausible and consistent knowledge is incorporated [43]. Despite these advances, the approach is prone to prompt-sensitivity bias, in which small variations in the prompt can produce inconsistent outputs that focus on the entities the prompt emphasizes and ignore those that are less emphasized. This will result in biased and incomplete knowledge graphs, with some important concepts missing.

Furthermore, domain-specific biases exist in LLM-Augmented Knowledge graphs, where LLMs may overprioritize general concepts from their pre-trained knowledge over specialized and domain-specific knowledge. These biases can lead to inaccurate KGs in a specific domain, highlighting the importance of human-in-the-loop validation to ensure accuracy, fairness, and comprehensive coverage of specialized entities and relations [44].

One framework that addresses both positional bias and prompt-sensitivity bias is LLM-Align [45]. This framework, shown in Figure 6, proposes an Entity Alignment (EA) approach that leverages Large Language Models (LLMs) to improve entity matching across knowledge graphs by incorporating richer semantic context. Entity Alignment is the task of identifying and linking equivalent entities that may appear under different names or structural representations across multiple knowledge graphs (KGs). By utilizing the contextual reasoning capabilities of LLMs, the framework aims to enhance the accuracy and robustness of entity matching.

While previous EA methods often relied on carefully designed prompts and exemplar outputs to guide the model toward relevant attributes and relationships, LLM-Align leverages the reasoning capabilities of Large Language Models to infer alignments more effectively. This allows the model to focus on meaningful semantic attributes and relations rather than surface-level textual similarities. To further improve reliability, LLM-Align employs a multi-round voting mechanism in which candidate entities are presented to the LLM in different orders across multiple rounds. The predictions from each round are then aggregated, reinforcing consistent alignments while reducing errors caused by hallucinations or positional bias, improving the robustness of entity alignment results [45].

One persistent challenge in maintaining Knowledge Graphs (KGs) is ensuring that updates remain aligned with real-world knowledge demands. Previous approaches, such as information extraction and link prediction, often generate incomplete or misaligned triples that fail to capture the underlying relationships. Knowledge required for query answering is subsequently validated through retrieval from external sources using retrieval-augmented generation (RAG). After the verification process, these triples were incorporated into the KG, improving coverage and preserving factual accuracy. This process reduces hallucinations, enhances alignment with user needs, and strengthens the reliability of KGs in downstream applications such as knowledge-based question answering.

5. Evaluation Metrics

Evaluating the quality of knowledge graphs (KGs) generated or augmented by Large Language Models (LLMs) is inherently challenging. Traditional evaluation frameworks mainly rely on exact-match heuristics, where predicted $(h,r,t)$ triples are compared against ground-truth references using strict string equality. Although such rule-based methods are straightforward to implement, they are rigid and often underestimate performance by penalizing semantically equivalent triples that differ only in surface form (e.g., “AI” vs. “Artificial Intelligence”, or “located in” vs. “based in”) [46].

Recent studies have highlighted that this limitation becomes pronounced in LLM-augmented KG construction, where linguistic variation and paraphrasing are intrinsic to the generation process. To mitigate this issue, prior work has adopted hybrid evaluation strategies that move beyond exact lexical matching by integrating semantic similarity measures with adapted soft string-matching metrics. For example, the PiVe framework introduced by Huang et al. [47] combines one embedding-based semantic metric with two flexible lexical metrics to capture both conceptual alignment and surface-level fidelity in generated graphs. Surface-level fidelity refers to the extent to which the generated triples preserve the terminology, phrasing, and lexical structure of the reference knowledge, which is mainly important in terminology-sensitive or domain-specific settings.

5.1. Semantic-Level Alignment: G-BERTScore (G-BS)

To account for the high linguistic variability in LLM outputs, prior studies employ G-BERTScore (G-BS), an embedding-based metric that measures semantic similarity between predicted and reference triples. Different from traditional n-gram matching metrics, G-BERTScore leverages contextualized representations to compute cosine similarity, enabling robust comparison between semantically equivalent triples expressed using different wording [48].

Let $T_p$ denote the set of predicted triples and $T_{gt}$ the set of ground-truth triples. Each triple is linearized into a short textual representation and embedded using a pretrained language model. For each predicted triple, its maximum semantic similarity with any ground-truth triple is computed, and vice versa, resulting in graph-level precision and recall, as defined in (9) and (10), respectively:

$${\mathrm{Precision}}_{\mathrm{G}-\mathrm{BS}}={\displaystyle \frac{1}{|T_p|}}\sum _{t_p\in T_p}\underset{t_{gt}\in T_{gt}}{max}\mathrm{sim}(t_p,t_{gt})$$

(9)

$${\mathrm{Recall}}_{\mathrm{G}-\mathrm{BS}}={\displaystyle \frac{1}{|T_{gt}|}}\sum _{t_{gt}\in T_{gt}}\underset{t_p\in T_p}{max}\mathrm{sim}(t_{gt},t_p)$$

(10)

The harmonic mean of these two quantities results in the F1 score, defined in (11):

$${\mathrm{F}1}_{\mathrm{G}-\mathrm{BS}}={\displaystyle \frac{2·{\mathrm{Precision}}_{\mathrm{G}-\mathrm{BS}}·{\mathrm{Recall}}_{\mathrm{G}-\mathrm{BS}}}{{\mathrm{Precision}}_{\mathrm{G}-\mathrm{BS}}+{\mathrm{Recall}}_{\mathrm{G}-\mathrm{BS}}}}$$

(11)

This formulation implicitly performs a soft alignment between predicted and reference triples by matching each triple to its most semantically similar counterpart. As a result, G-BERTScore provides a robust indicator of conceptual consistency at the graph level, even in the presence of paraphrasing or minor lexical variation. While G-BERTScore serves as a primary measure for the semantic alignment, its utility extends to exposing deeper representational asymmetries within LLM-synthesized knowledge graphs, for instance, in a scenario where triples linked to marginalized entity groups or rare relation categories yield consistently depressed alignment scores relative to the dataset’s dominant concepts. Such a disparity acts as a proxy indicator for representational skew, suggesting that the model’s knowledge extraction is fundamentally uneven. By scrutinizing these score fluctuations across distinct entity types or subgraph clusters, researchers gain a diagnostic lens through which to pinpoint, and eventually mitigate, the semantic gaps introduced during the LLM-driven extraction process.

5.2. Flexible Lexical Matching: G-BLEU and G-ROUGE

While semantic similarity is essential for assessing conceptual correctness, maintaining a degree of lexical fidelity remains important, especially for domain-specific knowledge graphs where precise terminology is required [49]. Accordingly, prior work complements semantic evaluation with adapted versions of BLEU and ROUGE, denoted as G-BLEU (G-BL) and G-ROUGE (G-RO), which operate on the linearized representations of KG triples [47].

G-BLEU measures surface-level precision by evaluating the overlap of generated n-grams with those appearing in the reference triples, thereby penalizing unsupported or extraneous lexical content. The metric is defined in (12):

$$\mathrm{G}-\mathrm{BLEU}={\displaystyle \frac{{\sum }_{n=1}^N{\sum }_{g\in G_p}min({\mathrm{count}}_n\left(g\right),{\mathrm{count}}_n\left(G_{gt}\right))}{{\sum }_{n=1}^N{\sum }_{g\in G_p}{\mathrm{count}}_n\left(g\right)}}$$

(12)

In contrast, G-ROUGE focuses on recall, quantifying the extent to which the ground-truth lexical content is successfully covered by the generated triples, as defined in (13):

$$\mathrm{G}-\mathrm{ROUGE}={\displaystyle \frac{{\sum }_{g\in G_{gt}}min(\mathrm{count}\left(g\right),\mathrm{count}\left(G_p\right))}{{\sum }_{g\in G_{gt}}\mathrm{count}\left(g\right)}}$$

(13)

G-BLEU and G-ROUGE provide complementary perspectives on lexical accuracy and coverage, capturing how faithfully the generated graph preserves the terminology and phrasing of the reference knowledge. Moreover, G-BLEU and G-ROUGE serve as critical indicators of lexical and terminological bias. By disaggregating these scores across specific entity clusters or thematic domains, researchers can identify systematic discrepancies in how the model handles domain-specific nomenclature. For instance, consistently depressed G-BLEU scores within marginalized categories may reveal a failure to maintain lexical consistency or the use of imprecise, non-standard terminology. Because these metrics are highly sensitive to surface-level token overlaps, they provide a quantitative signature of lexical skew that semantic-heavy metrics like G-BERTScore might overlook.

5.3. Additional Evaluation Considerations for LLM-Augmented Knowledge Graphs

Beyond semantic and lexical alignment, recent literature emphasizes that evaluating LLM-augmented knowledge graphs requires additional considerations related to factuality, structural quality, and robustness. In particular, hallucination remains a central concern [50], as LLMs may generate triples that are syntactically plausible and semantically coherent yet factually incorrect. To address this issue, several studies propose factuality-oriented analyses that assess whether newly generated triples can be verified against trusted sources, existing graph constraints, or external knowledge bases. For example, recent work on knowledge graph–based demonstrates how KGs can be used to detect and mitigate hallucinated content in LLM outputs by enforcing factual consistency during post-generation validation [51].

Structural properties of the generated graph also play an important role in evaluation. In augmentation settings, it is necessary to distinguish between genuinely novel knowledge and redundant paraphrases of existing facts. Metrics that quantify triple uniqueness or redundancy have therefore been discussed as complementary indicators of graph quality, particularly in large-scale or iterative augmentation scenarios. Additionally, in bias-sensitive applications, evaluation extends beyond static accuracy to include consistency under prompt variation, where significant fluctuations in extracted triples across semantically equivalent prompts may indicate prompt-induced bias. Finally, human-in-the-loop evaluation is frequently used as a complementary validation mechanism, enabling expert assessment of factual accuracy, redundancy, and subtle biases that may not be fully captured by automated metrics alone.

In addition to lexical and semantic alignment, structural bias metrics can be used to quantify imbalances in the topology of generated knowledge graphs. A commonly discussed indicator is the Popularity Bias Score (PopBS), which measures the extent to which graph connectivity is concentrated among highly connected entities [52]. Assume $d\left(v\right)$ refers to the degree of node v, and let $V_h$ represent the set of high-degree nodes. PopBS can be defined as:

$$PopBS={\displaystyle \frac{{\sum }_{v\in V_h}d\left(v\right)}{{\sum }_{v\in V}d\left(v\right)}}$$

(14)

Higher PBS values indicate that a disproportionate share of graph connectivity is concentrated among a small subset of entities, which reflects the structural popularity bias that may disadvantage long-tail nodes. These indicators are useful when evaluating LLM-augmented KG construction pipelines, where entity extraction and relation generation may amplify existing structural imbalances in the graph. For instance, consider a Knowledge Graph (KG) derived from biomedical literature. In this context, nodes like “COVID-19” or “SARS-CoV-2” often dominate the topology, appearing in a vast majority of triples. This leaves the remaining entities on the periphery, sparsely connected by comparison. Under these conditions, the PBS value naturally climbs. This spike occurs because the graph’s connectivity is heavily skewed toward a narrow cohort of high-degree ‘hubs’, a clear signature of structural popularity bias within the synthesized graph.

6. Datasets for Bias Analysis in LLMs and Knowledge Graphs

Investigating bias in LLMs and KGs requires reliable benchmarks, as these issues often stem from imbalanced data, gaps in domain-specific coverage, or social biases in the original training sets. Because bias can manifest in different ways, researchers have begun using a variety of datasets designed to address specific problems. These range from examining how LLMs represent certain groups in text generation to identifying structural and relational gaps within knowledge graphs [53].

Table 6. Representative datasets used to evaluate bias in LLMs, knowledge graphs, and LLM–KG integrated systems.

Dataset	System Type	Primary Bias Type	Evaluation Signal	Domain
LLM Bias Benchmarks
CrowS-Pairs [54]	LLM	Social/Stereotypical	Stereotype preference score	General
StereoSet [55]	LLM	Representational	Bias score (intra-/inter-sentence)	General
BOLD [56]	LLM	Demographic/Ideological	Bias in generated text	General
BBQ [57]	LLM	Social/Contextual	QA answer distribution bias	General
Knowledge Graph Bias Benchmarks and Sources
FB15k-237 [58]	KG	Structural/Popularity	Link prediction disparity	General
WN18RR [59]	KG	Relational/Hierarchical	Relational consistency	Lexical
YAGO [60]	KG	Taxonomic/Coverage	Ontology coverage imbalance	Multilingual
DBpedia [61]	KG	Coverage/Representational	Entity frequency imbalance	Multilingual
LLM–KG Integrated Benchmarks
HaluEval [62]	LLM–KG	Hallucination/Factuality	Response-level factual verification	General
MultiHal [63]	LLM–KG	Hallucination/Grounding	KG grounding consistency	Multilingual
PrimeKG [64]	LLM–KG	Domain coverage	Biomedical KG completion	Biomedical
PubMed corpus [65]	LLM–KG	Domain-specific	Scientific entity and relation extraction	Biomedical
Multimodal and Algorithmic Bias Benchmarks
CelebA [66]	Multimodal	Representational	Attribute bias across facial features	Vision
COMPAS [67]	Tabular	Historical/Demographic	Fairness disparity across groups	Criminal justice

summarizes representative datasets used for bias evaluation across LLMs, knowledge graphs, and hybrid systems.

6.1. Benchmarks for LLM Social Bias

Research into LLM bias has mostly centred on detecting social issues, stereotypes, and factual errors in model outputs. The datasets used for these analyses are typically task-oriented and built to test how a model behaves when given specific prompts or scenarios.

• CrowS-Pairs [54]: This dataset is widely used to measure how much models favor stereotypical over anti-stereotypical statements across nine categories, including race, gender, and religion.
• StereoSet [55]: Similar to CrowS-Pairs, this benchmark uses crowd-sourced sentence pairs to evaluate both intra-sentence and inter-sentence biases.
• BOLD (Bias in Open-Ended Language Generation) [56]: This dataset focuses on measuring fairness when models generate text about different professions, religions, and political ideologies.
• BBQ (Bias Benchmark for QA) [57]: This benchmark targets social biases in question answering tasks. It is particularly useful for measuring how LLMs rely on stereotypes when the provided context is ambiguous, covering eleven social categories such as age, disability status, and sexual orientation.

6.2. Benchmarks for Structural Bias in Knowledge Graphs

Bias within knowledge graphs appears as structural imbalances where certain entities or relations are overrepresented compared to others. Such datasets are frequently derived from large-scale knowledge bases or domain-specific graphs and are used to study phenomena such as coverage bias, popularity bias, and relational imbalance.

• FB15k-237 [58]: Originally a subset of Freebase, this is a standard benchmark used to evaluate how well models can predict missing links without relying on simple patterns like inverse relations.
• WN18RR [59]: Derived from WordNet [68], this dataset is used to test models on hierarchical and semantic relations while ensuring that the evaluation is not skewed by data leakage.
• YAGO [60]: As a large-scale multilingual knowledge base derived from Wikipedia and WordNet, YAGO is often used to study taxonomic bias and the uneven distribution of attributes across different cultural or geographical entities.
• DBpedia [61]: This dataset serves as a major hub for the Linked Open Data cloud. Researchers use it to analyze coverage bias, specifically focusing on how the structured representation of information can be skewed by the crowdsourced nature of its underlying wiki sources.

6.3. Benchmarks for Bias Analysis in LLM-KG Systems

For systems that integrate LLMs and KGs, datasets must account for both factual grounding and the risk of hallucination within the graph structure. Since these integrated systems rely on the LLM to interpret or populate the KG, any bias in the model can lead to corrupted or untrustworthy knowledge representation.

• HaluEval [62]: This is a large-scale collection that includes both generated and human-annotated samples. It is designed to test whether an LLM can recognize its own hallucinations, which is vital when using these models to augment a knowledge graph with new facts.
• MultiHal [63]: This benchmark uses specific knowledge graph paths from Wikidata to verify the factuality of model responses. It is particularly useful for checking if the integration maintains consistency across different languages.
• PrimeKG [64] and PubMed corpus [65]: These resources support domain-specific evaluation of factual consistency and knowledge extraction in biomedical settings.

6.4. Benchmarks for Multimodal and Algorithmic Bias

Since the construction of semantically rich knowledge graphs often involves processing various data types, it is important to consider benchmarks that address bias outside of pure text. This is especially relevant for systems that integrate visual or tabular information into the graph structure.

• CelebA [66]: Although it is primarily a facial attributes dataset, CelebA is frequently used in multimodal research to see how models associate physical traits with social or professional labels. For an LLM tasked with generating KG entities from visual data, this benchmark helps identify representational skews that could lead to biased node attributes in the resulting graph.
• COMPAS [67]: This dataset is a foundational benchmark for studying algorithmic fairness and historical data imbalance. While it originated in risk assessment, it is used here to represent the challenges of processing tabular data. It highlights how underlying skews in training data can lead to disparate treatment of different racial or demographic groups, a critical concern when KGs are used for automated decision-making or case retrieval in engineering.

7. Research Gaps and Motivation

Despite growing interest in integrating Large Language Models (LLMs) into knowledge graph (KG) construction pipelines, existing research treats bias in a fragmented and largely surface-level manner. Prior studies have examined positional bias in long-context reasoning, prompt sensitivity in generation, and domain knowledge limitations of pretrained models. However, these biases are typically analyzed in isolation and at the level of textual output, rather than within the structured, cumulative setting of knowledge graph construction.

Crucially, little work systematically investigates how such biases propagate across pipeline stages. This begins with input- and model-level generation and then proceeds through entity extraction, relation generation, alignment, and graph completion. While positional and prompt-induced variations are known to affect output stability, their downstream structural consequences remain underexplored. There is limited quantitative analysis linking these biases to graph-level distortions such as entity omission, relation sparsity, edge imbalance, or topology skew.

Domain knowledge bias presents an additional and particularly critical challenge. In expert-driven or low-resource domains, incomplete or systematically skewed extraction can result in structurally marginalized subgraphs, ultimately degrading the quality of reasoning and decision support. Yet current mitigation strategies, primarily prompt engineering or retrieval augmentation, lack unified evaluation protocols that measure structural bias accumulation across domains.

Moreover, prevailing evaluation practices focus on output correctness, fluency, or task-level accuracy, overlooking bias-induced incompleteness, representational imbalance, and structural asymmetry within the constructed graph. To date, no comprehensive framework jointly characterizes positional, prompt-induced, and domain-specific biases and quantifies their structural manifestations and propagation throughout LLM-augmented KG pipelines.

These limitations highlight the need for a unified taxonomy of bias propagation and principled evaluation strategies that explicitly capture structural, domain, and multi-stage distortions. Addressing these challenges is essential for improving the robustness, completeness, and trustworthiness of LLM-augmented knowledge graph systems.

8. Discussion and Conclusions

In this work, we systematically analyze bias propagation in LLM-augmented knowledge graph pipelines, mainly focusing on text-driven extraction workflows while discussing extensions to multimodal inputs such as images, audio, and structured documents. We demonstrate that biases originating at the input or model level, such as prompt sensitivity, positional effects, and domain coverage limitations, propagate and amplify through pipeline stages, including entity extraction, relation generation, schema alignment, and graph completion. Crucially, prior studies suggest that these biases may manifest differently across modalities: visual and audio inputs can introduce distinct distortions, and interactions across modalities can further exacerbate structural imbalances in the knowledge graph. Our survey of existing datasets reveals persistent patterns of structural distortion, including entity omission, relation sparsity, and topology skew, which are particularly pronounced in low-resource or domain-specific knowledge graphs.

To address these challenges, we propose a unified taxonomy of bias propagation and introduce structural bias metrics that quantify distortions across modalities and pipeline stages. Our framework enables systematic comparison of LLM- and KG-centric datasets, distinguishing between surface-level generation biases and deeper structural or relational biases. By explicitly considering multimodal inputs, it provides a comprehensive perspective on how biases accumulate, transform, and amplify throughout the pipeline, underscoring the importance of pipeline-aware, cross-domain evaluation for constructing reliable knowledge graphs with LLMs.

This survey also highlights critical directions for future research. Automated mitigation strategies must account for modality-specific and cross-modal bias, while evaluation frameworks should extend to temporal and hybrid knowledge graphs. Integrating domain-specific considerations into LLM training and KG construction is essential for reducing structural distortions and improving downstream reasoning. By systematically characterizing bias propagation across stages and modalities, this work lays a foundation for more robust, fair, and trustworthy LLM-augmented knowledge graphs in real-world applications.

From a practical perspective, several mitigation strategies can be applied at different stages of the LLM–KG pipeline. For example, prompt ensembling and prompt normalization can reduce prompt sensitivity during entity extraction and relation generation. Retrieval-augmented generation (RAG) and grounding in external knowledge sources can help mitigate hallucination and domain coverage bias, while schema constraints and post-generation validation rules can limit structural distortions in the resulting knowledge graph. In real deployments, however, these techniques introduce trade-offs related to computational cost, the availability of high-quality domain knowledge sources, and governance requirements such as human-in-the-loop auditing in high-stakes applications. Future work could include case studies in real-world domains, such as healthcare or scientific literature, to observe bias propagation in practice. Such studies would enable quantitative assessment of how model-level and pipeline-level biases affect the accuracy, completeness, and reliability of the generated knowledge graphs, as well as their downstream applications in question answering, decision-making, or recommendation tasks.