Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
126 tokens/sec
GPT-4o
47 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

A Survey of Large Language Models Attribution (2311.03731v2)

Published 7 Nov 2023 in cs.CL

Abstract: Open-domain generative systems have gained significant attention in the field of conversational AI (e.g., generative search engines). This paper presents a comprehensive review of the attribution mechanisms employed by these systems, particularly LLMs. Though attribution or citation improve the factuality and verifiability, issues like ambiguous knowledge reservoirs, inherent biases, and the drawbacks of excessive attribution can hinder the effectiveness of these systems. The aim of this survey is to provide valuable insights for researchers, aiding in the refinement of attribution methodologies to enhance the reliability and veracity of responses generated by open-domain generative systems. We believe that this field is still in its early stages; hence, we maintain a repository to keep track of ongoing studies at https://github.com/HITsz-TMG/awesome-LLM-attributions.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (84)
  1. Do language models know when they’re hallucinating references? CoRR, abs/2305.18248.
  2. Palm 2 technical report. ArXiv, abs/2305.10403.
  3. Anonymous. 2023. Learning to plan and generate text with citations. In Submitted to The Twelfth International Conference on Learning Representations. Under review.
  4. Self-rag: Learning to retrieve, generate, and critique through self-reflection. CoRR, abs/2310.11511.
  5. Amos Azaria and Tom M. Mitchell. 2023. The internal state of an LLM knows when its lying. CoRR, abs/2304.13734.
  6. Training a helpful and harmless assistant with reinforcement learning from human feedback. ArXiv, abs/2204.05862.
  7. Attributed question answering: Evaluation and modeling for attributed large language models. CoRR, abs/2212.08037.
  8. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 2206–2240. PMLR.
  9. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pages 632–642. The Association for Computational Linguistics.
  10. Sergey Brin and Lawrence Page. 1998. The anatomy of a large-scale hypertextual web search engine. Comput. Networks, 30:107–117.
  11. Reading wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 1870–1879. Association for Computational Linguistics.
  12. Understanding retrieval augmentation for long-form question answering.
  13. Complex claim verification with evidence retrieved in the wild. CoRR, abs/2305.11859.
  14. Factool: Factuality detection in generative AI - A tool augmented framework for multi-task and multi-domain scenarios. CoRR, abs/2307.13528.
  15. Quantifying and attributing the hallucination of large language models via association analysis. CoRR, abs/2309.05217.
  16. Finding news citations for wikipedia. In Proceedings of the 25th ACM International Conference on Information and Knowledge Management, CIKM 2016, Indianapolis, IN, USA, October 24-28, 2016, pages 337–346. ACM.
  17. Citebench: A benchmark for scientific citation text generation.
  18. RARR: researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 16477–16508. Association for Computational Linguistics.
  19. Enabling large language models to generate text with citations. CoRR, abs/2305.14627.
  20. Improving alignment of dialogue agents via targeted human judgements. CoRR, abs/2209.14375.
  21. Learning to fake it: Limited responses and fabricated references provided by chatgpt for medical questions. Mayo Clinic Proceedings: Digital Health, 1(3):226–234.
  22. Nianlong Gu and Richard H. R. Hahnloser. 2022. Controllable citation text generation. CoRR, abs/2211.07066.
  23. Textbooks are all you need. CoRR, abs/2306.11644.
  24. A survey on automated fact-checking. Trans. Assoc. Comput. Linguistics, 10:178–206.
  25. Understanding in-context learning via supportive pretraining data. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 12660–12673. Association for Computational Linguistics.
  26. Xiaochuang Han and Yulia Tsvetkov. 2022. ORCA: interpreting prompted language models via locating supporting data evidence in the ocean of pretraining data. CoRR, abs/2205.12600.
  27. Rethinking with retrieval: Faithful large language model inference. CoRR, abs/2301.00303.
  28. Jie Huang and Kevin Chen-Chuan Chang. 2023. Citation: A key to building responsible and accountable large language models. CoRR, abs/2307.02185.
  29. Retrieving supporting evidence for generative question answering. CoRR, abs/2309.11392.
  30. Gautier Izacard and Edouard Grave. 2021. Leveraging passage retrieval with generative models for open domain question answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021, pages 874–880. Association for Computational Linguistics.
  31. Alon Jacovi and Yoav Goldberg. 2020. Towards faithfully interpretable NLP systems: How should we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4198–4205, Online. Association for Computational Linguistics.
  32. 1-pager: One pass answer generation and evidence retrieval. CoRR, abs/2310.16568.
  33. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12):248:1–248:38.
  34. HAGRID: A human-llm collaborative dataset for generative information-seeking with attribution. arXiv:2307.16883.
  35. Wice: Real-world entailment for claims in wikipedia. CoRR, abs/2303.01432.
  36. LAMBADA: Backward chaining for automated reasoning in natural language. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6547–6568, Toronto, Canada. Association for Computational Linguistics.
  37. Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over BERT. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, pages 39–48. ACM.
  38. Rethinking explainability as a dialogue: A practitioner’s perspective. CoRR, abs/2202.01875.
  39. Towards reliable and fluent large language models: Incorporating feedback learning loops in qa systems. arXiv preprint arXiv:2309.06384.
  40. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6086–6096. Association for Computational Linguistics.
  41. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
  42. Unifying model explainability and robustness for joint text classification and rationale extraction. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 10947–10955. AAAI Press.
  43. Llatrieval: Llm-verified retrieval for verifiable generation. arXiv preprint arXiv:2311.07838.
  44. Towards verifiable generation: A benchmark for knowledge-aware language model attribution.
  45. Establishing trustworthiness: Rethinking tasks and model evaluation.
  46. Frederick Liu and Besim Avci. 2019. Incorporating priors with feature attribution on text classification. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6274–6283. Association for Computational Linguistics.
  47. Evaluating verifiability in generative search engines. ArXiv, abs/2304.09848.
  48. Expertqa: Expert-curated questions and attributed answers. ArXiv, abs/2309.07852.
  49. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. CoRR, abs/2303.08896.
  50. Teaching language models to support answers with verified quotes. ArXiv, abs/2203.11147.
  51. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. CoRR, abs/2305.14251.
  52. Evaluating and modeling attribution for cross-lingual question answering. CoRR, abs/2305.14332.
  53. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332.
  54. OpenAI. 2022. Chatgpt: Optimizing language models for dialogue.
  55. OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774.
  56. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155.
  57. The pagerank citation ranking : Bringing order to the web. In The Web Conference.
  58. The refinedweb dataset for falcon llm: Outperforming curated corpora with web data, and web data only. ArXiv, abs/2306.01116.
  59. Denis Peskoff and Brandon Stewart. 2023. Credible without credit: Domain experts assess generative language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 427–438. Association for Computational Linguistics.
  60. Improving wikipedia verifiability with AI. CoRR, abs/2207.06220.
  61. The ROOTS search tool: Data transparency for llms. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics: System Demonstrations, ACL 2023, Toronto, Canada, July 10-12, 2023, pages 304–314. Association for Computational Linguistics.
  62. Webbrain: Learning to generate factually correct articles for queries by grounding on large web corpus. CoRR, abs/2304.04358.
  63. WebCPM: Interactive web search for Chinese long-form question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8968–8988, Toronto, Canada. Association for Computational Linguistics.
  64. Measuring attribution in natural language generation models. CoRR, abs/2112.12870.
  65. A survey of hallucination in large foundation models. ArXiv, abs/2309.05922.
  66. Smartbook: Ai-assisted situation report generation. CoRR, abs/2303.14337.
  67. SEMQA: semi-extractive multi-source question answering. CoRR, abs/2311.04886.
  68. Retrieval augmentation reduces hallucination in conversation. In Findings of the Association for Computational Linguistics: EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pages 3784–3803. Association for Computational Linguistics.
  69. Recitation-augmented language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.
  70. Transformer memory as a differentiable search index. In NeurIPS.
  71. Lamda: Language models for dialog applications. ArXiv, abs/2201.08239.
  72. FEVER: a large-scale dataset for fact extraction and verification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long Papers), pages 809–819. Association for Computational Linguistics.
  73. Haoran Wang and Kai Shu. 2023. Explainable claim verification via knowledge-grounded reasoning with large language models.
  74. " according to…" prompting language models improves quoting from pre-training data. arXiv preprint arXiv:2305.13252.
  75. Towards generating citation sentences for multiple references with intent control. CoRR, abs/2112.01332.
  76. Adaptive chameleon or stubborn sloth: Unraveling the behavior of large language models in knowledge clashes. ArXiv, abs/2305.13300.
  77. Automatic generation of citation texts in scholarly papers: A pilot study. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 6181–6190. Association for Computational Linguistics.
  78. Search-in-the-chain: Towards the accurate, credible and traceable content generation for complex knowledge-intensive tasks. CoRR, abs/2304.14732.
  79. Cognitive mirage: A review of hallucinations in large language models. ArXiv, abs/2309.06794.
  80. Effective large language model adaptation for improved grounding. arXiv preprint arXiv:2311.09533.
  81. Automatic evaluation of attribution by large language models. CoRR, abs/2305.06311.
  82. Mitigating language model hallucination with interactive question-knowledge alignment. CoRR, abs/2305.13669.
  83. Siren’s song in the ai ocean: A survey on hallucination in large language models. ArXiv, abs/2309.01219.
  84. Chatgpt hallucinates when attributing answers.
Citations (39)
View on Semantic Scholar

Summary

  • The paper presents a systematic taxonomy of attribution methods in LLMs to reduce hallucinations.
  • It analyzes various datasets and techniques, including direct generation and post-retrieval approaches, to provide verifiable citations.
  • The evaluation employs human and automatic metrics, establishing accuracy metrics like precision and F1-score for source credibility.

A Survey of LLMs Attribution: A Comprehensive Overview

The paper "A Survey of LLMs Attribution" proposes an extensive exploration into the mechanisms and methodologies that underpin attribution within LLMs. Attribution, within the context of LLMs, pertains to the provision of source citations for generated content, thus enhancing factual accuracy and credibility. This paper delineates the current landscape of attribution methods, highlighting the challenges and advances within this domain.

Introduction and Motivation

The paper outlines the growing prominence of open-domain generative systems, powered by LLMs. Despite the improved capabilities of these models, a looming challenge remains: the "hallucination" problem, where outputs include distorted facts without credible sourcing. Attribution is suggested as a realistic approach to mitigating hallucinations. By employing citations, LLMs can provide verifiable sources, thus allowing both developers and users to assess the factuality of the generated content.

Core Classification and Methodological Approaches

The paper employs a systematic taxonomy to categorize sources, datasets, and methods for attribution in LLMs.

  1. Sources of Attribution:
    • Pre-training Data: Attribution is inherently linked to the vast corpora from which LLMs learn. Identifying specific data subsets that influence model decisions can enhance the interpretability of these systems.
    • Out-of-model Knowledge: The inclusion of external databases such as the web or knowledge graphs enhances model reliability by anchoring generated content to verifiable sources.
  2. Datasets for Attribution:
    • Datasets play a key role in evaluating attribution. This paper references various datasets that have been adapted for examining the accuracy of citation mechanisms within LLMs. These datasets vary in domain focus and citation granularity, ranging from general open-domain datasets to domain-specific ones.
  3. Approach to Attribution:
    • Direct Generated Attribution: Utilizing parametric knowledge from the model itself to provide attributions, although this method faces challenges in accuracy and relevance.
    • Post-retrieval Answering: This involves leveraging retrieved information to form the basis of model-generated answers. While effective, it necessitates clear boundaries between retrieved and inherent model knowledge.
    • Post-Generation Attribution: Conducting retrieval post-answer generation for sources, thereby ensuring that all claims are verifiable and grounded in reality.
  4. Attribution Systems:
    • Systems such as GopherCite and WebGPT implement real-time retrieval to improve accuracy and remain up-to-date with reliable references. These systems showcase the integration of external web sources and underscore the benefits of reinforcement learning for refining citation processes.

Evaluation and Challenges

The evaluation of attribution is primarily approached through both human and automatic metrics. Studies focus on recall, precision, and F1-score to determine the accuracy of cited references concerning generated text. Challenges in attribution revolve around granularity, mistaken synthesis, hallucination, excessive citation, bias, and evolving content. A fine balance is required to ensure accurate attribution while maintaining model dexterity.

Implications and Future Directions

The implications of this research are multifaceted:

  • Practical Impact: Improved citation mechanisms bolster the trust in LLM outputs, fostering reliability in applications across knowledge-intensive fields.
  • Theoretical Impact: Theoretical advancements in attribution can guide the refinement of LLM architectures and training processes.

Looking forward, the authors advocate for continuous updates to LLMs, akin to the real-time data refreshes seen in search engines. This progression ensures outputs remain timely and relevant. Additionally, a call is made for establishing robust frameworks to consistently evaluate source credibility.

Conclusion

This detailed examination of attribution within LLMs provides critical insights into improving model reliability. By systematically addressing attribution challenges and implementing structured approaches, the surveyed frameworks aim not only to enhance interpretability but also to advance the trustworthiness and practical utility of LLMs in real-world applications. As this field continues to evolve, robust attribution mechanisms will likely serve as a cornerstone for the integrity and advancement of AI systems.

File Document Download Save Streamline Icon: https://streamlinehq.com PDF File Document Download Save Streamline Icon: https://streamlinehq.com Markdown
Github Logo Streamline Icon: https://streamlinehq.com

GitHub

  1. GitHub - HITsz-TMG/awesome-llm-attributions: A Survey of Attributions for Large Language Models (203 stars)
Youtube Logo Streamline Icon: https://streamlinehq.com

YouTube