Measuring Attribution in Natural Language Generation Models
Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/2023.cl-4.2.pdf
Summary
Neural text generators often produce fluent statements that their underlying sources do not support. Prior work on hallucination and faithfulness lacked a shared, formally grounded way to measure this. The authors propose Attributable to Identified Sources (AIS), a framework that judges whether generated text about the world is backed by a specific supplied source. A system utterance is first resolved into its full context-independent meaning (an explicature), and then evaluated with a simple test: would "According to P, s" hold for source P? Outputs whose meaning cannot be recovered are handled separately.
The framework is operationalized as a two-stage human annotation process. Raters first decide whether the output is interpretable without seeing the source. Only interpretable outputs move on to an attribution judgment against the source. The authors adapt AIS to conversational QA, summarization, and table-to-text, and apply it to outputs from several models. Annotators reached moderate to high agreement, and model scores differed in expected ways. AIS is meant to complement, not replace, metrics for fluency or source quality.
Key points
- AIS defines attribution through explicatures, building on Carston and Wilson & Sperber and on decontextualization (Choi et al. 2021). For example, "he was 25" becomes a standalone claim naming George Harrison and the album.
- The annotation pipeline has two stages:
- Interpretability: the source is hidden from raters.
- Attribution: the source is shown; this stage is skipped if the output was judged uninterpretable.
- Validation covered four datasets across three tasks: two conversational QA sets, one summarization set, and one table-to-text set. The table-to-text set is ToTTo, with highlighted cells preserved in the interface.
- AIS deliberately avoids absolute factuality judgments. The source is treated as the reference, and trustworthiness of the source is left to other measures.
- Full guidelines and interfaces appear in the appendices, including a flag for malformed or expert-only items. The annotated data is released at github.com/google-research-datasets/AIS.
Cited by
- Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models
- Chain-of-Verification Reduces Hallucination in Large Language Models
- Enabling Large Language Models to Generate Text with Citations
- RARR: Researching and Revising What Language Models Say, Using Language Models
- Gemini: A Family of Highly Capable Multimodal Models
- TRUE: Re-evaluating Factual Consistency Evaluation
- Evaluating Open-Domain Question Answering in the Era of Large Language Models
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- Evaluating Verifiability in Generative Search Engines
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- LaMDA: Language Models for Dialog Applications
- Survey on Factuality in Large Language Models: Knowledge, Retrieval and Domain-Specificity
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Cite
@article{rashkin2023measuring,
title = {{Measuring Attribution in Natural Language Generation Models}},
author = {Rashkin, Hannah and Nikolaev, Vitaly and Lamm, Matthew and Aroyo, Lora and Collins, Michael and Das, Dipanjan and Petrov, Slav and Tomar, Gaurav Singh and Turc, Iulia and Reitter, David},
journal = {Computational Linguistics},
year = {2023},
doi = {10.1162/coli\_a\_00486},
url = {https://doi.org/10.1162/coli_a_00486}
}References
- Adiwardana, Daniel, Minh-Thang Luong, David R. So, Jamie Hall, Noah Fiedel, Romal Thoppilan, Zi Yang, Apoorv Kulshreshtha, Gaurav Nemade, Yifeng Lu, and Quoc V. Le. 2020. Towards a human-like open-domain chatbot. arXiv preprint arXiv:2001.09977. find our card
- Anantha, Raviteja, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguis… find
- Carston, Robyn. 1988. Implicature, explicature, and truth-theoretic semantics. In Ruth Kempson, editor, Mental Representations: The Interface Between Language and Reality. Cambridge University Press, pages 155–181. find
- Choi, Eunsol, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. 2018. QuAC: Question answering in context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2174–2184. https://doi.org/10.18653/v1 /D18-1241 Cho… find our card
- Dalton, Jeffrey, Chenyan Xiong, Vaibhav Kumar, and Jamie Callan. 2020. Cast-19: A dataset for conversational information seeking. In Proceedings of SIGIR, pages 1985–1988. https://doi.org /10.1145/3397271.3401206 Dinan, Emily, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston.… ↗
- Durmus, Esin, He He, and Mona Diab. 2020. FEQA: A question answering evaluation framework for faithfulness assessment in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5055–5070. https://doi.org/10 .18653/v1/2020.acl-main.… find
- Elsevier Science. Honovich, Or, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and Omri Abend. 2021. q 2 : Evaluating factual consistency in knowledge-grounded dialogues via question generation and question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural L… find
- Kwiatkowski, Tom, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural… find our card
- Raffel, Colin, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(1):5485–5551. find
- Rashkin, Hannah, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Nat… find
- Thorne, James and Andreas Vlachos. 2018. Automated fact checking: Task formulations, methods and future directions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3346–3359. find
- Thorne, James, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. 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, V… find our card
- Wiseman, Sam, Stuart Shieber, and Alexander Rush. 2017. Challenges in data-to-document generation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2253–2263. https://doi.org/10.18653/v1/D17 -1239 Zaheer, Manzil, Guru Guruganesh, Kumar Avinava Dubey, J… ↗ our card
- In Advances in Neural Information Processing Systems, volume 33, pages 17283–17297. Zhong, Ming, Pengfei Liu, Yiran Chen, Danqing Wang, Xipeng Qiu, and Xuanjing Huang. 2020. Extractive summarization as text matching. In Proceedings of the 58th Annual Meeting of the Association for Computational Ling… find our card