Elamur Research⌕
☰
home / papers / 2026 / From Graph to Text and Back: Semantic Fidelity in Automated Industrial Knowledge Graphs

From Graph to Text and Back: Semantic Fidelity in Automated Industrial Knowledge Graphs

Summary

When LLMs turn knowledge-graph triples into prose, they can invent facts or drop relations. That is unacceptable in industrial settings, and checking every output by hand is too expensive. The authors propose a self-supervised round-trip loop to solve this. One LLM writes text from a set of triples, and a second LLM extracts triples back from that text. The original and recovered graphs are then compared with a score that blends embedding cosine similarity and lexical overlap. This score also steers decoding: temperature and top-p go down when fidelity is low and up when it is high, which encourages varied outputs. Only candidates above a strict threshold are kept. On 250 small Wikipedia-derived graphs, the round-trip score agreed closely with human ratings, so it can stand in for expert review. Fine-tuning small LLaMA models on the filtered synthetic text improved triple extraction on automatic metrics and in blinded human evaluation. The approach needs no manual labels and no large teacher model.

Key points

  • Data: 250 Wikipedia-based KGs, each with five triples about one subject. The authors say this size was chosen so experts could check outputs triple by triple, and that 5-triple chunks resemble subgraph retrieval in GraphRAG.
  • Similarity score: it uses multilingual-e5-large-instruct embeddings plus a lexical metric, weighted by α. Thresholds and step sizes were set by grid search on a dev set, and the generation loop runs for N = 100 iterations per sample.
  • Fine-tuned models: LLaMA2-7B, LLaMA3.2-3B and LLaMA3.2-1B, evaluated with BERTScore F1, ROUGE-L F1, BLEU-4 and A–E human ratings. Fine-tuned models received mostly top ratings, while base models clustered at the low end.
  • Training cost and data volume: notable gains already appear with 25% of the synthetic data, which the authors attribute to data quality rather than quantity. Fine-tuning the 3B model takes about 15 minutes, and inference cost is unchanged.
  • Release: code, checkpoints and datasets are published under the MIT license (github.com/KamyarZeinalipour/round-trip-kg). The authors suggest light expert spot-checks to limit bias drift and claim this still cuts annotation cost by over 90%.

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: Machine translation · Natural language processing

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.140

@inproceedings{zeinalipour2026graph,
  title = {{From Graph to Text and Back: Semantic Fidelity in Automated Industrial Knowledge Graphs}},
  author = {Zeinalipour, Kamyar and Severini, Silvia and Borghini, Alessia and Cardarelli, Sara and Maggini, Marco and Gori, Marco},
  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.140},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.140}
}

References

53 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation. In ACL Workshop. find
  2. André Bauer, Simon Trapp, Michael Stenger, Robert Leppich, Samuel Kounev, Mark Leznik, Kyle Chard, and Ian Foster. 2024. Comprehensive exploration of synthetic data generation: A survey. arXiv preprint arXiv:2401.02524. ↗
  3. Thiago et al. Castro Ferreira. 2018. Lessons learned from combining principal methods for natural language generation. In INLG. find
  4. Robert Dale and Ehud Reiter. 1998. Building natural language generation systems. Journal of Natural Language Engineering. find
  5. Bhuwan et al. Dhingra. 2019. Handling divergent reference texts when evaluating table-to-text generation. In ACL. find our card
  6. Ondřej Dušek, Jekaterina Novikova, and Verena Rieser. 2019. Semantic noise affects the quality of data-to-text training data. In INLG. find
  7. Tommaso et al. Furlanello. 2018. Born-again neural networks. In ICML. find
  8. Heng et al. Gong. 2020. Tablegpt: Few-shot table-to-text generation with table structure reconstruction and content matching. In COLING. ↗
  9. Qipeng et al. Guo. 2020. Cyclegt: Unsupervised graph-to-text and text-to-graph generation via cycle training. In WebNLG Workshop. ↗
  10. Hamza Harkous, Isabel Groves, and Amir Saffari. 2020. Have your text and use it too! end-to-end neural data-to-text generation with semantic fidelity. In COL-ING. find
  11. Di et al. He. 2016. Dual learning for machine translation. In NeurIPS. find
  12. Geoffrey et al. Hinton. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. ↗ our card
  13. Ari et al. Holtzman. 2020. The curious case of neural text degeneration. In ICLR. ↗
  14. Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM computing surveys, 55(12):1–38. ↗ our card
  15. Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Structgpt: A general framework for large language model to reason over structured data. arXiv preprint arXiv:2305.09645. ↗
  16. Pei et al. Ke. 2021. Jointgt: Graph-text joint representation learning for text generation from knowledge graphs. In ACL Findings. ↗
  17. Daehee Kim, Deokhyung Kang, Sangwon Ryu, and Gary Geunbae Lee. 2024. Ontology-free general-domain knowledge graph-to-text generation dataset synthesis using large language model. arXiv preprint arXiv:2409.07088. ↗
  18. Rik Koncel-Kedziorski, Dhanush Bekal, Yi Luan, Mirella Lapata, and Hannaneh Hajishirzi. 2019. Text generation from knowledge graphs with graph transformers. In NAACL. ↗ our card
  19. Mateusz Lango and Ondrej Dušek. 2023. Critic-driven decoding for mitigating hallucinations in data-to-text generation. In EMNLP. find
  20. Junyi et al. Li. 2021. Few-shot knowledge graph-to-text generation with pretrained language models. In ACL Findings. ↗
  21. Wei Li, Wenhao Wu, Moye Chen, Jiachen Liu, Xinyan Xiao, and Hua Wu. 2022. Faithfulness in natural language generation: A systematic survey of analysis, evaluation and optimization methods. arXiv preprint arXiv:2203.05227. ↗
  22. Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In ACL Workshop. find
  23. Ruibo Liu, Jerry Wei, Fangyu Liu, Chenglei Si, Yanzhe Zhang, Jinmeng Rao, Steven Zheng, Daiyi Peng, Diyi Yang, Denny Zhou, and Andrew M. Dai. 2024. Best practices and lessons learned on synthetic data. In First Conference on Language Modeling. ↗
  24. Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. 2024. On llms-driven synthetic data generation, curation, and evaluation: A survey. arXiv preprint arXiv:2406.15126. ↗ our card
  25. Lin et al. Long. 2024. On llms-driven synthetic data generation, curation, and evaluation: A survey. In ACL Findings. ↗ our card
  26. Diego Marcheggiani and Laura Perez-Beltrachini. 2018. Deep graph convolutional encoders for structured data-to-text generation. In INLG. find
  27. Hongyuan Mei, Mohit Bansal, and Matthew Walter. 2016. Data-to-text generation with macro planning. In ACL. find
  28. Amit et al. Moryossef. 2019. Step-by-step: Separating planning from realization in neural data-to-text generation. In NAACL. ↗ our card
  29. Chinonso Cynthia Osuji, Thiago Castro Ferreira, and Brian Davis. 2024. A systematic review of data-to-text nlg. arXiv preprint arXiv:2402.08496. ↗
  30. Yutong Shao and Arun Kumar. 2022. Structured data representation in natural language interfaces. A Quarterly bulletin of the Computer Society of the IEEE Technical Committee on Data Engineering, 45(3). find
  31. Shashi Narayan et al. Surya. 2019. Unsupervised sentence simplification using monolingual corpora in multiple languages. In EMNLP. find
  32. Craig Thomson and Ehud Reiter. 2020. Hallucination and omission errors in neural nlg. Computational Linguistics. find
  33. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971. ↗
  34. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288. ↗
  35. Liang et al. Wang. 2022a. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533. ↗ our card
  36. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL. find our card
  37. Shanshan Wang, Chun Zhang, and Ning Zhang. 2024. Mgsa: Multi-granularity graph structure attention for knowledge graph-to-text generation. In 2024 12th International Conference on Information Systems and Computing Technology (ISCTech), pages 1–5. IEEE. find
  38. Ratish Puduppully, Li Dong, and Mirella Lapata. 2019. Data-to-text generation with content selection and planning. In AAAI. ↗
  39. Yizhong et al. Wang. 2022b. Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560. ↗
  40. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. find
  41. Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017. Challenges in data-to-document generation. In EMNLP. find our card
  42. Caroline Rebuffel, Carolina Scarton, and Ehud Reiter. 2022a. Controlling numerical hallucinations in data-to-text generation. In EMNLP. find
  43. Clément Rebuffel, Marco Roberti, Laure Soulier, Geoffrey Scoutheeten, Rossella Cancelliere, and Patrick Gallinari. 2022b. Controlling hallucinations at word level in data-to-text generation. Data Mining and Knowledge Discovery, pages 1–37. ↗
  44. Ehud et al. Reiter. 1997. Generating summaries of patient records. In AMIA. find
  45. Raphael et al. Schumann. 2020. Discrete-event sequence to text with cyclical generative adversarial networks. In COLING. find
  46. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In ACL. find
  47. Kamyar Zeinalipour, Laurent Barriere, Daniele Ghe-lardi, and Marco Gori. 2025. Application of machine learning models for carbon monoxide and nitrogen oxides emission prediction in gas turbines. arXiv preprint arXiv:2501.17865. ↗
  48. Kamyar Zeinalipour, Andrea Fusco, Andrea Zanollo, Marco Maggini, and Marco Gori. 2024a. Harnessing LLMs for educational content-driven Italian crossword generation. In Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC-it 2024). ↗
  49. Kamyar Zeinalipour, Neda Jamshidi, Monica Bianchini, Marco Maggini, and Marco Gori. 2024b. Design proteins using large language models: Enhancements and comparative analyses. In Proceedings of the 1st Workshop on Language + Molecules (L+M 2024), pages 34–47. threatens the delicate balance its ecosystem, conservation efforts have become more pressing than ever, ensuring the long-term survival of this magnificent creature, a poignant reminder that preserving biodiversity is crucial for maintaining the intricate web of life that sustains our planet. find
  50. Kamyar Zeinalipour, Yusuf Gökberk Keptiğ, Marco Maggini, Leonardo Rigutini, and Marco Gori. 2024c. A Turkish educational crossword puzzle generator. arXiv preprint arXiv:2405.07035. ↗
  51. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675. ↗
  52. Yuxiang Zhang, Lei Li, and Tao Jiang. 2024. Chatgraph: Large language model prompting for knowledge graph-to-text generation. arXiv preprint arXiv:2403.01234. ↗
  53. Wayne Xin et al. Zhao. 2023. Knowgpt: Augmenting llms with external knowledge graphs for text generation. arXiv preprint arXiv:2310.00000. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zeinalipour-2026-graph-text-back-semantic-fidelity/