Home

Awesome

A Survey of Large Language Models Attribution [ArXiv preprint]

🌟 Introduction

Open-domain dialogue systems, driven by large language models, have changed the way we use conversational AI. However, these systems often produce content that might not be reliable. In traditional open-domain settings, the focus is mostly on the answer’s relevance or accuracy rather than evaluating whether the answer is attributed to the retrieved documents. A QA model with high accuracy may not necessarily achieve high attribution.

Attribution refers to the capacity of a model, such as an LLM, to generate and provide evidence, often in the form of references or citations, that substantiates the claims or statements it produces. This evidence is derived from identifiable sources, ensuring that the claims can be logically inferred from a foundational corpus, making them comprehensible and verifiable by a general audience. The primary purposes of attribution include enabling users to validate the claims made by the model, promoting the generation of text that closely aligns with the cited sources to enhance accuracy and reduce misinformation or hallucination, and establishing a structured framework for evaluating the completeness and relevance of the supporting evidence in relation to the presented claims.

In this repository, we focus on unraveling the sources that these systems tap into for attribution or citation. We delve into the origins of these facts, their utilization by the models, the efficacy of these attribution methodologies, and grapple with challenges tied to ambiguous knowledge reservoirs, inherent biases, and the pitfalls of excessive attribution.

✨ Work in progress. We would like to appreciate any contributions via PRs, issues from NLP community.

1. Attribution Definition & Position Paper

<!--* [2023/03] **TRAK: Attributing Model Behavior at Scale** *Sung Min Park et al. arXiv.* [[paper](https://arxiv.org/abs/2303.14186)][[code](https://github.com/MadryLab/trak)] ``` Attributing Model: trace model predictions back to training data. This paper introduces a data attribution method that is both effective and computationally tractable for large-scale, differentiable models. ```-->

2. Attribution Paper Before the Era of Large Language Models and Related Task

2.1 Fact Checking & Claim Verificication & Natural Language Inference

2.2 Feature Attribution and Interpretability of Models for NLP

2.3 Attribution in Mutli-modal Systems

2.4 Wiki

2.5 Model based Information Retrieval

2.6 Small Language Model

3. Sources of Attribution

3.1 Pre-training Data

3.2 Out-of-model Knowledge and Retrieval-based Question Answering & Knowledge-Grounded Dialogue

4. Datasets for Attribution

5. Approaches to Attribution

5.1 Direct Generated Attribution

5.2 Retrieval-then-Answering

<!--* [2023/05] **WikiChat: Stopping the Hallucination of Large Language Model Chatbots by Few-Shot Grounding on Wikipedia** *Sina Semnani et al. Findings of EMNLP'23** [[paper]](https://arxiv.org/abs/2305.14292) [[code]](https://github.com/stanford-oval/WikiChat) [[demo]](https://wikichat.genie.stanford.edu)-->

5.3 Post-Generation Attribution

5.4 Attribution Systems & End-to-End Attribution Models

6. Attribution Evaluation

<!--``` This paper defines a new task of Knowledge-aware Language Model Attribution (KaLMA), and builds a dataset in biography domain BioKaLMA via a well-designed evolutionary question generation strategy. ```-->

7. Limitations, Future Directions and Challenges in Attribution

  a. hallucination of attribution i.e. does attribution faithfully to its content?
  b. Inability to attribute parameter knowledge of model self.
  c. Validity of the knowledge source - source trustworthiness. Faithfulness ≠ Factuality
  d. Bias in attribution method
  e. Over-attribution & under-attribution
  f. Knowledge conflict

Cite

@misc{li2023llmattribution,
      title={A Survey of Large Language Models Attribution}, 
      author={Dongfang Li and Zetian Sun and Xinshuo Hu and Zhenyu Liu and Ziyang Chen and Baotian Hu and Aiguo Wu and Min Zhang},
      year={2023},
      eprint={2311.03731},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      howpublished={\url{https://github.com/HITsz-TMG/awesome-llm-attributions}},
}

For finding survey of hallucination please refer to:

Project Maintainers & Contributors