Elamur Research⌕
☰
home / papers / 2026 / Robust Explanations for User Trust in Enterprise NLP Systems

Robust Explanations for User Trust in Enterprise NLP Systems

Paper · 2026 · Guilin Zhang, Kai Zhao, Jeffrey Friedman et al.

Summary

Enterprise teams often reach NLP models only through APIs. That rules out gradient- or attention-based explainers and makes it hard to check, before deployment, whether token-level rationales stay stable when users introduce noise. The question matters more as companies move from encoder classifiers to decoder LLMs. The authors, from Workday AI, propose a black-box protocol. It builds explanations with leave-one-out (LOO) occlusion, which requires only prediction calls. It then measures robustness as the top-1 token Flip Rate under realistic edits: character swaps and deletions, WordNet synonym substitution, word deletion, local shuffling, and back-translation, each at several severity levels. They ran 64,800 paired evaluations covering six models (BERT, RoBERTa, Qwen 7B/14B, Llama 8B/70B) on SST-2, AG News and IMDB. Decoder LLMs gave much more stable rationales than the encoders, and stability kept improving with scale. Encoders almost always kept the same label, yet their rationales shifted often. The authors link these results to inference cost and propose a three-tier selection guide for compliance-sensitive settings such as HR and finance.

Key points

  • Flip Rate counts how often the single most important word, ranked by LOO occlusion, changes after perturbation. Severity levels are 5%, 10% and 20% of positions, and back-translation goes English→German→English.
  • On average, decoders show about 73% lower Flip Rate than encoders. Within decoders, moving from 7B to 70B gives a further ~44% improvement. The decoder advantage is largest for deletion and reordering.
  • The authors call this the "Inconsistency Paradox": BERT and RoBERTa keep their predicted label in nearly all cases (~0.99–1.0 consistency), yet their top explanation token changes in close to half of cases (average FR 0.470).
  • The three deployment tiers are:
    • Speed-first: encoders, FR 0.470, 1× cost.
    • Balanced: 7–8B decoders, FR 0.149, ~64× cost.
    • Regulatory/audit: Llama 70B, FR 0.083, ~636× cost. The authors connect this tier to EU AI Act consistency requirements.
  • The study has three acknowledged limitations. Architecture and scale are confounded, since the encoders have 110M–125M parameters and the decoders 7B–70B. Stability does not imply faithfulness. Only text classification is evaluated.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

Cite

Cite as Guilin Zhang, Kai Zhao, Jeffrey Friedman, Xu Chu and Amine Anoun. 2026. Robust Explanations for User Trust in Enterprise NLP Systems. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.81

@inproceedings{zhang2026robust,
  title = {{Robust Explanations for User Trust in Enterprise NLP Systems}},
  author = {Guilin Zhang, Kai Zhao, Jeffrey Friedman, Xu Chu and Amine Anoun},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.81},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.81}
}

References

27 references, 20 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Samuel Ackerman, Ella Rabinovich, Eitan Farchi, and Ateret Anaby Tavor. 2024. A novel metric for measuring the robustness of large language models in non-adversarial scenarios. In Findings of EMNLP, pages 2794–2802. find
  2. David Alvarez-Melis and Tommi S Jaakkola. 2018. Towards robust interpretability with self-explaining neural networks. In Advances in Neural Information Processing Systems, volume 31. find
  3. Pepa Atanasova, Oana-Maria Camburu, Christina Lioma, Thomas Lukasiewicz, Jakob Grue Simonsen, and Isabelle Augenstein. 2023. Faithfulness tests for natural language explanations. In Proceedings of ACL, pages 11893–11908.
  4. Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, and 1 others. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.
  5. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877–1901.
  6. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186.
  7. Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. 2020. ERASER: A benchmark to evaluate rationalized NLP models. In Proceedings of ACL, pages 4443–4458.
  8. Joakim Edin, Andreas Geert Motzfeldt, Casper L. Christensen, Tuukka Ruotsalo, Lars Maaloe, and Maria Maistro. 2025. Normalized AOPC: Fixing misleading faithfulness metrics for feature attribution explainability. In Proceedings of ACL. find
  9. European Parliament and Council. 2024. Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union. find
  10. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, and 1 others. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  11. Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is BERT really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of AAAI, volume 34, pages 8018–8025.
  12. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361.
  13. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692.
  14. Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, volume 30.
  15. Qing Lyu, Marianna Tan, Hongming Xie, Kaitao Pei, Hongyi Su, Yan Bai, Xinyu Chen, Jing Gao, and Liang Fang. 2024. Towards faithful model explanation in NLP: A survey. Computational Linguistics, 50(2):657–723.
  16. Andrew L Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of ACL, pages 142–150.
  17. George A Miller. 1995. WordNet: A lexical database for English. Communications of the ACM, 38(11):39–41.
  18. John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020. TextAttack: A framework for adversarial attacks, data augmentation, and adversarial training in NLP. In Proceedings of EMNLP: System Demonstrations, pages 119–126.
  19. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. “Why should I trust you?”: Explaining the predictions of any classifier. In Proceedings of KDD, pages 1135–1144.
  20. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of EMNLP, pages 1631–1642.
  21. Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319–3328. find
  22. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
  23. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.
  24. Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2022. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652.
  25. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, and 1 others. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. find
  26. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28.
  27. Yihua Zhang, Changsheng Wang, Yiwei Chen, Chongyu Fan, Jinghan Jia, and Sijia Liu. 2025. The fragile truth of saliency: Improving LLM input attribution via attention bias optimization. In Advances in Neural Information Processing Systems, volume 38. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-robust-explanations-user-trust-enterprise/