Pith. sign in

REVIEW 3 major objections 6 minor 54 references

Open-domain triplet extraction improves when the LLM optimizes its own prompt from restoration-based feedback.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 08:58 UTC pith:CUHW2KBB

load-bearing objection Same-corpus tuning means the reported F1 gains are not yet evidence of generalization, but KRPO's design is a coherent, novel framework that deserves a careful held-out rerun. the 3 major comments →

arxiv 2601.15037 v2 pith:CUHW2KBB submitted 2026-01-21 cs.CL cs.AI

Knowledge Restoration-driven Prompt Optimization: Unlocking LLM Potential for Open-Domain Relational Triplet Extraction

Xiaonan Jing , Gongqing Wu , Xingrui Zhuo , Lang Sun , Jiapu Wang This is my paper
classification cs.CL cs.AI
keywords open-domain relational triplet extractionprompt optimizationknowledge restorationself-evaluationnatural language inferencerelation canonicalizationknowledge graph constructionLLM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that open-domain relational triplet extraction—pulling (subject, relation, object) triples out of text without a predefined relation list—does not have to rely on fixed, hand-written prompts. It proposes KRPO, a two-phase, label-free framework: in phase one, the LLM's extracted triplets are turned back into natural-language sentences and scored against the original input by natural language inference; those scores become 'textual gradients' that rewrite the extraction prompt batch by batch. In phase two, raw relations are aligned to a dynamically growing memory of canonical relation schemas using a cross-encoder plus an LLM judge. The paper reports that this pipeline beats a strong static-prompt baseline on three benchmarks and five backbones, with the largest gains under strict/exact matching and on smaller models. A sympathetic reader would care because it offers a path to self-adapting LLM extraction without any gold labels.

Core claim

On its own terms, the paper's central claim is that the supervision deficit in open-domain extraction can be overcome by substituting intrinsic consistency for ground truth. Each extracted triplet is restored to a sentence (e.g., 'West Lake lies in Hangzhou'), and an NLI step labels it entailment (+1), neutral (0), or contradiction (−0.5) against the source sentence. This aggregate score is decomposed into two LLM-generated textual gradients: one says how to improve the evaluation metrics, the second how to improve the ORTE prompt, and a third instruction rewrites the prompt from a batch of such gradients. The paper further claims that a cross-encoder trained on restored relation schemas, co

What carries the argument

The load-bearing mechanism is a closed loop of knowledge restoration and textual-gradient prompt updates. Restoration turns a triplet into a sentence and masks subject/object to form a relation schema; NLI then supplies a scalar score in {1, 0, −0.5}. The score is treated as a gradient with respect to the sentence, propagated through the triplets to the prompt via two chained LLM prompts (I1 and I2), and a third prompt (I3) rewrites the extraction prompt on a batch of accumulated gradients. The companion mechanism is a Memory-augmented Relation Canonicalizer: a cross-encoder scores masked relation schemas against a schema memory, the top-K candidates are presented to an LLM, and if none fits

Load-bearing premise

The evaluation optimizes the prompt on the same unlabeled sentences it later scores, so the reported F1 gains may come from adapting to those exact sentences rather than from a prompt that generalizes to unseen text.

What would settle it

Run Phase 1 prompt optimization on one split of the corpus and evaluate the optimized prompt on a disjoint held-out split; if the F1 advantage over the static-prompt baseline disappears, the generalization claim is not supported.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • LLM-based triplet extraction can improve without gold annotations, using consistency between the source sentence and restored triplet text as the only feedback.
  • Prompt optimization and relation canonicalization compound: strict/exact F1 improves more than partial F1, indicating better structural fidelity, not just more surface matches.
  • Smaller LLMs benefit more from the loop, narrowing the gap to larger models in extraction tasks.
  • Dynamic schema memory plus an LLM decision step reduces relation redundancy better than static embedding similarity, which matters for knowledge-graph consistency.
  • The performance gap over a static prompt widens as more samples are processed, suggesting the prompt accumulates useful experience rather than plateauing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because Phase 1 optimizes the prompt on the same corpus X that Phase 2 evaluates, the headline numbers should be read as transductive adaptation; a held-out split would test whether the optimized prompt generalizes to unseen sentences.
  • The restoration→NLI→gradient recipe is not tied to triplet extraction; the same loop could be applied to other under-supervised structured outputs, such as event argument extraction or table filling.
  • Replacing the LLM NLI judge with a standard NLI model would isolate whether the gains come from the feedback signal itself or from the judge's competence—a direct ablation the paper does not run.
  • One testable risk: NLI-based rewards may favor conservative, textually explicit triplets over valid inferences, so precision could rise while recall of implicit relations drops on corpora where relations are implied.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes KRPO, a two-phase, label-free framework for open-domain relational triplet extraction (ORTE). In Phase 1, an LLM extracts triplets using an initial prompt; each triplet is restored to a natural-language sentence and scored against the source text by an NLI-style self-evaluation (Eqs. 7–9); these scores are converted via LLM-generated textual gradients into an updated prompt. In Phase 2, the optimized prompt is used for extraction, and a memory-augmented relation canonicalizer (XLM-RoBERTa Cross-Encoder plus an LLM decision module) aligns raw relations to a dynamically expanded schema memory. The authors evaluate on WebNLG, REBEL, and Wiki-NRE with five LLM backbones, reporting F1 gains over EDC/REGEN/GenIE, and provide ablations for the prompt optimization and relation canonicalization modules.

Significance. KRPO targets a real limitation of LLM-based ORTE: static prompts cannot adapt to open-domain variation, and free-form relation surface forms create redundancy. The knowledge-restoration self-evaluation idea is interesting and practical, and the memory-augmented canonicalizer with Cross-Encoder scoring is a sensible extension of embedding-based canonicalization. The manuscript is broad in scope: three datasets, five LLM backbones, three matching settings, and detailed appendices with ablations and case studies. However, the central empirical claim is not currently supported by the evaluation protocol. Because Phase 1 prompt optimization is run on the same corpus on which F1 is later measured, with no held-out split, the reported gains are compatible with transductive adaptation to the exact test sentences rather than with a prompt that transfers to unseen text. The self-evaluation analysis also uses the very quantity being optimized as evidence of extraction quality. A properly held-out evaluation would substantially strengthen the paper; as submitted, the main generalization claim is not yet established.

major comments (3)
  1. [Algorithm 1, lines 2–30; §5.1] Phase 1 optimizes the prompt over the unlabeled corpus X in batches (lines 2–15), and Phase 2 then extracts triplets from the same X using the optimized prompt (lines 18–30). No held-out split, cross-validation, or separate optimization corpus is described anywhere in §5 or Appendix C. Since EDC is a static-prompt baseline that never adapts to X, the F1 comparisons in Tables 1/4–6 conflate 'a prompt that transfers to new text' with 'a prompt that has been tuned on the exact sentences being scored.' This directly bears on the abstract's 'generalization ability' claim and §5.2's 'consistently outperforms EDC' claim. Please either (i) evaluate on a held-out test set after optimization on a disjoint corpus (or nested cross-validation), reporting F1 on never-seen sentences, or (ii) explicitly reframe the contribution as transductive on-corpus adaptation and remove the generalization claim. Th
  2. [§5.4, Table 2; Eqs. (7)–(9)] The 'Entailment Triplet Analysis' uses the entailment proportion produced by the same NLI-based self-evaluation mechanism that serves as the optimization signal in Phase 1. Arguing that KRPO 'retains more entailment triplets' and therefore that 'self-evaluation is better aligned' is circular, because the optimizer explicitly maximizes that quantity. The external F1 columns in Table 2 are relevant, but the textual claim rests on the optimized proxy. To support the self-evaluation quality claim, the authors should report agreement between the LLM-based NLI labels and human judgments on a sample, or compare against an independently trained NLI model, and show that improved entailment proportion corresponds to improved extraction fidelity under a non-circular measure.
  3. [Table 1, Tables 4–6] Several cells in the main results tables are malformed with concatenated numbers and missing delimiters, e.g., Table 1 row Deepseek-V3: '75.255.2 50.552.0' and '67.467.465.2'; REBEL row Deepseek-V3: '50.551.7 52.752.0'; Wiki-NRE row Deepseek-V3: '67.8 68.3 68.0 67.3 67.7 67.4' with missing separators in places. As typeset, the central empirical claim cannot be verified from the tables. In addition, no standard deviation, confidence interval, or significance test is reported for any F1 number, despite stochastic LLM sampling and API variability. Please regenerate the tables with clear cell boundaries and include at least three independent runs (or bootstrap CIs) for the main KRPO-versus-EDC comparison.
minor comments (6)
  1. [Abstract and §1] The title and abstract use 'Knowledge Restoration-driven Prompt Optimization' and 'knowledge restoration', while the Introduction introduces 'Knowledge Reconstruction-driven Prompt Optimization'. Please unify the terminology.
  2. [Footnote 1] The anonymous code link is garbled/corrupted in the rendered text. A clean, working link is needed for reproducibility.
  3. [Appendix A.2, Eq. (19)] The Bayesian justification is heuristic and contains notational slips: the text says 'optimizing the reverse probability P_M(T | x)' where the intended quantity is P_M(P | T, x), and the approximation drops the prompt-dependent prior 1/P_M(P|x) without discussion. A clearer statement that this is a design assumption rather than a formal equivalence would be appropriate.
  4. [Eq. (9)] The notation 'Metrics(\hat T)' uses M both for the LLM and for the number of triplets. Using a different symbol for the count would improve readability.
  5. [Figure 5 and §5.5] The case study demonstrates the optimized prompt on the same sentence that was used to generate the optimization feedback. This is consistent with the current protocol but should be explicitly framed as on-corpus adaptation if no held-out evaluation is added, and the caption should note the sentence is from the optimization corpus.
  6. [§5.1] The batch size is given as 5 and Top-K as 5, but the number of optimization iterations/epochs over the corpus is not specified. Please state how many passes over X Phase 1 performs.

Circularity Check

1 steps flagged

Ancillary self-evaluation validation is circular; main F1 comparison is external.

specific steps
  1. fitted input called prediction [Section 5.4 (Impact of Self-evaluation in Prompt Optimization), Table 2; with Eq. (8)-(9) and Algorithm 1 lines 5-13]
    "To evaluate the reliability of self-evaluation, we analyze triplets classified as entailment against a static prompt baseline EDC under partial matching. ... KRPO consistently retains more entailment triplets and achieves higher F1 scores across all datasets, whereas EDC shows limited gains, confirming that self-evaluation is better aligned with the optimized prompt, enabling more accurate discrimination between correct and incorrect triplets. (Table 2 caption: "Prop. is the entailment proportion by self-evaluation.")"

    The entailment proportion is exactly the quantity optimized in Phase 1: Algorithm 1 line 6 computes S = sum_t NLI(x, Restore(t)) and line 13 updates the prompt from gradients derived from S (Eqs. 8-9). Thus finding that KRPO has a higher entailment proportion than the static EDC baseline is a by-construction consequence of the optimization, not an independent confirmation of extraction quality. The same self-scored NLI labels serve as both the training signal and the evidence in Table 2, so the claim that this 'confirms' better alignment is circular; only the simultaneously reported gold-F1 numbers are external evidence.

full rationale

The main empirical claim—KRPO outperforms EDC in F1—is evaluated against gold ground truth and compares an adapted prompt to a static-prompt baseline, so it is not circular. The optimization signal (NLI-based entailment) is a proxy rather than the evaluation metric. The circular component is restricted to Section 5.4/Table 2, where the paper validates the self-evaluation mechanism using the entailment proportion, which is precisely the objective maximized in Phase 1 (Eqs. 8-9; Algorithm 1 lines 6, 13). Higher entailment proportion after optimization is expected by construction, so it cannot independently confirm 'better alignment' or 'more accurate discrimination.' The F1 column in Table 2 is external, but the surrounding text treats the self-scored proportion as confirmatory, which is circular. I do not count the lack of a held-out split (Algorithm 1 optimizes over the same X used for final extraction) as circularity, though it is a validity limitation on the 'generalization ability' wording. There are no load-bearing self-citations or imported uniqueness theorems.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical or formal entities are introduced. The framework's load-bearing premises are the five axioms above, with the same-corpus evaluation protocol being the most consequential. The free parameters are small in number, but the NLI score mapping and Top-K/batch-size choices directly shape the prompt-optimization signal and the canonicalization behavior.

free parameters (4)
  • NLI score mapping = +1 / 0 / -0.5
    Hand-chosen discrete weights for entailment/neutral/contradiction (Eq. 8); directly sets the magnitude of the self-evaluation signal that drives prompt updates.
  • Batch size B = 5
    Used in Algorithm 1 for aggregating textual gradients before updating the prompt; no sensitivity analysis is reported.
  • Top-K candidate relations = 5
    Number of canonical relation candidates retrieved by the cross-encoder before LLM decision (Section 5.1); only a partial k-sweep is shown.
  • Ranking-loss margin m = not reported
    Margin hyperparameter in the cross-encoder ranking loss (Appendix B.1); unspecified, so the canonicalizer training is not fully reproducible.
axioms (5)
  • domain assumption LLM-based NLI between original text and restored triplet is a reliable proxy for triplet faithfulness in the absence of gold labels.
    Section 4.1, Eqs. 7–9; the whole prompt-optimization loop depends on this signal being informative and not drifting.
  • ad hoc to paper arg max_P P_M(P|T,x) approximates arg max_P P_M(T|x,P) because the language prior log P_M(P|x) is a weak regularizer.
    Appendix A.2, Eqs. 17–19; this is a heuristic transfer of objective, not a proof, and is used to justify optimizing prompts rather than outputs.
  • domain assumption LLM-generated textual feedback (I1, I2, I3) behaves like a gradient and improves the ORTE prompt enough to help F1.
    Section 4.1, Eqs. 10–13; there is no guarantee LLM feedback optimizes the prompt, and overfitting to the self-evaluation signal is possible.
  • ad hoc to paper Optimizing on the unlabeled evaluation corpus and measuring F1 on the same corpus is a valid evaluation protocol.
    Algorithm 1 Phase 1 and Phase 2 both operate on dataset X; no held-out split is described in Section 5.1.
  • domain assumption A cross-encoder fine-tuned on TEKGEN yields semantic relevance scores that transfer to ORTE relation canonicalization.
    Appendix B.1 constructs the training set from TEKGEN; transfer to WebNLG/REBEL/Wiki-NRE relations is assumed, not separately validated.

pith-pipeline@v1.3.0-alltime-deepseek · 24872 in / 13836 out tokens · 140838 ms · 2026-08-03T08:58:22.930776+00:00 · methodology

0 comments
read the original abstract

Open-domain Relational Triplet Extraction (ORTE) aims to mine structured knowledge without predefined relation schemas. Large Language Models (LLMs) have advanced ORTE toward a prompt-driven paradigm through powerful in-context learning. However, adapting their extraction behavior to varying open-domain contexts remains challenging. Existing methods typically rely on manually crafted prompts that remain fixed across inputs, despite substantial variation in linguistic expressions and contextual structures. This mismatch may lead to unsupported triplets, while the absence of ground-truth annotations makes such deficiencies difficult to identify and correct. Moreover, free-form relation generation produces non-canonical relation surface forms, undermining knowledge graph consistency. To address these challenges, we propose Knowledge Restoration-driven Prompt Optimization (KRPO), a framework for label-free target-corpus adaptation. KRPO restores extracted triplets into textual statements and evaluates their semantic consistency with the source inputs, deriving intrinsic feedback without gold annotations. This feedback is transformed into natural-language optimization guidance for batch-wise prompt optimization and adaptation. KRPO further introduces a Memory-augmented Relation Canonicalizer that aligns free-form relations with a dynamically updated schema memory, improving relation consistency. Experiments on three ORTE benchmarks with multiple LLM backbones demonstrate strong overall performance, with KRPO achieving the best average F1 score across the evaluated settings.

Figures

Figures reproduced from arXiv: 2601.15037 by Gongqing Wu, Jiapu Wang, Lang Sun, Xiaonan Jing, Xingrui Zhuo.

Figure 1
Figure 1. Figure 1: Comparison between KRPO (Ours) and EDC [Zhang and Soh, 2024]. KRPO dynamically optimizes prompts via self￾reflection, releasing LLM’s potential for ORTE, whereas EDC relies on static prompts and suffers from performance stagnation. Pan et al., 2017], providing structured knowledge for down￾stream tasks such as information retrieval [Xiong et al., 2017], logical reasoning [Zhuo et al., 2024], and ques￾tion … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of KRPO. Comprises four modules: (1) Relational Triplet Extraction (RTE), extracting triplets by LLM with an opti [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study on WebNLG, REBEL, and Wiki-NRE. We [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: A case study comparing the extraction results between the Initial and Updated prompts. The optimized prompt (right) incorporates [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: F1 Evolution over Samples with Iterative Prompt Opti [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 5 linked inside Pith

  1. [1]

    Knowledge graph based synthetic corpus generation for knowledge-enhanced language model pre-training

    Oshin Agarwal, Heming Ge, Siamak Shakeri, and Rami Al - Rfou. Knowledge graph based synthetic corpus generation for knowledge-enhanced language model pre-training. In NAACL , pages 3554--3565, 2021

    2021

  2. [2]

    Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni

    Michele Banko, Michael J. Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni. Open information extraction from the web. In IJCAI , pages 2670--2676, 2007

    2007

  3. [3]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, et al. Language models are few-shot learners. In NeurIPS , 2020

    2020

  4. [4]

    REBEL: relation extraction by end-to-end language generation

    Pere - Llu \' s Huguet Cabot and Roberto Navigli. REBEL: relation extraction by end-to-end language generation. In EMNLP , pages 2370--2381, 2021

    2021

  5. [5]

    Exploiting syntactico-semantic structures for relation extraction

    Yee Seng Chan and Dan Roth. Exploiting syntactico-semantic structures for relation extraction. In ACL , pages 551--560, 2011

    2011

  6. [6]

    SAC-KG: exploiting large language models as skilled automatic constructors for domain knowledge graph

    Hanzhu Chen, Xu Shen, Qitan Lv, Jie Wang, Xiaoqi Ni, and Jieping Ye. SAC-KG: exploiting large language models as skilled automatic constructors for domain knowledge graph. In ACL , pages 4345--4360, 2024

    2024

  7. [7]

    BGE m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. BGE m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. CoRR , abs/2402.03216, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Overview of MUC-7

    Nancy Chinchor. Overview of MUC-7 . In Seventh Message Understanding Conference (MUC-7) , 1998

    1998

  9. [9]

    Unsupervised cross-lingual representation learning at scale

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \' a n, et al. Unsupervised cross-lingual representation learning at scale. In ACL , pages 8440--8451, 2020

    2020

  10. [10]

    Rlprompt: Optimizing discrete text prompts with reinforcement learning

    Mingkai Deng, Jianyu Wang, Cheng - Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, et al. Rlprompt: Optimizing discrete text prompts with reinforcement learning. In EMNLP , pages 3369--3391, 2022

    2022

  11. [11]

    Span-level model for relation extraction

    Kalpit Dixit and Yaser Al - Onaizan. Span-level model for relation extraction. In ACL , pages 5308--5314, 2019

    2019

  12. [12]

    Dognin, Inkit Padhi, Igor Melnyk, and Payel Das

    Pierre L. Dognin, Inkit Padhi, Igor Melnyk, and Payel Das. Regen: Reinforcement learning for text and knowledge base generation using pretrained language models. In EMNLP , pages 1084--1099, 2021

    2021

  13. [13]

    From local to global: A graph RAG approach to query-focused summarization

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. From local to global: A graph RAG approach to query-focused summarization. CoRR , abs/2404.16130, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    The 2020 bilingual, bi-directional webnlg+ shared task overview and evaluation results (webnlg+ 2020)

    Thiago Castro Ferreira, Claire Gardent, Nikolai Ilinykh, Chris Van Der Lee, Simon Mille, Diego Moussallem, and Anastasia Shimorina. The 2020 bilingual, bi-directional webnlg+ shared task overview and evaluation results (webnlg+ 2020). In Proceedings of the 3rd International Workshop on Natural Language Generation from the Semantic Web (WebNLG+) , 2020

    2020

  15. [15]

    Suchanek

    Luis Gal \' a rraga, Geremy Heitz, Kevin Murphy, and Fabian M. Suchanek. Canonicalizing open knowledge bases. In CIKM , pages 1679--1688, 2014

    2014

  16. [16]

    PPDB: the paraphrase database

    Juri Ganitkevitch, Benjamin Van Durme, and Chris Callison - Burch. PPDB: the paraphrase database. In NAACL , pages 758--764, 2013

    2013

  17. [17]

    Table filling multi-task recurrent neural network for joint entity and relation extraction

    Pankaj Gupta, Hinrich Sch \" u tze, and Bernt Andrassy. Table filling multi-task recurrent neural network for joint entity and relation extraction. In COLING , pages 2537--2547, 2016

    2016

  18. [18]

    Knowledge Graphs

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia d'Amato, Gerard de Melo, Claudio Gutierrez, et al. Knowledge Graphs . Synthesis Lectures on Data, Semantics, and Knowledge. 2021

    2021

  19. [19]

    Relink: Constructing query-driven evidence graph on-the-fly for graphrag

    Manzong Huang, Chenyang Bu, Yi He, Xingrui Zhuo, and Xindong Wu. Relink: Constructing query-driven evidence graph on-the-fly for graphrag. arXiv preprint arXiv:2601.07192 , 2026

    arXiv 2026

  20. [20]

    Xu, Jun Araki, and Graham Neubig

    Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. How can we know what language models know. Trans. Assoc. Comput. Linguistics , 8:423--438, 2020

    2020

  21. [21]

    Genie: Generative information extraction

    Martin Josifoski, Nicola De Cao, Maxime Peyrard, Fabio Petroni, and Robert West. Genie: Generative information extraction. In NAACL , pages 4626--4643, 2022

    2022

  22. [22]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al - Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP , pages 3045--3059, 2021

    2021

  23. [23]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL/IJCNLP , pages 4582--4597, 2021

    2021

  24. [24]

    P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In ACL , pages 61--68, 2022

    2022

  25. [25]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv. , 55(9):195:1--195:35, 2023

    2023

  26. [26]

    George A. Miller. Wordnet: A lexical database for english. Commun. ACM , 38(11):39--41, 1995

    1995

  27. [27]

    Kanatsoulis, and Sanmi Koyejo

    Belinda Mo, Kyssen Yu, Joshua Kazdan, Proud Mpala, Lisa Yu, Chris Cundy, Charilaos I. Kanatsoulis, and Sanmi Koyejo. Kggen: Extracting knowledge graphs from plain text with language models. CoRR , abs/2502.09956, 2025

    arXiv 2025

  28. [28]

    Pan, Guido Vetere, Jos \' e Manu \' e l G \' o mez - P \' e rez, and Honghan Wu

    Jeff Z. Pan, Guido Vetere, Jos \' e Manu \' e l G \' o mez - P \' e rez, and Honghan Wu. Exploiting Linked Data and Knowledge Graphs in Large Organisations . 2017

    2017

  29. [29]

    Unifying large language models and knowledge graphs: A roadmap

    Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. Unifying large language models and knowledge graphs: A roadmap. IEEE Trans. Knowl. Data Eng. , 36(7):3580--3599, 2024

    2024

  30. [30]

    Aligning open IE relations and KB relations using a siamese network based on word embedding

    Rifki Afina Putri, Giwon Hong, and Sung - Hyon Myaeng. Aligning open IE relations and KB relations using a siamese network based on word embedding. In IWCS , pages 142--153, 2019

    2019

  31. [31]

    Onerel: Joint entity and relation extraction with one module in one step

    Yuming Shang, Heyan Huang, and Xianling Mao. Onerel: Joint entity and relation extraction with one module in one step. In AAAI , pages 11285--11293, 2022

    2022

  32. [32]

    Logan IV, Eric Wallace, and Sameer Singh

    Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In EMNLP , pages 4222--4235, 2020

    2020

  33. [33]

    Overview of results of the MUC-6 evaluation

    Beth M Sundheim. Overview of results of the MUC-6 evaluation . In Sixth Message Understanding Conference (MUC-6) , 1995

    1995

  34. [34]

    Neural relation extraction for knowledge base enrichment

    Bayu Distiawan Trisedya, Gerhard Weikum, Jianzhong Qi, and Rui Zhang. Neural relation extraction for knowledge base enrichment. In ACL , pages 229--240, 2019

    2019

  35. [35]

    Talukdar

    Shikhar Vashishth, Prince Jain, and Partha P. Talukdar. CESI: canonicalizing open knowledge bases using embeddings and side information. In WWW , pages 1317--1327, 2018

    2018

  36. [36]

    Somin Wadhwa, Silvio Amir, and Byron C. Wallace. Revisiting relation extraction in the era of large language models. In ACL , pages 15566--15589, 2023

    2023

  37. [37]

    Tplinker: Single-stage joint extraction of entities and relations through token pair linking

    Yucheng Wang, Bowen Yu, Yueyang Zhang, Tingwen Liu, Hongsong Zhu, and Limin Sun. Tplinker: Single-stage joint extraction of entities and relations through token pair linking. In COLING , pages 1572--1582, 2020

    2020

  38. [38]

    A survey on temporal knowledge graph completion: Taxonomy, progress, and prospects

    Jiapu Wang, Boyue Wang, Meikang Qiu, Shirui Pan, Bo Xiong, Heng Liu, Linhao Luo, Tengfei Liu, Yongli Hu, Baocai Yin, et al. A survey on temporal knowledge graph completion: Taxonomy, progress, and prospects. arXiv preprint arXiv:2308.02457 , 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Ime: Integrating multi-curvature shared and specific embedding for temporal knowledge graph completion

    Jiapu Wang, Zheng Cui, Boyue Wang, Shirui Pan, Junbin Gao, Baocai Yin, and Wen Gao. Ime: Integrating multi-curvature shared and specific embedding for temporal knowledge graph completion. In WWW , pages 1954--1962, 2024

    1954

  40. [40]

    Large language models-guided dynamic adaptation for temporal knowledge graph reasoning

    Jiapu Wang, Kai Sun, Linhao Luo, Wei Wei, Yongli Hu, Alan W Liew, Shirui Pan, and Baocai Yin. Large language models-guided dynamic adaptation for temporal knowledge graph reasoning. NeurIPS , 37:8384--8410, 2024

    2024

  41. [41]

    Zero-shot information extraction via chatting with chatgpt

    Xiang Wei, Xingyu Cui, Ning Cheng, Xiaobin Wang, Xin Zhang, Shen Huang, Pengjun Xie, Jinan Xu, Yufeng Chen, Meishan Zhang, Yong Jiang, and Wenjuan Han. Zero-shot information extraction via chatting with chatgpt. CoRR , abs/2302.10205, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    Explicit semantic ranking for academic search via knowledge graph embedding

    Chenyan Xiong, Russell Power, and Jamie Callan. Explicit semantic ranking for academic search via knowledge graph embedding. In WWW , pages 1271--1279, 2017

    2017

  43. [43]

    Large language models for generative information extraction: a survey

    Derong Xu, Wei Chen, Wenjun Peng, Chao Zhang, Tong Xu, Xiangyu Zhao, et al. Large language models for generative information extraction: a survey. Frontiers Comput. Sci. , 18(6):186357, 2024

    2024

  44. [44]

    Le, Denny Zhou, and Xinyun Chen

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. CoRR , abs/2309.03409, 2023

    Pith/arXiv arXiv 2023

  45. [45]

    u ksekg \

    Mert Y \" u ksekg \" o n \" u l, Federico Bianchi, Joseph Boen, Sheng Liu, Pan Lu, Zhi Huang, Carlos Guestrin, and James Zou. Optimizing generative AI by backpropagating language model feedback. Nat. , 639(8055):609--616, 2025

    2025

  46. [46]

    Kernel methods for relation extraction

    Dmitry Zelenko, Chinatsu Aone, and Anthony Richardella. Kernel methods for relation extraction. J. Mach. Learn. Res. , 3:1083--1106, 2003

    2003

  47. [47]

    Extract, define, canonicalize: An llm-based framework for knowledge graph construction

    Bowen Zhang and Harold Soh. Extract, define, canonicalize: An llm-based framework for knowledge graph construction. In EMNLP , pages 9820--9836, 2024

    2024

  48. [48]

    Smola, and Le Song

    Yuyu Zhang, Hanjun Dai, Zornitsa Kozareva, Alexander J. Smola, and Le Song. Variational reasoning for question answering with knowledge graph. In AAAI , pages 6069--6076, 2018

    2018

  49. [49]

    Large language models are human-level prompt engineers

    Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In ICLR , 2023

    2023

  50. [50]

    Llms for knowledge graph construction and reasoning: recent capabilities and future opportunities

    Yuqi Zhu, Xiaohan Wang, Jing Chen, Shuofei Qiao, Yixin Ou, Yunzhi Yao, Shumin Deng, Huajun Chen, and Ningyu Zhang. Llms for knowledge graph construction and reasoning: recent capabilities and future opportunities. World Wide Web (WWW) , 27(5):58, 2024

    2024

  51. [51]

    Geometric-contextual mutual infomax path aggregation for relation reasoning on knowledge graph

    Xingrui Zhuo, Gongqing Wu, Zan Zhang, and Xindong Wu. Geometric-contextual mutual infomax path aggregation for relation reasoning on knowledge graph. IEEE Trans. Knowl. Data Eng. , 36(7):3076--3090, 2024

    2024

  52. [52]

    Progressive prefix-memory tuning for complex logical query answering on knowledge graphs

    Xingrui Zhuo, Shirui Pan, Jiapu Wang, Gongqing Wu, Zan Zhang, Rui Li, Zizhong Wei, and Xindong Wu. Progressive prefix-memory tuning for complex logical query answering on knowledge graphs. In IJCAI , pages 3716--3724, 2025

    2025

  53. [53]

    Effective instruction parsing plugin for complex logical query answering on knowledge graphs

    Xingrui Zhuo, Jiapu Wang, Gongqing Wu, Shirui Pan, and Xindong Wu. Effective instruction parsing plugin for complex logical query answering on knowledge graphs. In WWW , pages 4780--4792, 2025

    2025

  54. [54]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...