Survey of Hallucination in Natural Language Generation¶
Ji et al. (2023) provide a comprehensive survey of hallucination across natural language generation tasks, published in ACM Computing Surveys. The paper covers hallucination in abstractive summarisation, dialogue generation, generative question answering, data-to-text generation, machine translation, vision-language generation, and — in a section updated in 2024 — large language models. It establishes a foundational taxonomy and reviews contributors, metrics, and mitigation methods.
Hallucination Taxonomy¶
The paper distinguishes two fundamental types:
Intrinsic hallucination: the generated output contradicts the source content. For example, a summary that states a vaccine was approved in 2021 when the source says 2019. These are always erroneous and detectable by source comparison.
Extrinsic hallucination: the generated output cannot be verified from the source — it is neither supported nor contradicted. For example, adding information about clinical trials not mentioned in the source document. Extrinsic hallucinations are not always incorrect — they may draw on valid world knowledge — but they are unverifiable from the given context, making them risky in contexts where faithfulness to source material matters.
This distinction matters for safety-critical systems because different tasks have different hallucination tolerances. Summarisation and data-to-text tasks require strict faithfulness (low tolerance for either type), while dialogue systems tolerate more extrinsic hallucination because user engagement benefits from drawing on broader knowledge.
Contributors to Hallucination¶
Data-related contributors: source-reference divergence in training data (the target output in training contains information not in the source); heuristic data collection that pairs mismatched sources and targets; noisy or inconsistent training data.
Training and inference contributors: the standard likelihood maximisation objective can encourage the model to generate likely text rather than faithful text; exposure bias (the model trains on gold targets but at inference conditions on its own previous output); decoder strategies like beam search can amplify hallucination patterns.
Metrics and Evaluation¶
The paper categorises hallucination metrics into: - Statistical metrics (e.g., ROUGE, BLEU) — measure surface overlap but miss semantic hallucination - Model-based metrics (e.g., BERTScore, entailment classifiers, QA-based evaluation) — better at detecting semantic divergence but require task-specific calibration - Human evaluation — remains the gold standard but is expensive and inconsistent across annotators
Mitigation Methods¶
- Data-related methods: improving training data quality, filtering source-reference divergent pairs, augmenting with faithful examples
- Architecture and training methods: faithful attention mechanisms, multi-task learning with entailment objectives, reinforcement learning with faithfulness rewards
- Inference methods: constrained decoding, knowledge-grounded generation (see retrieval-augmented-generation), post-hoc verification and editing
Relevance to This Wiki¶
This survey provides the theoretical foundation for the hallucination page's treatment of hallucination types and causes. The intrinsic/extrinsic taxonomy is particularly valuable for safety-critical system design: intrinsic hallucinations (contradicting provided information) are categorically unacceptable in advisory systems, while extrinsic hallucinations (adding unverifiable information) require more nuanced assessment depending on the application context. The survey complements Huang et al. (2023), which focuses specifically on LLM hallucination, by providing the broader NLG context and establishing that hallucination is a fundamental property of neural text generation, not a problem unique to large language models.