PaperlayerView exact revision on arXiv

Source-linked AI summary

Survey of Hallucination in Natural Language Generation

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Delong Chen, Wenliang Dai, Ho Shu Chan, Andrea Madotto, Pascale Fung

arXiv:2202.03629v7cs.CL

TL;DR

Hallucination in neural NLG can produce fluent but misleading text, yet research had not been comprehensively synthesized across major NLG tasks. This survey organizes evaluation metrics, mitigation methods, challenges, and task-specific findings, concluding that intrinsic and extrinsic hallucinations require different mitigation strategies.

  • Problem

    Existing surveys focused on specific tasks, leaving hallucination research across major NLG tasks insufficiently understood from a broader perspective.

  • Method

    The survey synthesizes hallucination terminology, evaluation metrics, mitigation methods, contributors, and task-specific research across major NLG tasks and LLMs.

  • Results

    The survey identifies intrinsic and extrinsic hallucinations as distinct categories requiring different mitigation strategies.

  • Takeaways & Limitations

    Hallucination mitigation should use diverse strategies matched to the hallucination category and NLG task.

  • Takeaways & Limitations

    Hallucination detection and mitigation in visual-language generation remain early-stage, with limited analyses and inadequate general-purpose automatic metrics.

Abstract

from arXiv · show

Natural Language Generation (NLG) has improved exponentially in recent years thanks to the development of sequence-to-sequence deep learning technologies such as Transformer-based language models. This advancement has led to more fluent and coherent NLG, leading to improved development in downstream tasks such as abstractive summarization, dialogue generation and data-to-text generation. However, it is also apparent that deep learning based generation is prone to hallucinate unintended text, which degrades the system performance and fails to meet user expectations in many real-world scenarios. To address this issue, many studies have been presented in measuring and mitigating hallucinated texts, but these have never been reviewed in a comprehensive manner before. In this survey, we thus provide a broad overview of the research progress and challenges in the hallucination problem in NLG. The survey is organized into two parts: (1) a general overview of metrics, mitigation methods, and future directions; (2) an overview of task-specific research progress on hallucinations in the following downstream tasks, namely abstractive summarization, dialogue generation, generative question answering, data-to-text generation, machine translation, and visual-language generation; and (3) hallucinations in large language models (LLMs). This survey serves to facilitate collaborative efforts among researchers in tackling the challenge of hallucinated texts in NLG.

1 INTRODUCTION

NLG has advanced through deep-learning and Transformer-based models across major generation tasks, but generated outputs can still degenerate or hallucinate, undermining performance and creating real-world safety and privacy risks. This survey addresses the gap in broad, cross-task understanding by organizing definitions, causes, metrics, mitigation methods, and task-specific hallucination research.

  • Motivation: Transformer-based NLG supports downstream tasks including summarization, dialogue generation, generative question answering, data-to-text generation, and machine translation.The introduction presents NLG as a crucial but challenging NLP subfield accelerated by deep-learning technologies.
  • Motivation: Likelihood-maximization training and decoding can cause degeneration, producing bland, incoherent, or repetitively looping outputs.Early studies identified degeneration as a limitation of standard likelihood-maximization objectives.
  • Motivation: Hallucination threatens real-world NLG applications by degrading performance and creating medical safety and privacy risks.The passage gives examples involving hallucinatory medical summaries, machine-translated medicine instructions, and potential privacy violations.
  • Contribution: Although hallucination research is active across NLG tasks, existing surveys have not broadly unified understanding across all major tasks.The survey argues that cross-task analysis could clarify relationships among hallucinations and encourage unified research efforts.
  • Organization: The survey covers hallucination definitions, categories, contributors, metrics, mitigation methods, and task-specific research across multiple NLG applications.Its organization includes general foundations followed by sections on abstractive summarization, dialogue generation, generative question answering, and data-to-text generation, among others.

2 DEFINITIONS

In NLG, hallucination denotes generated content that is nonsensical or unfaithful to the provided source, often appearing fluent, natural, and grounded despite being difficult to verify. The survey distinguishes intrinsic hallucinations, which contradict the source, from extrinsic hallucinations, which cannot be verified from it, while noting terminology and definitions vary across tasks.

  • Core definition: Hallucinated text can appear fluent, natural, and grounded in the provided context despite being unfaithful, nonsensical, or difficult to verify.This resemblance to psychological hallucination explains the terminology and makes hallucinated text hard to identify at first glance.
  • Core definition: NLG hallucination is generated content that is nonsensical or unfaithful to the provided source content.This is the most inclusive and standard definition in NLP, although task-specific variations exist.
  • Hallucination types: Intrinsic hallucination is generated output that contradicts the source content.The survey illustrates this with a summary claiming the first Ebola vaccine was approved in 2021 when the source states 2019.
  • Hallucination types: Extrinsic hallucination is generated output that cannot be verified from the source because it is neither supported nor contradicted by it.Extrinsic hallucination is not always erroneous because it may derive from factually correct external information.
  • Terminology and task variation: Faithfulness means remaining consistent and truthful to the provided source, making it an antonym of hallucination and a target for hallucination reduction.Judgments of faithfulness and hallucination can differ across tasks; source definitions also vary among summarization, data-to-text, dialogue, GQA, and machine translation.

3 CONTRIBUTORS TO HALLUCINATION IN NLG

Hallucination in NLG primarily arises from source-reference divergence in datasets, but also from task-inherent divergence and neural-model training and modeling choices. These contributors can produce outputs that are not grounded in or faithful to the provided source.

  • Data contributors: Source-reference divergence in training data can encourage models to generate text unsupported by or unfaithful to the source.Divergence may result from heuristic data collection or from the nature of some NLG tasks.
  • Data contributors: 62% of WIKIBIO first sentences contain additional information not supported by their source infoboxes.WIKIBIO pairs Wikipedia infoboxes with the first sentences of corresponding pages as source and reference.
  • Data contributors: Duplicate pretraining examples can bias models toward repeating memorized phrases from those examples.The problem is difficult to detect manually in very large text corpora.
  • Data contributors: Some NLG tasks inherently permit source-reference divergence because they value diverse outputs, including open-domain dialogue with chit-chat, subjective responses, or relevant external facts.Such responses can improve dialogue engagingness and diversity.
  • Model contributors: Hallucination can persist despite little dataset divergence because neural-model training and modeling choices introduce additional contributors.The survey identifies imperfect representation learning, erroneous decoding, exposure bias, and parametric knowledge bias as model-related contributors.

4 METRICS MEASURING HALLUCINATION

The section surveys lexical, model-based, entailment-oriented, language-model, and human-evaluation approaches for measuring hallucination in NLG. It emphasizes that conventional metrics can miss hallucinations, while learned evaluators improve coverage but introduce model-error and transfer limitations.

  • Motivation: Conventional metrics such as ROUGE, BLEU, and METEOR are inadequate for hallucination measurement; abstractive summarization systems hallucinated content in 25% of generated summaries.These findings motivate metrics designed specifically to assess faithfulness rather than general writing quality.
  • Lexical-based metrics: Lexical metrics estimate hallucination from n-gram overlap or contradiction, while PARENT matches generated text against both source tables and target references using lexical entailment.Lexical matching is simple but cannot handle syntactic or semantic variation; PARENT combines source- and target-based evidence.
  • Model-based metrics: Model-based metrics capture complex syntactic and semantic mismatches, but errors in the underlying neural models can propagate and impair hallucination quantification.IE-based metrics simplify verifiable knowledge into relational tuples, whereas QA-based metrics compare answers generated from questions about the text.
  • Entailment-based metrics: NLI-based metrics assess whether source knowledge entails generated content and are more robust to lexical variability than token-matching approaches.However, off-the-shelf NLI models transfer poorly to abstractive summarization, sentence-level entailment may not localize errors, and QG or QA components can propagate errors.
  • Language-model metrics: Language-model metrics classify tokens as hallucinatory when an unconditional target-only model assigns lower loss than a source-and-target conditional model during forced-path decoding.The approach uses two language models to estimate whether each generated token is supported by the input.
  • Human evaluation: Because automatic hallucination evaluation remains challenging and imperfect, human evaluation commonly uses rating-based scoring or comparisons with baselines and ground-truth references.Faithfulness-specific datasets can better match hallucination judgments than generic NLI labels, which do not always distinguish unsupported information.

5 HALLUCINATION MITIGATION METHODS

Hallucination mitigation methods are divided into data-related methods and modeling/inference methods. Data methods improve faithfulness by constructing, filtering, correcting, or augmenting training inputs, while modeling methods strengthen source representation, alignment, planning, and generation control.

  • Mitigation methods fall into data-related methods and modeling and inference methods, reflecting two main contributors of hallucinations.
  • Data-related methods construct faithful datasets by writing clean targets, rewriting existing sentences, or filtering and correcting noisy or contradictory training pairs.Instance-level corpus filtering ranks hallucination scores and removes untrustworthy samples; input correction is especially applicable to data-to-text because structured data are easier to correct.
  • Input augmentation with external knowledge, explicit alignments, or extra training data improves source-target correlation and semantic understanding, but can introduce semantic and format gaps.
  • Modeling methods modify encoders, attention, and decoders to improve input representation, source focus, long-range dependencies, and constrained or uncertainty-aware generation.
  • Planning and sketching control generated content and ordering, while reinforcement learning uses rewards to optimize policies and reduce hallucinations caused by exposure bias.Planning can be implemented as a separate step or integrated end-to-end, but separate two-step generation may progressively amplify hallucinations.

6 FUTURE DIRECTIONS

Future work should develop more fine-grained, human-aligned, generalizable metrics and robust fact-checking, while improving mitigation for numerals, reasoning, long texts, extrinsic hallucinations, and controllability.

  • Fine-grained Metrics: Fine-grained metrics should locate hallucinatory substrings and distinguish intrinsic from extrinsic hallucinations for richer insight and better explainability.QA-based metrics may not identify individual hallucinatory substrings, making accurate localization an essential first step.
  • Fact-Checking: Automatic fact-checking requires trustworthy evidence retrieval beyond Wikipedia and verification models robust to adversarial attacks, negation, numerical, and comparative changes.Fact-checking comprises knowledge evidence selection and claim verification, both of which remain challenging.
  • Generalization: Task-agnostic, cross-domain-robust metrics and open-source platforms could support unified benchmarks and standardized hallucination evaluation across NLG tasks.Current source and output formats vary across downstream tasks, motivating general metrics and collaborative evaluation infrastructure.
  • Incorporation of Human Cognitive Perspective: Metrics should incorporate human cognitive priorities, such as treating named entities and numerals as especially salient, to improve correlation with human judgment.Automatic check-worthy detection is proposed as a way to approximate human-like judgments and potentially mitigate hallucination.
  • Mitigation Directions: Future mitigation should explicitly model numerals, add reasoning abilities, address long-input and self-contradictory outputs, and distinguish intrinsic from overlooked extrinsic hallucinations.Numeral correctness is important for dates, quantities, and scalars; reasoning should support both source understanding and backward faithfulness checks.
  • Controllability: Controllability should let models balance faithfulness and diversity while adapting acceptable hallucination levels to different NLG tasks.Chit-chat may permit some factual hallucinatory content, whereas the desirability of factual hallucinations in abstractive summarization remains unsettled.

7 HALLUCINATION IN ABSTRACTIVE SUMMARIZATION

In abstractive summarization, hallucination is source-unsupported content, including contradictions and unverifiable additions, and factual errors can be categorized into semantic frame, discourse, and content verifiability errors. Research addresses detection through unsupervised and semi-supervised metrics and mitigation through model, training, and post-editing approaches.

  • Definition: A summary is hallucinated when it contains spans unsupported by the input document, with intrinsic content contradicting the source and extrinsic content unverifiable from it.This definition distinguishes two hallucination types according to their relationship with source evidence.
  • Definition: Pagnoni et al. classify factual errors in summaries as semantic frame error, discourse error, or content verifiability error.Because the relevant facts derive from source knowledge, the survey treats factual errors as a hallucination subtype.
  • Automatic Evaluation: Hallucination metrics are mainly model-based and divided into unsupervised and semi-supervised categories, while automatically distinguishing intrinsic from extrinsic hallucinations remains difficult.Unsupervised metrics include information extraction-, natural language inference-, and question answering-based methods.
  • Automatic Evaluation: FactCC evaluates factual consistency using jointly trained tasks for consistency checking, supporting-span extraction, and entailment-related prediction, and its performance surpasses classifiers trained on MNLI or FEVER.FactCC is transferred from synthetic summarization data to machine-generated summaries.
  • Mitigation: Mitigation methods modify sequence-to-sequence architectures, training objectives, or generated drafts through factual correction modules to reduce hallucinated content.Reported approaches include graph-based source encoding, focus attention, contrastive learning with hallucinatory negatives, and SpanFact post-editing.

8 HALLUCINATION IN DIALOGUE GENERATION

Dialogue-generation hallucinations concern contradictions with dialogue history or external knowledge, while some non-factual generation may be acceptable in open-domain chit-chat. Research distinguishes open-domain and task-oriented settings, with emphasis on consistency, factual evaluation, mitigation, and fact-checking.

  • Task scope: Dialogue generation produces responses to user utterances and comprises task-oriented and open-domain subtasks.Responses are expected to be fluent, coherent, and consistent with dialogue history.
  • Hallucination definition: Intrinsic hallucination occurs when a response contradicts the dialogue history or external knowledge.The survey illustrates this with generated facts that conflict with inputs or cannot be verified from available information.
  • Open-domain dialogue: Open-domain dialogue faces self-inconsistency, persona inconsistency, and factual inconsistency with external sources.Knowledge-grounded dialogue provides external references, increasing active discussion of hallucination and factual consistency.
  • Evaluation: Dialogue hallucination evaluation has no standard metric, so systems are commonly assessed by humans for factual consistency or correctness.Automatic approaches include knowledge-overlap metrics such as KF1, NLI-based consistency measures, learnable evaluators, and coherence test sets.
  • Mitigation: Mitigation methods include filtering untrustworthy training samples, retrieval augmentation, control codes, and inductive attention that removes potentially uninformative links.These methods target self-consistency, knowledge grounding, evidence awareness, or dialogue modeling during generation.
  • Future directions: Fact-checking is proposed as a future direction combining verifiable claim detection with evidence retrieval from external sources.It can support both evaluation and identification of hallucination-prone dialogue.

9 HALLUCINATION IN GENERATIVE QUESTION ANSWERING

Hallucination is a major, under-explored challenge in generative question answering because answers synthesize potentially redundant, complementary, or contradictory retrieved documents. Research mainly measures faithfulness through human judgment and explores improved metrics, retrieval, and generation models.

  • Background: GQA generates abstractive answers from information retrieved across multiple documents, where redundancy, complementarity, and contradictions make hallucination common.Unlike extractive question answering, GQA targets long, explanatory answers that cannot be directly copied from existing spans.
  • Evaluation: GQA lacks a standard hallucination definition, so studies primarily use human evaluation of factual correctness and answer faithfulness as proxies.Recent work also frames the problem as semantic drift, indicating how an answer deviates from the intended meaning.
  • Evaluation: ROUGE and F1 are poor hallucination measures because their N-gram overlap has weak correlation with human judgments, motivating semantic-overlap, factual-correctness, and factual-consistency metrics.Alternative approaches include extracting facts for comparison, using QA models to test faithfulness, and measuring short-answer recall in generated long-form answers.
  • Causes: Hallucination can arise from irrelevant retrieval and from intrinsic or extrinsic errors in the conditional generation model, with both sources interconnected.This makes GQA hallucination more complicated than hallucination in tasks whose source documents are directly related to the target.
  • Mitigation and future directions: Mitigation research has explored knowledge-source selection, multi-source integration, local knowledge graphs, rationale extraction, and fine-tuning rather than merely scaling models.Future work calls for better automatic metrics, retrieval of relevant information, and generation models that synthesize accurate answers from multi-source documents.

10 HALLUCINATION IN DATA-TO-TEXT GENERATION

Data-to-text generation models hallucinate because mapping structured data to natural language can cause semantic misunderstanding and erroneous correlations, producing contradictions or unsupported extra information. Research addresses this through specialized metrics, faithful data and modeling or inference methods, while highlighting numerical handling and reasoning as future priorities.

  • Motivation: Data-to-text hallucinations arise from the gap between structured inputs and natural language, causing semantic misunderstanding and erroneous correlations.The task conditions descriptions on tables, database records, and knowledge graphs, while low fault tolerance makes errors consequential.
  • Definitions: Intrinsic hallucinations contradict input data, whereas extrinsic hallucinations add information irrelevant to or absent from the source.Examples include incorrect table-grounded facts and claims not mentioned in the source table.
  • Metrics: Hallucination measurement includes PARENT and PARENT-T, entity-centric coverage and hallucinated-ratio metrics, Data-QuestEval, NLI-based entailment, and LM-based hallucinated-token ratios.These approaches evaluate table-text alignment, entity faithfulness, question-answer consistency, entailment in both directions, or token-level conditionality.
  • Mitigation: Mitigation methods use faithful corpora, data processing, planning and skeleton generation, faithfulness-oriented losses, controllable hallucination levels, and inference-time decoding.A Multi-Branch Decoder separately controls content, hallucination, and fluency using word-level alignment labels; the methods can collaborate.
  • Future Directions: Future work should treat numbers explicitly because numerical errors frequently cause hallucinations, while reasoning can improve entity relationships and reduce errors in logical data-to-text generation.Logical tasks require inference, calculation, and comparison, and current automatic hallucination metrics do not specifically treat numbers.

11 HALLUCINATIONS IN NEURAL MACHINE TRANSLATION · 11.1 Hallucinations Definition and Categories in NMT

Neural machine translation produces fluent, human-level translations but can hallucinate content disconnected from the source, risking user misinformation. The survey distinguishes intrinsic hallucinations, which alter source information, from extrinsic hallucinations, which add unsupported content, while also noting perturbation-based categories.

  • 11 HALLUCINATIONS IN NEURAL MACHINE TRANSLATION: NMT generates target-language translations from parallel training data and is typically more fluent and human-level than SMT.This fluency increases the danger that hallucinations will misinform users.
  • 11.1 Hallucinations Definition and Categories in NMT: NMT hallucination research emerged with early deployments, when models were observed to sacrifice adequacy for fluency, especially on out-of-domain tests.Later work generally characterizes hallucinated translations as disconnected from the source.
  • 11.1 Hallucinations Definition and Categories in NMT: The survey adopts intrinsic and extrinsic hallucination categories for NMT because the literature uses overlapping terminology unlike other NLG tasks.Both categories are based on how translated output is disconnected from the source.
  • 11.1 Hallucinations Definition and Categories in NMT: Intrinsic hallucinations contain incorrect information relative to the source, such as changing “Mike” to “Jerry” or removing a source negation.The example changes “Mike” to “Jerry” and translates “doesn’t go” as “goes.”
  • 11.1 Hallucinations Definition and Categories in NMT: Extrinsic hallucinations add content without regard to the source, such as “happily” and “with his friend” in a translation about Mike visiting the bookstore.These additions are not supported by the source sentence.
  • 11.1 Hallucinations Definition and Categories in NMT: Raunak et al. alternatively divide NMT hallucinations into hallucinations under perturbations and natural hallucinations.Perturbation hallucinations occur when outputs differ drastically between perturbed and unperturbed test sets.

11.2 Hallucination Metrics in NMT

Hallucinations in neural machine translation are difficult to define and are not reliably reflected by BLEU, motivating automated statistical metrics. Proposed approaches assess adequacy or faithfulness through similarity, attention entropy, token-level prediction, overlap, and output repetition.

  • 11.2 Hallucination Metrics in NMT: Hallucinations are often identified manually because their qualitative definition is subjective, and their appearance does not affect BLEU scores.Researchers have consequently pursued statistical methods to automate and quantify hallucination detection.
  • 11.2 Hallucination Metrics in NMT: The BVSS metric identifies sentence-level adequacy loss when the reference contains more information than the MT output, or vice versa.This frames hallucination-related assessment as detecting information imbalance between reference and translation.
  • 11.2 Hallucination Metrics in NMT: Faithfulness can be evaluated with an auxiliary decoder using word-by-word translation probabilities alongside translation fluency.The evaluation module returns a loss that supports faithfulness assessment.
  • 11.2 Hallucination Metrics in NMT: Fine-grained detectors classify hallucinated tokens using synthetic noisy data, source-target embedding alignments, perturbed-sentence similarity, or bilingual synonym overlap.Overlap heuristics can miss valid paraphrases and synonyms, while similarity-based methods better accommodate diverse translations.

11.3 Hallucination Mitigation Methods in NMT · 11.4 Future Directions in NMT

NMT hallucination mitigation spans data augmentation, divergence-aware training, objective and inference changes, loss filtering, regularization, and phrase alignment. Future work emphasizes quantifying hallucinations and making detection and mitigation methods less costly and more practical.

  • 11.3 Hallucination Mitigation Methods in NMT: NMT hallucinations motivate mitigation because they are difficult for non-fluent target-language readers to detect and can cause errors or dangers.
  • 11.3 Hallucination Mitigation Methods in NMT: Data augmentation commonly mitigates hallucinations by adding perturbed sentences, with beginning-of-sentence insertion of common tokens reported as especially successful.The method requires understanding the hallucination types produced by a model to select an appropriate augmentation.
  • 11.3.1 Data-Related.: Semantic-factor tags identify fine-grained divergences between source and target tokens, mitigating hallucinations while improving overall NMT performance.
  • 11.3.1 Data-Related.: Minimum risk training and scheduled sampling are proposed to mitigate overexposure bias caused by teacher forcing and reliance on potentially incorrect past predictions.
  • 11.3.2 Modeling and Inference.: Hallucination labels support token loss truncation and hidden-state masking during NMT self-training by discarding or masking losses for predicted hallucinations.
  • 11.3.2 Modeling and Inference.: Tilted empirical risk minimization, dropout, L2E regularization, clipping, and improved phrase alignment are reported as approaches for reducing hallucinations or improving accuracy and source-content identification.
  • 11.4 Future Directions in NMT: Future NMT research should define a metric-based cutoff separating translation errors from hallucinated content, while recognizing that fluent but inadequate sentences may have other causes.A proposed fluency–adequacy threshold is described as close to this goal but not focused specifically on hallucinated content.
  • 11.4 Future Directions in NMT: Detection and mitigation methods also need greater efficiency and practicality because some are computationally expensive, require perturbed test sets, or require one-to-many architectures.

12 HALLUCINATION IN VISION-LANGUAGE GENERATION

Vision-language hallucination research is at an early stage, focusing primarily on object hallucination in image captioning while observing hallucinations in other vision-language tasks. Existing work defines and measures object hallucination, proposes limited mitigation methods, and identifies detection, mitigation, and evaluation as open research directions.

  • Scope: Vision-language models can hallucinate in image captioning, visual question answering, and other vision-grounded text-generation tasks.Hallucination research in this domain remains at an early stage, and measurement and mitigation are open questions.
  • Future Directions: Future research needs stronger hallucination detection, mitigation, and evaluation across tasks including visual storytelling, visual commonsense reasoning, and video captioning.The survey highlights limited empirical and theoretical analyses and the need for more effective evaluation metrics beyond CHAIR.
  • Image Captioning: Object hallucination occurs when image captions contain non-existent or inaccurate objects from the input image.Intrinsic hallucinations are incorrect or definitely non-existent objects, whereas extrinsic hallucinations cannot be verified from the image.
  • Image Captioning: CHAIR measures object hallucination through instance-level and sentence-level rates of hallucinated objects or captions.CHAIR_i measures the fraction of hallucinated object instances, while CHAIR_s measures the fraction of captions containing at least one hallucinated object.
  • Image Captioning: Limited mitigation methods use data augmentation, uncertainty-aware beam search, and object masked language modeling to reduce image-caption hallucination.These approaches target systematic object co-occurrence, predictive uncertainty during decoding, and hallucination in vision-language pre-training.

13 HALLUCINATION IN LARGE LANGUAGE MODELS

In LLMs, hallucination extends beyond source deviations to conflicts with internal training knowledge, making extrinsic hallucination especially important. Research develops reference-dependent and reference-free metrics, mitigation methods such as RLHF, and broader directions addressing multimodal, low-resource, and theoretical challenges.

  • 13.1 Definition: LLM hallucination includes deviations from training data and internal parametric knowledge, not only deviations from source inputs.This broader definition makes hallucination more oriented toward the extrinsic type.
  • 13.1 Definition: Hallucination should be distinguished from factuality because training datasets can contain non-factual and contradictory information.Many studies instead assume training data is factual and use it as a proxy for truth.
  • 13.2 Measurement: LLM hallucination metrics measure generated-content hallucination or individual-model risk through reference-dependent and reference-free approaches.Reference-dependent benchmarks commonly use multiple-choice or fill-in-the-blank QA, whereas reference-free metrics target settings without ground truth.
  • 13.2.1 Hallucination Benchmarks: Human-annotated benchmarks prioritize low-frequency, long-tailed knowledge because common knowledge and basic commonsense are readily memorized.Automated benchmarking addresses the labor and scalability limits of manual question collection and answer annotation.
  • 13.3 Mitigation: 41% to 21%: RLHF fine-tuning of GPT-3 produced InstructGPT and significantly reduced hallucination in closed-domain tasks.Safety fine-tuning and RLHF are prevalent approaches for mitigating hallucination-related risks in LLMs.
  • 13.4 Future Directions: Future work must address hallucination in multimodal, low-resource, and long-tail settings while developing reliable theoretical frameworks and mathematical formulations.Large multimodal models pose greater risks because visual inputs encode object, attribute, spatial, and higher-level information; empirical understanding remains limited.

14 CONCLUSION

The survey presents a comprehensive overview of hallucination in NLG, covering evaluation metrics, mitigation methods, contributors, and future challenges. It emphasizes that fluent hallucinations from neural NLG can mislead users and cause harm in some scenarios.

  • The survey provides the first comprehensive overview of hallucination in NLG, including evaluation metrics, mitigation methods, and remaining future challenges.
  • Hallucination is an artifact of neural-based NLG that can appear fluent and therefore mislead users.
  • In some scenarios and tasks, hallucination can cause harm.
  • The survey examines contributors to hallucination, including noisy data and erroneous parametric knowledge.
Loading 2202.03629v7…