Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Survey of Large Language Models Attribution

Dongfang Li, Zetian Sun, Xinshuo Hu, Zhenyu Liu, Ziyang Chen, Baotian Hu, Aiguo Wu, Min Zhang

TL;DR

This paper surveys attribution mechanisms for open-domain large language models, focusing on grounding model outputs in identifiable sources to mitigate hallucinations. It categorizes attribution sources into pre-training data and out-of-model knowledge, and reviews data resources, methodologies (direct, post-retrieval, post-generation), and practical attribution systems. It discusses evaluation paradigms, including human and automatic metrics, and analyzes common attribution errors, limitations, and biases. The authors outline future directions such as continuous knowledge refresh, improved reliability, and balancing citation with creative generation to enhance trustworthiness.

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 large language models. 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.

A Survey of Large Language Models Attribution

TL;DR

This paper surveys attribution mechanisms for open-domain large language models, focusing on grounding model outputs in identifiable sources to mitigate hallucinations. It categorizes attribution sources into pre-training data and out-of-model knowledge, and reviews data resources, methodologies (direct, post-retrieval, post-generation), and practical attribution systems. It discusses evaluation paradigms, including human and automatic metrics, and analyzes common attribution errors, limitations, and biases. The authors outline future directions such as continuous knowledge refresh, improved reliability, and balancing citation with creative generation to enhance trustworthiness.

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 large language models. 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.
Paper Structure (20 sections, 4 figures, 4 tables)

This paper contains 20 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Task Definition
  3. Sources of Attribution
  4. Pre-training Data
  5. Out-of-model Knowledge
  6. Datasets for Attribution
  7. Approaches to Attribution
  8. Direct Generated Attribution
  9. Post-retrieval Answering
  10. Post-Generation Attribution
  11. Attribution Systems
  12. Attribution Evaluation
  13. Discussion
  14. Attribution Error Analysis
  15. Limitations of Attributions
  16. ...and 5 more sections

Figures (4)

  • Figure 1: By providing attribution, both developers and users can view the possible source of an answer and evaluate factuality and reliability to form their own assessment. Attribution as a more realistic way to reduce hallucinations bypasses the task of directly determining the "truthfulness" of statements, a feat difficult to achieve except for the most basic queries.
  • Figure 2: Taxonomy of Large Language Models Attribution.
  • Figure 3: Three ways to attribute model-generated content. In direct model-driven attribution way, the reference document is derived from model itself and is used to cite generated answer. In post-retrieval answering way, model generates answer with citations based on the retrieved documents. In post-generation attribution way, an answer is first generated then citation and attribution are purposed.
  • Figure 4: Workflow of post-generation attribution. Retrieval is performed after an answer being generated. The retrieved documents are used to perform citation and attribution, subsequently used to do fact verification and post-editing.