Back

Generate-on-Graph: Treat LLM as both Agent and KG for Incomplete Knowledge Graph Question Answering

Yao XuShizhu HeJiabei ChenZihao WangYangqiu SongHanghang TongGuang LiuJun ZhaoKang Liu

article2024EMNLP112 citations

Proposes Generate-on-Graph, a training-free framework that enables large language models to answer complex questions over incomplete knowledge graphs by alternating between exploring existing graph triples and generating missing factual connections using internal model knowledge.

Listen

Large Language Models often struggle with factual inaccuracies, hallucinations, and gaps in specialized knowledge. While connecting language models to external knowledge graphs can ground responses in structured facts, traditional integration benchmarks assume these knowledge bases are completely comprehensive. In real-world enterprise environments, external knowledge graphs are frequently incomplete, causing standard methods—such as semantic parsing and path retrieval—to fail or produce ungrounded guesses when critical factual connections are missing.

The main objective of the article is to establish a realistic benchmark for question answering over incomplete knowledge graphs and to propose and evaluate Generate-on-Graph, a training-free framework that enables language models to act both as exploring agents and as dynamic knowledge generators.

To simulate real-world conditions, the authors created incomplete benchmark datasets from standard question-answering collections by randomly deleting key factual links at missingness rates between 20% and 80%. They evaluated the Generate-on-Graph framework against standard prompting, semantic parsing, and retrieval-augmented baselines using multiple language model backbones, including GPT-3.5, GPT-4, and open-source models. The framework operates through an iterative cycle where the model plans the next reasoning step, searches the graph using structured tools, and actively generates missing factual triples using internal knowledge whenever the graph lacks necessary links.

The evaluation yielded several key findings. First, Generate-on-Graph consistently outperformed all baseline methods across both complete and incomplete graph settings, achieving top accuracy scores of 84.4% on WebQSP and 75.2% on Complex WebQuestions when powered by GPT-4. Second, conventional methods experienced severe performance drops under incomplete graphs; for example, semantic parsing accuracy fell from 76.5% to 39.3% when 40% of key facts were missing. Third, Generate-on-Graph maintained robust performance across varying degrees of missing information, delivering an average accuracy improvement of 5.0% over competing methods on complex questions. Finally, ablation studies showed that dynamically generating missing facts with contextual graph cues significantly enhanced accuracy, though retrieving excessive, irrelevant graph context introduced noise that degraded performance.

These findings indicate that treating language models as both navigators and supplemental knowledge sources bridges critical data gaps in automated reasoning systems. For organizations deploying knowledge-assisted artificial intelligence, this approach reduces the cost and burden of maintaining exhaustive databases while mitigating the risk of query failures caused by incomplete records. However, because generating missing facts relies on model memory, systems remain susceptible to hallucinations and phrasing mismatches.

Organizations seeking to implement knowledge graph question answering should adopt hybrid agent-generator architectures rather than rigid semantic parsers. Before full deployment, teams should run controlled pilots to tune context retrieval thresholds and implement automated validation routines to check model-generated facts against known organizational constraints.

bipartite graphsdirected graphsgeneral graphIncomplete knowledge graph question answeringLarge language modelsToG reasoning paths
Xu et al (2024).pdf

No sufficiently relevant recommendations were found.

Cover for Generate-on-Graph: Treat LLM as both Agent and KG for Incomplete Knowledge Graph Question Answering

Topics

bipartite graphsdirected graphsgeneral graphintelligent agentsquestion answeringsoftware agentsweighted graphknowledge graphs

Keywords

Incomplete knowledge graph question answeringLarge language modelsToG reasoning pathsKnowledge graphPath retrieval

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Preliminary
  • 4 Incomplete Knowledge Graph Question Answering (IKGQA)
  • 4.1 Task Introduction
  • 4.2 Datasets Construction
  • 5 Generate-on-Graph (GoG)
  • 5.1 Thinking
  • 5.2 Searching
  • 5.3 Generating
  • 6 Experiments
  • 6.1 Experiments Setup
  • 6.2 Main Results
  • 6.3 Performance under Different Degrees of KGIncompleteness
  • 6.4 Performance with Different LLMs
  • 6.5 Ablation Study
  • The Effect of the Number of Related Triples
  • The Effect of Generate Action
  • 7 Conclusion
  • Limitation
  • Ethics Statement
  • Acknowledgment
  • References
  • A Semantic Parsing Methods Details
  • B Retrieval Augmented Methods Details
  • C Prompt List
  • D Settings for Baselines
  • E Statistics of Topic Entities in IKGs
  • F Compound Value Type (CVT) node
  • G Result Analysis
  • G.1 Performance under Generate Action
  • G.2 Error Analysis
  • H Case Study
  • H.1 Comparison between ToG and GoG under CKG setting
  • H.2 Comparison between ToG and GoG under IKG setting
  • H.3 Cases of Generate Action
  • H.4 Bad Cases of GoG

Knowls

  1. Knowl 1 — Incomplete-KG question answering requires combining graph evidence with missing facts

    definition

    Incomplete Knowledge Graph Question Answering (IKGQA) asks a system to answer a question from a knowledge graph that may omit one or more facts needed to connect a mentioned topic entity to an answer entity. In ordinary KGQA, the triples along the question’s gold relation path are assumed to be present; in IKGQA, at least one such triple may be absent. The system must therefore use available graph structure together with its internal knowledge or reasoning to recover the missing information.

  2. Knowl 2 — Construction of the WebQSP and CWQ IKGQA datasets

    experimental setup

    The paper constructs IKGQA versions of WebQuestionsSP (WebQSP) and Complex WebQuestions (CWQ), both using Freebase, by sampling 1,000 questions from each source dataset. For each question, its SPARQL query is executed against the original graph to identify triples on the answer path; property-node triples are filtered out, and each remaining crucial triple is independently selected for deletion with probability pp. The resulting settings remove 20%, 40%, 60%, or 80% of crucial triples per question. In addition to each selected triple, the construction also removes other relations between that triple’s two endpoint entities. Topic entities with no remaining neighbors are excluded. This yields question-specific incomplete graphs in which the original answer path is no longer guaranteed to be fully available.

  3. Knowl 3 — Generate-on-Graph alternates reasoning, graph search, and fact generation

    model/method

    Generate-on-Graph (GoG) is a training-free method for IKGQA that uses an LLM both as an agent exploring the supplied knowledge graph and as a source of candidate facts. It repeatedly produces a thought that decomposes the question or selects a next subproblem, then takes one of three actions: search the graph for relevant evidence, generate candidate factual triples using retrieved evidence and the LLM’s internal knowledge, or finish with an answer (or unknown). Search results and accepted generated triples become observations available to subsequent steps. The process continues until the agent judges it has sufficient information to answer.

  4. Knowl 4 — GoG searches one-hop graph neighborhoods and filters relations by the current subproblem

    model/method

    For a search action targeting an entity, GoG first uses predefined SPARQL queries to retrieve the entity’s one-hop relations and neighboring entities. An LLM then ranks the retrieved relations for relevance to the agent’s latest thought, and GoG retains the selected relations and their associated triples as the observation. The observation is added to the context for later reasoning. The method description considers one target entity per search action for simplicity, while its prompt permits multiple entities. If the agent returns unknown, GoG rolls back and searches one additional hop from the last target entity.

  5. Knowl 5 — GoG generates, checks, and links candidate triples

    model/method

    When graph observations do not directly answer the current subproblem, GoG invokes a generation action. It uses BM25 to select relevant triples from prior observations, then prompts the LLM to generate new triples from those facts and its internal knowledge; generation is repeated nn times to reduce errors and hallucinations. An LLM verifier selects generated triples considered more likely to be accurate, and accepted triples become observations. If generation introduces an entity not previously explored, GoG retrieves similar graph entities and their types using BM25, then asks an LLM to select the most relevant candidate so the entity can be linked to a Freebase machine identifier.

  6. Knowl 6 — GoG outperforms the compared methods on complete and 40%-incomplete graphs

    data/table

    The following Hits@1 scores (percent) compare methods on complete knowledge graphs (CKG) and graphs with 40% of crucial triples removed (IKG-40%). LLM-only prompting has no graph input; the other rows use a knowledge graph. GoG with GPT-3.5 exceeds the other listed prompt-based methods on both datasets in both graph settings. GoG with GPT-4 also exceeds ToG with GPT-4 in all four settings. Fine-tuned RoG is stronger than GPT-3.5 GoG on CKG, but its scores fall on IKG-40%; the comparison shows the reported performance under the stated graph settings, not a guarantee of generalization beyond these datasets.

    Method CWQ CKG CWQ IKG-40% WebQSP CKG WebQSP IKG-40%
    IO prompt 37.6 63.3
    CoT 38.8 62.2
    CoT+SC 45.4 61.1
    RoG 66.1 54.2 88.6 78.2
    ChatKBQA 76.5 39.3 78.1 49.5
    KB-BINDER – – 50.7 38.4
    StructGPT – – 76.4 60.1
    ToG (GPT-3.5) 47.2 37.9 76.9 63.4
    GoG (GPT-3.5) 55.7 44.3 78.7 66.6
    ToG (GPT-4) 71.0 56.1 80.3 71.8
    GoG (GPT-4) 75.2 60.4 84.4 80.3

  7. Knowl 7 — GoG retains an advantage as crucial triples are increasingly removed

    data/table

    These Hits@1 scores (percent) compare prompt-based methods using GPT-3.5 across complete graphs and graphs with 20%, 40%, 60%, or 80% of crucial triples removed. GoG has the highest score among the listed methods at every incompleteness level on both datasets. Its advantage over ToG is especially consistent on CWQ; on WebQSP, GoG also exceeds StructGPT and ToG at each reported setting. All methods lose accuracy as more crucial triples are removed, so the results indicate relative robustness rather than immunity to incompleteness.

    Dataset Method CKG IKG-20% IKG-40% IKG-60% IKG-80%
    CWQ ToG 47.2 40.5 37.9 33.7 31.4
    CWQ GoG 55.7 44.9 44.3 36.2 34.4
    WebQSP StructGPT 76.0 67.8 60.1 51.7 43.7
    WebQSP ToG 76.9 70.3 61.4 60.6 55.9
    WebQSP GoG 78.7 70.8 66.6 62.6 56.5

  8. Knowl 8 — GoG performance varies with the LLM backbone and its graph-access setting

    data/table

    The table reports GoG Hits@1 scores (percent) for four LLM backbones using complete graphs (CKG), 40%-incomplete graphs (IKG-40%), and no graph (NKG). GPT-4 gives the highest reported GoG score on both datasets with CKG and IKG-40%. The relative ordering of Qwen-1.5 and Llama-3 changes by graph setting: Qwen-1.5 scores higher with CKG, while Llama-3 scores higher with NKG. The authors take this contrast as evidence that backbone strength as a graph-searching agent need not match its strength as a source of internal knowledge.

    Dataset GoG backbone CKG IKG-40% NKG
    CWQ GPT-3.5 55.7 44.3 38.8
    CWQ Qwen-1.5 63.3 49.2 47.0
    CWQ Llama-3 59.6 54.6 54.0
    CWQ GPT-4 75.2 60.4 55.6
    WebQSP GPT-3.5 78.7 66.6 62.6
    WebQSP Qwen-1.5 77.9 70.2 65.1
    WebQSP Llama-3 77.4 74.4 70.8
    WebQSP GPT-4 84.4 80.3 75.7

  9. Knowl 9 — Ablations support using both generated facts and a focused evidence subgraph

    empirical result

    Removing GoG’s Generate action lowers Hits@1 on both datasets under both complete and IKG-40% settings, supporting the utility of supplementing graph search with LLM-generated triples. The method without generation remains competitive because it relies only on graph evidence, avoiding errors from generated facts while continuing to explore the graph. A separate experiment varies the number of BM25-selected related triples supplied to generation, using Qwen-1.5-72B-Chat: performance generally improves as relevant triples are added, but in most settings it first rises and then falls as more triples introduce noisy or unrelated information.

    Dataset GoG variant CKG Hits@1 (%) IKG-40% Hits@1 (%)
    CWQ Without Generate 62.7 48.6
    CWQ With Generate 63.3 50.6
    WebQSP Without Generate 74.7 69.4
    WebQSP With Generate 77.9 71.1

  10. Knowl 10 — GoG remains vulnerable to hallucination and extreme graph incompleteness

    limitation

    GoG can accept incorrect facts from its Generate action, and the paper identifies hallucination as an unavoidable risk with the evaluated LLMs. Its performance also has room to improve: when the graph is very incomplete, GoG can score below Chain-of-Thought prompting without graph access. The error analysis identifies generation errors, decomposition errors (losing the original question’s goal over multiple search rounds), unsupported final answers, and alias-based false negatives. Excluding false negatives, hallucinations account for the largest share of actual errors; alias mismatches are more likely under incomplete graphs because generated answers may use a different name from the reference.

Coverage note — Detailed per-example case studies are omitted because they illustrate GoG’s behaviors and failure modes rather than adding generalizable methods or results.

References

  1. 1.Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.
  2. 2.Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity. ArXiv:2302.04023 [cs].
  3. 3.Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 1247–1250.
  4. 4.Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. ArXiv:2005.14165 [cs].
  5. 5.Sitao Cheng, Ziyuan Zhuang, Yong Xu, Fangkai Yang, Chaoyun Zhang, Xiaoting Qin, Xiang Huang, Ling Chen, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, and Qi Zhang. 2024. Call me when necessary: LLMs can efficiently and faithfully reason over structured environments.
  6. 6.Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, Lei Li, and Zhifang Sui. 2023. A Survey on In-context Learning. ArXiv:2301.00234 [cs].
  7. 7.Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2023. Complexity-Based Prompting for Multi-Step Reasoning. ArXiv:2210.00720 [cs].
  8. 8.Qimeng Guo, Xue Wang, Zhenfang Zhu, Peiyu Liu, and Liancheng Xu. 2023. A knowledge inference model for question answering on an incomplete knowledge graph. Applied Intelligence, 53(7):7634–7646.
  9. 9.Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2023. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ArXiv:2311.05232 [cs].
  10. 10.Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and Philip S. Yu. 2021. A Survey on Knowledge Graphs: Representation, Acquisition and Applications. arXiv:2002.00388 [cs]. ArXiv: 2002.00388.
  11. 11.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. In EMNLP 2023. arXiv. ArXiv:2305.09645 [cs].
  12. 12.Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. 2023. Decomposed Prompting: A Modular Approach for Solving Complex Tasks. In NIPS 2023. arXiv. ArXiv:2210.02406 [cs].
  13. 13.Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023a. HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models. ArXiv:2305.11747 [cs].
  14. 14.Shiyang Li, Yifan Gao, Haoming Jiang, Qingyu Yin, Zheng Li, Xifeng Yan, Chao Zhang, and Bing Yin. 2023b. Graph Reasoning for Question Answering with Triplet Retrieval. ArXiv:2305.18742 [cs].
  15. 15.Tianle Li, Xueguang Ma, Alex Zhuang, Yu Gu, Yu Su, and Wenhu Chen. 2023c. Few-shot In-context Learning on Knowledge Base Question Answering. In ACL 2023, pages 6966–6980, Toronto, Canada. Association for Computational Linguistics.
  16. 16.Xingxuan Li, Ruochen Zhao, Yew Ken Chia, Bosheng Ding, Shafiq Joty, Soujanya Poria, and Lidong Bing. 2023d. Chain-of-Knowledge: Grounding Large Language Models via Dynamic Knowledge Adapting over Heterogeneous Sources. ArXiv:2305.13269 [cs].
  17. 17.Haoran Luo, Zichen Tang, Shiyao Peng, Yikai Guo, Wentai Zhang, Chenghao Ma, Guanting Dong, Meina Song, Wei Lin, et al. 2024. Chatkbqa: A generate-then-retrieve framework for knowledge base question answering with fine-tuned large language models. arXiv preprint arXiv:2310.08975.
  18. 18.Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. 2023. Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning. ArXiv:2310.01061 [cs].
  19. 19.Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. 2023. Unifying Large Language Models and Knowledge Graphs: A Roadmap. ArXiv:2306.08302 [cs].
  20. 20.Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: Bm25 and beyond. Found. Trends Inf. Retr., 3(4):333–389.
  21. 21.Mikhail Salnikov, Hai Le, Prateek Rajput, Irina Nikishina, Pavel Braslavski, Valentin Malykh, and Alexander Panchenko. 2023. Large language models meet knowledge graphs to answer factoid questions. arXiv preprint arXiv:2310.02166.
  22. 22.Apoorv Saxena, Aditay Tripathi, and Partha Talukdar. 2020. Improving multi-hop question answering over knowledge graphs using knowledge base embeddings. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 4498–4507.
  23. 23.Haitian Sun, Tania Bedrax-Weiss, and William W Cohen. 2019. Pullnet: Open domain question answering with iterative retrieval on knowledge bases and text. arXiv preprint arXiv:1904.09537.
  24. 24.Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel M. Ni, Heung-Yeung Shum, and Jian Guo. 2023. Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph. ArXiv:2307.07697 [cs].
  25. 25.Alon Talmor and Jonathan Berant. 2018. The web as a knowledge-base for answering complex questions. arXiv preprint arXiv:1803.06643.
  26. 26.Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.
  27. 27.Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. ArXiv:2203.11171 [cs].
  28. 28.Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. ArXiv:2201.11903 [cs].
  29. 29.Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. ArXiv:2210.03629 [cs].
  30. 30.Wen-tau Yih, Matthew Richardson, Christopher Meek, Ming-Wei Chang, and Jina Suh. 2016a. The value of semantic parse labeling for knowledge base question answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 201–206.
  31. 31.Wen-tau Yih, Matthew Richardson, Christopher Meek, Ming-Wei Chang, and Jina Suh. 2016b. The value of semantic parse labeling for knowledge base question answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 201–206.
  32. 32.Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context generators.
  33. 33.Daoguang Zan, Sirui Wang, Hongzhi Zhang, Kun Zhou, Wei Wu, Wayne Xin Zhao, Bingchao Wu, Bei Guan, and Yongji Wang. 2022. Complex question answering over incomplete knowledge graph as n-ary link prediction. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE.
  34. 34.Fen Zhao, Yinguo Li, Jie Hou, and Ling Bai. 2022. Improving question answering over incomplete knowledge graphs with relation prediction. Neural Computing and Applications, pages 1–18.

Citation

MLA
Xu, Y., et al. “Generate-on-Graph: Treat LLM as Both Agent and KG for Incomplete Knowledge Graph Question Answering”. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024, pp. 18410–30, https://doi.org/10.18653/v1/2024.emnlp-main.1023.
APA
Xu, Y., (何世柱), S. H., Chen, J., Wang, Z., Song, Y., Tong, H., Liu, G., Zhao, J., & Liu, K. (2024). Generate-on-Graph: Treat LLM as both Agent and KG for Incomplete Knowledge Graph Question Answering. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 18410–18430. https://doi.org/10.18653/v1/2024.emnlp-main.1023
Chicago
Xu, Y., S. H. (何世柱), J. Chen, et al. 2024. “Generate-on-Graph: Treat LLM as Both Agent and KG for Incomplete Knowledge Graph Question Answering”. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 18410–30. https://doi.org/10.18653/v1/2024.emnlp-main.1023.
Harvard
Xu, Y. et al. (2024) “Generate-on-Graph: Treat LLM as both Agent and KG for Incomplete Knowledge Graph Question Answering”, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp. 18410–18430. Available at: https://doi.org/10.18653/v1/2024.emnlp-main.1023.
Vancouver
1. Xu Y, (何世柱) SH, Chen J, Wang Z, Song Y, Tong H, Liu G, Zhao J, Liu K (2024) Generate-on-Graph: Treat LLM as both Agent and KG for Incomplete Knowledge Graph Question Answering. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pp 18410–18430

BibTeX

@inproceedings{xu-etal-2024-generate,
    title = "Generate-on-Graph: Treat {LLM} as both Agent and {KG} for Incomplete Knowledge Graph Question Answering",
    author = "Xu, Yao  and
      He, Shizhu  and
      Chen, Jiabei  and
      Wang, Zihao  and
      Song, Yangqiu  and
      Tong, Hanghang  and
      Liu, Guang  and
      Zhao, Jun  and
      Liu, Kang",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.1023/",
    doi = "10.18653/v1/2024.emnlp-main.1023",
    pages = "18410--18430"
}
Metadata:ACL Anthology

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: https://creativecommons.org/licenses/by/4.0/