FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
By Sewon Min, Kalpesh Krishna, Xinxi Lyu, Michael Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke S. Zettlemoyer, Hannaneh Hajishirzi. Conference on Empirical Methods in Natural Language Processing, 2023
1,471 citations recorded by the source corpus.
Evaluating the factuality of long-form text generated by large language models (LMs) is non-trivial because (1) generations often contain a mixture of supported and unsupported pieces of information, making binary judgments of quality inadequate, and (2) human evaluation is time-consuming and costly. In this paper, we introduce FACTSCORE, a new evaluation that breaks a generation into a series of atomic facts and computes the percentage of atomic facts supported by a reliable knowledge source. We conduct an extensive human evaluation to obtain FACTSCOREs of people biographies generated by several state-of-the-art commercial LMs -- InstructGPT, ChatGPT, and the retrieval-augmented PerplexityAI -- and report new analysis demonstrating the need for such a fine-grained score (e.g., ChatGPT only achieves 58%). Since human evaluation is costly, we also introduce an automated model that estimates FACTSCORE using retrieval and a strong language model, with less than a 2% error rate. Finally, we use this automated metric to evaluate 6,500 generations from a new set of 13 recent LMs that would have cost $26K if evaluated by humans, with various findings: GPT-4 and ChatGPT are more factual than public models, and Vicuna and Alpaca are some of the best public models. FACTSCORE is available for public use via `pip install factscore`.
Subfields: Artificial Intelligence, General Social Sciences
View Paper: https://www.semanticscholar.org/paper/bd5deadc58ee45b5e004378ba1d54a96bc947b4a
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Mohit Bansal
- Hannaneh Hajishirzi
- Yejin Choi
- Luke S. Zettlemoyer
- Noah A. Smith
- Jianfeng Gao
- Percy Liang
- Graham Neubig
In General Social Sciences
- Dan Roth
- Matei A. Zaharia
- Omar Khattab
- Noah A. Smith
- Christopher Potts
- Leyang Cui
- Qun Liu
- Yizhong Wang
Citations in the corpus
- ARES: An Automated Evaluation Framework for Retrieval-Augmented Generation Systems (2023, background): “For our fine-tuned judges, ARES relies on generating cheap but quality synthetic queries and answers using LLMs.”
- Aligning Large Language Models with Human: A Survey (2023, background): “, 2023d), LM Examiner (Chiang and Lee, 2023) FactScore (Min et al., 2023), Alignscore (Zha et al.”
- Chain-of-Verification Reduces Hallucination in Large Language Models (2023, methodology): “We evaluate the efficacy of our approach using the F ACT S CORE metric (Min et al., 2023) developed in that work, which uses a retrieval-augmented language model to fact-check the response (Instruct-Llama, “Llama + Retrieval + NP”), which they showed correlates well with human judgments.”
- WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild (2024, background): “While our focus is on evaluating LLM capabilities, other evaluation dimensions, such as safety [19], fairness [8], agentic planning [18, 23, 16], and hallucination detection [24, 25, 11], are equally important.”
- SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models (2023, background): “For example, the following work by Min et al. (2023) considers a fine-grained factuality evaluation by decomposing sentences into atomic facts.”
- Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models (2023, positive): “Generation benchmarks consider hallucination as a generation characteristic (Lin et al., 2021; Lee et al., 2022; Min et al., 2023; Yu et al., 2023a) , similar to fluency (Napoles et al., 2017) and coherence (Du et al., 2022), and evaluate the generated texts from LLMs.”
- ExpertQA: Expert-Curated Questions and Attributed Answers (2023, methodology): “We first break down each claim into finer-grained atomic claims using few-shot prompting with text-davinci-003 , using the same prompt and setting as Min et al. (2023).”
- Large Language Models: A Survey (2024, neutral): “FactScore [155] is a recent example of a metric that can be used both for human and model-based evaluation.”
- Prometheus: Inducing Fine-grained Evaluation Capability in Language Models (2023, background): “Fine-grained Evaluation with User-Defined Score Rubrics Score of a specific domain or task 2023; Peng et al., 2023; Zheng et al., 2023; Ye et al., 2023b; Min et al., 2023).”
- From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge (2024, positive): “Min et al. (2023) develop FactScore, a fine-grained method for evaluating the factuality of long-form generation by first splitting content into atomic-level sentences and retrieving relevant corpus from Wikipedia to assess their factuality.”
- A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions (2023, neutral): “However, as the length of the generated content increases, the propensity for hallucination also grows, leading to challenges in evaluating such hallucinations (Min et al., 2023).”
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (2023, methodology): “For FactScore, we set the maximum new token length to 500 for baselines and 200 for S ELF -”
- A Survey on Evaluation of Large Language Models (2023, background): “[129] broke down text generated by LLMs into individual “atomic" facts, which were then evaluated for their correctness.”
- Measuring short-form factuality in large language models (2024, background): “…benchmarks include LongFact (Wei et al., 2024), a benchmark of open-ended prompts, and FreshQA (Vu et al., 2023), a QA benchmark to evaluate performance on fast-changing knowledge, among other work (Lin et al., 2022a; Li et al., 2023; Cheng et al., 2023; Min et al., 2023; Zhao et al., 2024).”
- RAGAs: Automated Evaluation of Retrieval Augmented Generation (2023, background): “Other approaches rely on linking the generated responses to facts from an external knowledge base (Min et al., 2023), but this is not always possible.”
- Fine-Grained Human Feedback Gives Better Rewards for Language Model Training (2023, background): “Different from (a), which collects human preferences on the overall quality of LM outputs, we ask annotators to mark which part of an output contains what type(s) of errors.”
- ASTRA-QA: A Benchmark for Abstract Question Answering over Documents (2026, unknown): “Early evaluation of answers to abstract questions relied on lexical or semantic reference matching, such as ROUGE [17], BERTScore [18], and QA-based or factuality metrics such as FEQA [27], QuestEval [19], QAFactEval [20], and FActScore [21].”