Pith. sign in

REVIEW 4 major objections 5 minor 31 references

AgREE: Agentic Reasoning for Knowledge Graph Completion on Emerging Entities

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A training-free agent that interleaves web searches with self-reflection can complete knowledge-graph triplets, beating fully trained baselines by up to 13.7% — and far more on post-cutoff entities.

desk verdict A training-free agent for KGC with a useful emerging-entity benchmark, but the headline gains are partly an artifact of an asymmetric answer prompt and the proposed metric has an internal inconsistency. read the letter →

arxiv 2508.04118 v1 pith:GI6PZPSU submitted 2025-08-06 cs.AI cs.CL

Ruochen Zhao , Simone Conia , Eric Peng , Min Li , Saloni Potdar This is my paper · ORCID
classification cs.AIcs.CL
keywords knowledgegraphcompletionemergingentitiesLLMagentsiterativeretrievalself-reflectionrelation-awareevaluationtraining-freewebsearch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's central claim is that knowledge-graph completion — filling in missing (subject, predicate, object) triplets — need not be a training problem at all. AgREE, an LLM agent that alternates web searches with self-reflection, beats fully trained baselines on the standard Wikidata5M and FB60K-NYT10 benchmarks (up to +13.7% Hits@$N$) with zero gradient updates, and on a new benchmark of 332 entities created between January and May 2025 — after the backbone model's July 2024 cutoff — it reaches 45.2% Hits@1 and 75.5% Hits@10 versus 30.2% for ReAct. The paper also argues that the standard Hits@$N$ metric misjudges one-to-many relations and proposes a relation-aware variant under which its advantage grows. A sympathetic reader would care because, if true, keeping a knowledge graph current becomes a live search-and-reason task rather than a recurring retraining expense.

What carries the argument

The load-bearing mechanism is the iterative loop that interleaves retrieval with self-reflection: after each tool call the agent judges whether the retrieved chunks answer the query, and detected insufficiency triggers a refined query or an escalation from the basic (Wikipedia) to the advanced (web search) retriever, capped at 20 iterations. Retrieved text is chunked into 3-sentence fragments, filtered by keyword overlap with the query, and re-ranked before the agent reads it. The paper's second key object is the relation-aware Hits@$N$ metric: for each relation $r$ it sets $N_{\mathrm{rel}} = \max_h |\{t : (h,r,t) \in G_{\mathrm{train}}\}|$ and evaluates a test triple only when $N \le N_{\m

What would settle it

Sample 100 random triples from the 3,664 emerging test instances and check whether, for each correct prediction, the ground-truth object appears verbatim in the chunks the agent retrieved before answering. If it does for nearly all correct answers, completion is document copying and the benchmark does not isolate reasoning. The mirror test: rerun AgREE with retrieval restricted to pages that mention the subject but never state the full triple — if accuracy collapses, the method needs the target written down; if it holds, genuine multi-step inference is demonstrated.

Watch

Extended reading notes

Core claim

AgREE is an LLM agent that turns a triplet-completion query (head, relation, missing tail) into a search-and-reason task. If parametric knowledge does not suffice, it issues queries, picks between a Wikipedia-backed and a web-search tool, re-ranks results, then self-reflects on sufficiency, escalating tools until satisfied. It then generates at least ten ranked candidates and entity-links them. The paper's claim: this zero-training loop beats fully trained KGE, PLM, and RL baselines on Wikidata5M and FB60K-NYT10 (up to +13.7% Hits@$N$), beats ReAct by 15 points at Hits@1 and 45 at Hits@10 on 332 post-cutoff entities, and its relation-aware Hits@$N$ widens the gap to 88.9–100% where ReAct sco

Load-bearing premise

The results rest on the premise that the emerging-entities benchmark isolates genuinely unseen facts: entities are selected because their Wikidata inception dates fall after the backbone model's July 2024 cutoff, but the agent retrieves from Wikipedia and Google — the same sources the ground-truth triples were mined from — so if a retrieved page simply states the gold triple, the high scores mostly reflect search-and-reading skill rather than knowledge-graph reasoning about e

Editorial extensions

If this is right

  • Keeping a knowledge graph current stops requiring periodic retraining: a search-and-reflect agent can absorb post-cutoff entities as soon as they are documented, at the cost of inference time rather than gradient updates.
  • Because the same zero-training agent matches or exceeds fully trained models on established benchmarks, it is a viable route for sparse or low-resource domains where training data are thin.
  • The self-reflection loop is the main moving part: removing it lowers Hits@1 (about 2 points on Wikidata5m, 17.7 points on the emerging set under relation-aware evaluation), and using only one retriever also hurts; the same ablation slightly raises Wikidata5m Hits@3, so the loop's benefit is concentrated at the top of the ranking.
  • Under the proposed relation-aware Hits@$N$, the gap to baselines widens on every dataset (to 88.9–100% versus ReAct on the emerging set), supporting the paper's case that standard Hits@$N$ under-rewards models that return several valid answers for one-to-many relations.
  • Backbone capability bounds the framework: swapping DeepSeek-V3 for a 12B model cuts Wikidata5m Hits@1 from 67.2% to 30.5%, so the loop is only as strong as the reasoner at its center.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the emerging-entity scores measure agentic retrieval-plus-reading as much as completion, since the gold triples come from Wikidata and the retrievers draw on the same underlying sources. A sharper test would withhold answers — build triples that require synthesizing facts no single retrieved page states — and see whether accuracy survives.
  • Editorial extension: the two metrics describe different skills. Per the paper's Table 10, relation-aware Hits@$N$ evaluates only 9 of the 3,664 emerging triples at Hits@1 and 483 at Hits@10, so the 88.9–100% relation-aware results and the full-set 45.2% Hits@1 cannot be compared directly; jointly they suggest the agent's edge is comprehensive recall of valid answers, not pinpoint top-1 accuracy.
  • Editorial extension: the Mistral-Nemo ablation, where removing self-reflection improved the smaller model, points to an untested adaptive design — gate reflection on a confidence estimate so weak backbones skip it and strong backbones keep it.
  • Editorial extension: the same loop should transfer to neighbouring open-knowledge tasks — news-event graph construction, fact verification, entity profiling — wherever the bottleneck is locating current evidence rather than recalling memorized facts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces AgREE, a training-free LLM agent that iteratively retrieves from Wikipedia/Google, self-reflects on evidence sufficiency, and generates ranked candidate entities for knowledge graph completion (KGC). The authors report strong results on Wikidata5M and FB60K-NYT10, and construct a new 'emerging entities' benchmark from Wikidata entities with inception dates between January and May 2025, claiming up to 45.3% Hits@N improvement over ReAct and superiority over fully trained KGC baselines. They also propose a relation-aware Hits@N metric intended to correct for one-to-many relation bias. The core claims in the abstract—zero-training outperformance, especially on emerging entities, and a new fair evaluation methodology—are the load-bearing contributions. After review, I find that the evaluation does not yet support these claims: the emerging-entity comparison is confounded by an asymmetric answer-generation protocol, the benchmark is partially circular because the retrieval sources are the same as the Wikidata source, and the relation-aware metric as defined is inconsistent with the reported support sizes.

Significance. If the claims were established, the contribution would be significant: a zero-training agent that outperforms fully trained KGC models on standard benchmarks and on post-training entities would address a real bottleneck in maintaining up-to-date knowledge graphs. The proposed emerging-entity benchmark and the relation-aware evaluation idea also address genuine gaps in KGC evaluation. Strengths of the work include a clearly described agent architecture, detailed qualitative trajectories, an explicit construction procedure for the new benchmark, and a commitment to release code and benchmark data. The weaknesses, however, are concentrated in the experimental protocol: the comparison with ReAct is not apples-to-apples, the emerging-entity evaluation may largely measure retrieval/reading comprehension rather than KG reasoning, and the relation-aware metric contains an internal inconsistency that invalidates the reported relation-aware numbers as stated. These issues are load-bearing for the abstract's central claims, so the paper requires major revision before its contributions can be assessed.

major comments (4)
  1. [§'Performance on Emerging Entities' (Table 5) and supplementary 'Answer Generation Prompt'] The headline emerging-entity comparison is not apples-to-apples. AgREE's prompt explicitly requires 'at least 10 possible answers, ranked from most to least confident' and instructs the model to include multiple granularity levels, while ReAct is described as generating one answer. Consequently ReAct's Hits@5 and Hits@10 are equal to its Hits@1 by construction (all 30.2% in Table 5), whereas AgREE's Hits@10 is inflated by the forced 10-candidate output. The abstract's claim of '45.3% hits@N improvements' is therefore an artifact of the output protocol, not a measured capability difference. Please re-run the comparison under identical output constraints (e.g., same 10-answer prompt for ReAct, or a single-answer protocol for both) and report Hits@1 separately for the multi-answer setting.
  2. [§'Emerging Entities: A New Benchmark Dataset' and Table 4] The emerging-entities evaluation is partially circular with respect to the retrieval sources. The benchmark is constructed from Wikidata by filtering entities created in Jan–May 2025, and the retrieval tools are the Wikipedia API and Google search API—the same sources from which Wikidata facts are ultimately derived. The qualitative trajectory in Table 4 (Example 1) shows the agent retrieving an exact cast list for 'Us' from the web, which is close to copying the ground truth. Thus, high Hits@N on this benchmark may reflect the agent's ability to locate and read an already-existing factual statement, rather than knowledge graph completion or reasoning about novel entities. To support the paper's claim of 'constructing knowledge graph triplets' for unseen entities, please either (a) compare against a retrieval-augmented baseline that is given the same tool access and output protocol, or (
  3. [§'Relation-aware Hits@N' (Eq. 4) and Table 10] There is an internal inconsistency between the definition of relation-aware Hits@N and the reported support. Eq. (4) defines T_valid = {(h,r,t) ∈ T : N ≤ N_rel}. Since N_rel is the maximum number of tails for any head in the training set, N_rel ≥ 1 for any relation that occurs in training, so for N=1 the condition should include all test triples with a defined N_rel. Yet Table 10 reports that the Emerging-Entities test set supports only 9 triples for Hits@1. This implies the implementation used the opposite inequality (N_rel ≤ N), or else the metric is defined over a different subset than stated. The results in Table 6 are therefore not interpretable: for the emerging set, Hits@1=88.9% and Hits@3=100% correspond to 8/9 and 12/12 triples respectively, and even the Hits@10 number is based on only 483 triples. Please correct the definition, report the actual filtering rule, and provide conf
  4. [§'Knowledge Graph Completion Performance' (Tables 1–2) and §'Experimental Setup'] The standard-benchmark comparison is reported as single runs with no error bars, no significance tests, and no control for the output format. AgREE returns a free-text list of up to 10 candidate strings that are later entity-linked, whereas the KGE/PLM baselines rank over the full entity vocabulary. The difference in candidate generation makes Hits@N and MRR not directly comparable unless the protocol is aligned (e.g., by mapping AgREE's ranked list to a rank over all entities, or by evaluating all methods with the same number of allowed candidates). The 'up to 13.7%' improvement in Table 2 also needs variance/statistical significance before it can support the abstract's claim. Please report multiple seeds (for non-deterministic components), significance tests, and a description of how candidate ranks are computed for free-text outputs.
minor comments (5)
  1. [§'Relation-Aware Hits Evaluation Results'] The text says 'in Table 3, almost all hits metrics range from 80% to 90%', but Table 3 contains qualitative examples, not results. This appears to be a reference to Table 6; please correct the cross-reference.
  2. [Ablations and Table 2] The model name is spelled inconsistently as 'Mistra-Nemo' in the text and 'Mistral-Nemo' in the table. Please unify.
  3. [Related work / References] The rStar citation appears to be incorrect: the reference 'Ma, L.; Su, Z.; Pan, Y.; Zhang, L.; and Liu, T. 2004. Rstar: An RDF storage and query system...' is a 2004 RDF storage paper, not the MCTS-based retrieval-augmented reasoning work discussed in the text. Please verify and correct the citation.
  4. [Supplementary, 'Relation-Aware Evaluations'] The sentence 'reported in Table for relation-aware hits@N' is incomplete; the table number is missing.
  5. [Reproducibility] The abstract and supplementary indicate code and benchmark 'will be released upon acceptance'; please state clearly whether the benchmark construction code and the exact retrieval snapshots will be released, since the evaluation depends on time-varying web content.

Circularity Check

3 steps flagged · score 6.0 of 10

Emerging-entity claim reduces to construction: benchmark triples come from Wikidata while the agent reads Wikipedia/Google — the same sources Wikidata is compiled from; Hits@N gains are inflated by a ≥10-answer prompt vs. ReAct's single answer; relation-aware numbers rest on a metric contradiction (9/12 supports, opposite inequality).

  1. self definitional [Experiments — "Emerging Entities: A New Benchmark Dataset" and "Experimental Setup" (tools)]
    "We systematically select entities from Wikidata whose inception date falls between January 1, 2025 and May 1, 2025, which is well beyond DeepSeek-V3's training cutoff of July 2024. ... For each selected entity, we retrieve all associated triples via the Wikidata API. ... the basic retriever powered by the Wikipedia API and the advanced retriever powered by the Google search API."

    The Emerging-Entities benchmark's ground truth is the set of triples returned by the Wikidata API, and AgREE's two tools retrieve from Wikipedia and Google — the same sources from which Wikidata is compiled. For a query (h, r, ?), the retrieved documents about h can therefore contain the ground-truth triple verbatim; Table 4 shows this happening (the Wikipedia snippet 'Love Scout (Korean: ...) is a 2025 South Korea ...' is echoed as the predicted genre 'romance drama'). The correct answer is read off the input rather than completed from independent evidence, so temporal filtering excludes only the LLM's parametric knowledge — it does not prevent the answer key from being present in the retrieval corpus by construction. The claimed 'rigorous test of generalization' and the 45.3% emerging-en

  2. other [Experiments — "Performance on Emerging Entities" (Table 5); Supplementary — "Answer Generation Prompt"]
    "ReAct shows performance at 30.2%, which is flat as it generates one answer. ... ANSWER REQUIREMENTS: - Provide at least 10 possible answers, ranked from most to least confident"

    The main Hits@N comparisons pair AgREE, whose prompt mandates 'at least 10 possible answers, ranked from most to least confident', against ReAct, which the paper itself states 'generates one answer' (hence its flat 30.2% across Hits@1/5/10). For any model, top-10 recall over a forced 10-candidate list is trivially greater than or equal to its top-1 recall; a single-answer baseline has no such pool, so its Hits@10 collapses to Hits@1. AgREE's Hits@10 of 75.5% versus ReAct's 30.2% — the basis of the 'up to 45.3%' improvement — is thus manufactured by the output protocol rather than by superior completion. The same asymmetry inflates the Wikidata5m and FB60K-NYT10 tables, where AgREE always emits ≥10 answers while trained seq2seq baselines emit ranked single or few candidates.

1 more flagged steps
  1. other [Experiments — "Relation-Aware Hits@N" (Equation 4); Supplementary Table 10]
    "Tvalid = {(h, r, t) ∈ T : N ≤ Nrel} (Equation 4) ... Table 10: Emerging-Entities 3664 9 12 483 ... Numbers shown are the number of data points evaluated for relation-aware hits@N in the test set."

    Under Equation (4)'s validity condition N ≤ Nrel, every emerging-entity triple qualifies at N=1 because Nrel ≥ 1 for any relation present in the Wikidata5m training set; so Hits@1-RA support should be 3,664, not 9. The reported supports (9, 12, 483) increase with N, which matches the opposite inequality Nrel ≤ N. The implemented evaluation therefore inverts the paper's own metric definition. The headline relation-aware gains on emerging entities — 88.9% Hits@1 and 100% Hits@3 (Table 6) — are merely 8/9 and 12/12 correct predictions on an ultra-restricted subset whose selection rule contradicts Equation (4). The 'even greater performance gaps' are an artifact of this redefined, minuscule evaluation set, so the proposed methodology's flagship numbers do not follow from the metric as written.

full rationale

Three load-bearing reductions by construction. (1) The central emerging-entity claim is self-definitional: the benchmark is built from Wikidata while the agent reads Wikipedia/Google — the sources Wikidata is compiled from — so correct answers are retrievable verbatim from the input, and 'completion' reduces to copying the answer key's provenance. Table 4's own trajectory exhibits this (the retrieved Love Scout snippet states the genre that is then output). (2) The Hits@N comparison is asymmetric by construction: AgREE's prompt forces ≥10 ranked answers, ReAct emits one, so ReAct's Hits@5/Hits@10 trivially equal its Hits@1 (flat 30.2%), and AgREE's Hits@10 is counted over a prompt-guaranteed 10-candidate pool. The 'up to 45.3%' headline improvement is therefore largely an output-format artifact. (3) The relation-aware evaluation contradicts its own definition: Equation (4)'s N ≤ Nrel would retain all 3,664 triples at N=1, yet Table 10 reports support of 9, 12, 483 — matching only the opposite inequality Nrel ≤ N; the 88.9%/100% figures are 8/9 and 12/12. There is no load-bearing self-citation chain (the reference list contains no works by the present authors), and independent content does exist: Wikidata5m and FB60K-NYT10 are external benchmarks, ablations isolate component contributions, and the agent still misses 55% of emerging-entity Hits@1. But the paper's own admissions — ReAct 'generates one answer', the 9/12/483 support table, and the copy-from-snippet trajectories — corroborate that the flagship emerging-entity and relation-aware results reduce largely by construction, warranting a partial-circularity score of 6.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-chosen prompt and benchmark parameters (10-answer requirement, max iterations, cardinality filters, relation-aware threshold), on the assumption that live web retrieval is comparable to parametric KGC reasoning, and on the assumption that the post-cutoff benchmark contains no leakage. The ledger makes these explicit.

free parameters (5)
  • required answer count = at least 10
    Answer Generation Prompt requires at least 10 ranked answers with granularity variants; this directly inflates Hits@N and biases comparison against single-answer ReAct.
  • max iterations = 20
    The agent loop stops after 20 iterations, with an average of 2.5 on Wikidata5m. Hand-chosen cap that affects retrieval sufficiency and cost.
  • relation-aware threshold Nrel = max |{t : (h,r,t) in G_train}|
    Defined in Eq. 3; filters the test set to Tvalid, leaving only 9, 12, or 483 emerging-entity triples at different N. This filter drives the reported relation-aware gains.
  • benchmark inception window = 2025-01-01 to 2025-05-01
    Defines the emerging-entities dataset and assumes no leakage into DeepSeek-V3, whose cutoff is July 2024.
  • cardinality filter = exclude (h,r) with more than 10 ground-truth tails
    Benchmark construction removes high-cardinality queries, which changes task difficulty and comparability with standard KGC benchmarks.
assumptions (4)
  • domain assumption Retrieved documents from Wikipedia and Google contain the relevant ground-truth facts for benchmark triples.
    The method depends on external retrieval; if the facts are absent from the web, the agent cannot complete them. Retriever statistics show heavy tool use on emerging entities.
  • domain assumption Wikidata inception date between January and May 2025 guarantees the entity was unseen by DeepSeek-V3 (cutoff July 2024).
    Temporal filtering in dataset construction assumes no leakage through related entities, descriptions, or training data.
  • domain assumption Relation cardinality in the training set is a valid basis for deciding which test triples are meaningful to evaluate.
    Relation-aware Hits@N uses Nrel from the training KG and discards most test triples, which may select easier cases.
  • domain assumption The LLM can reliably judge information sufficiency during self-reflection.
    Self-reflection step in Algorithm 1; the paper's own error analysis shows over-confidence and premature stopping, so this is load-bearing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AgREE: Agentic Reasoning for Knowledge Graph Completion on Emerging Entities." pith.science (2026). https://pith.science/paper/GI6PZPSU

@misc{pith2026250804118,
  author       = {Pith},
  title        = {Pith review of: AgREE: Agentic Reasoning for Knowledge Graph Completion on Emerging Entities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GI6PZPSU}},
  note         = {Machine review of arXiv:2508.04118}
}
read the original abstract

Open-domain Knowledge Graph Completion (KGC) faces significant challenges in an ever-changing world, especially when considering the continual emergence of new entities in daily news. Existing approaches for KGC mainly rely on pretrained language models' parametric knowledge, pre-constructed queries, or single-step retrieval, typically requiring substantial supervision and training data. Even so, they often fail to capture comprehensive and up-to-date information about unpopular and/or emerging entities. To this end, we introduce Agentic Reasoning for Emerging Entities (AgREE), a novel agent-based framework that combines iterative retrieval actions and multi-step reasoning to dynamically construct rich knowledge graph triplets. Experiments show that, despite requiring zero training efforts, AgREE significantly outperforms existing methods in constructing knowledge graph triplets, especially for emerging entities that were not seen during language models' training processes, outperforming previous methods by up to 13.7%. Moreover, we propose a new evaluation methodology that addresses a fundamental weakness of existing setups and a new benchmark for KGC on emerging entities. Our work demonstrates the effectiveness of combining agent-based reasoning with strategic information retrieval for maintaining up-to-date knowledge graphs in dynamic information environments.

Figures

Figures reproduced from arXiv: 2508.04118 by the authors.

Figure 1
Figure 1. An illustration of the AgREE methodology. existing KG, relevant triples are also shown to the agent to provide context. Secondly, the agent determines whether to utilize tools (go to step 2) or to directly generate the answer (go to step 4) if it is already confident regarding its internal knowledge about the entity. This preliminary step prevents unnecessary retrieval operations for head entities while en￾suring co… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 23 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bala z evi \'c , I.; Allen, C.; and Hospedales, T. M. 2019. Tucker: Tensor factorization for knowledge graph completion. arXiv preprint arXiv:1901.09590

    arXiv 2019

  4. [4]

    Bordes, A.; Usunier, N.; Garcia-Duran, A.; Weston, J.; and Yakhnenko, O. 2013. Translating embeddings for modeling multi-relational data. Advances in neural information processing systems, 26

    work page 2013

  5. [5]

    Das, R.; Dhuliawala, S.; Zaheer, M.; Vilnis, L.; Durugkar, I.; Krishnamurthy, A.; Smola, A.; and McCallum, A. 2017. Go for a walk and arrive at the answer: Reasoning over paths in knowledge bases using reinforcement learning. arXiv preprint arXiv:1711.05851

    arXiv 2017

  6. [6]

    Derczynski, L.; Nichols, E.; van Erp, M.; and Limsopatham, N. 2017. Results of the WNUT 2017 Shared Task on Novel and Emerging Entity Recognition. In Derczynski, L.; Xu, W.; Ritter, A.; and Baldwin, T., eds., Proceedings of the 3rd Workshop on Noisy User-generated Text, 140--147. Copenhagen, Denmark: Association for Computational Linguistics

    work page 2017

  7. [7]

    Dettmers, T.; Minervini, P.; Stenetorp, P.; and Riedel, S. 2018. Convolutional 2D Knowledge Graph Embeddings. Proceedings of the AAAI Conference on Artificial Intelligence, 32(1)

    work page 2018

  8. [8]

    Fu, C.; Chen, T.; Qu, M.; Jin, W.; and Ren, X. 2019 a . Collaborative policy learning for open knowledge graph reasoning. arXiv preprint arXiv:1909.00230

    work page Pith review arXiv 2019

Show all 31 references
  1. [9]

    Fu, C.; Chen, T.; Qu, M.; Jin, W.; and Ren, X. 2019 b . Collaborative Policy Learning for Open Knowledge Graph Reasoning. In Inui, K.; Jiang, J.; Ng, V.; and Wan, X., eds., Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inter...

    2019

  2. [10]

    Han, X.; Liu, Z.; and Sun, M. 2018. Neural knowledge acquisition via mutual attention between knowledge graph and text. In Proceedings of the AAAI conference on artificial intelligence, volume 32

    2018

  3. [11]

    Hoffart, J.; Altun, Y.; and Weikum, G. 2014. Discovering emerging entities with ambiguous names. In Proceedings of the 23rd International Conference on World Wide Web, WWW '14, 385–396. New York, NY, USA: Association for Computing Machinery. ISBN 9781450327442

    2014

  4. [12]

    D.; Gutierrez, C.; Kirrane, S.; Gayo, J

    Hogan, A.; Blomqvist, E.; Cochez, M.; d’Amato, C.; Melo, G. D.; Gutierrez, C.; Kirrane, S.; Gayo, J. E. L.; Navigli, R.; Neumaier, S.; et al. 2021. Knowledge graphs. ACM Computing Surveys (Csur), 54(4): 1--37

    2021

  5. [13]

    Ji, S.; Pan, S.; Cambria, E.; Marttinen, P.; and Yu, P. S. 2021. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE transactions on neural networks and learning systems, 33(2): 494--514

    2021

  6. [14]

    Jiang, P.; Agarwal, S.; Jin, B.; Wang, X.; Sun, J.; and Han, J. 2023. Text Augmented Open Knowledge Graph Completion via Pre-Trained Language Models. In Rogers, A.; Boyd-Graber, J.; and Okazaki, N., eds., Findings of the Association for Computational Linguistics: ACL 2023, 111...

    2023

  7. [15]

    Kochsiek, A.; Saxena, A.; Nair, I.; and Gemulla, R. 2023 a . Friendly Neighbors: Contextualized Sequence-to-Sequence Link Prediction. In Can, B.; Mozes, M.; Cahyawijaya, S.; Saphra, N.; Kassner, N.; Ravfogel, S.; Ravichander, A.; Zhao, C.; Augenstein, I.; Rogers, A.; Cho, K.; ...

    2023

  8. [16]

    Kochsiek, A.; Saxena, A.; Nair, I.; and Gemulla, R. 2023 b . Friendly Neighbors: Contextualized Sequence-to-Sequence Link Prediction. In Can, B.; Mozes, M.; Cahyawijaya, S.; Saphra, N.; Kassner, N.; Ravfogel, S.; Ravichander, A.; Zhao, C.; Augenstein, I.; Rogers, A.; Cho, K.; ...

    2023

  9. [17]

    Kolomiyets, O.; and Moens, M.-F. 2011. A survey on question answering technology from an information retrieval perspective. Information Sciences, 181(24): 5412--5434

    2011

  10. [18]

    Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    2024 arXiv

  11. [19]

    Lv, X.; Lin, Y.; Cao, Y.; Hou, L.; Li, J.; Liu, Z.; Li, P.; and Zhou, J. 2022 a . Do Pre-trained Models Benefit Knowledge Graph Completion? A Reliable Evaluation and a Reasonable Approach. In Muresan, S.; Nakov, P.; and Villavicencio, A., eds., Findings of the Association for ...

    2022

  12. [20]

    Lv, X.; Lin, Y.; Cao, Y.; Hou, L.; Li, J.; Liu, Z.; Li, P.; and Zhou, J. 2022 b . Do pre-trained models benefit knowledge graph completion? a reliable evaluation and a reasonable approach. Association for Computational Linguistics

    2022

  13. [21]

    Ma, L.; Su, Z.; Pan, Y.; Zhang, L.; and Liu, T. 2004. Rstar: An rdf storage and query system for enterprise resource management. In Proceedings of the thirteenth ACM international conference on Information and knowledge management, 484--491

    2004

  14. [22]

    Saxena, A.; Kochsiek, A.; and Gemulla, R. 2022. Sequence-to-Sequence Knowledge Graph Completion and Question Answering. In Muresan, S.; Nakov, P.; and Villavicencio, A., eds., Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Lo...

    2022

  15. [23]

    Schick, T.; Dwivedi-Yu, J.; Dessì, R.; Raileanu, R.; Lomeli, M.; Zettlemoyer, L.; Cancedda, N.; and Scialom, T. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. (arXiv:2302.04761). ArXiv:2302.04761 [cs]

    2023 arXiv

  16. [24]

    Tran, H.; Yao, Z.; Wang, J.; Zhang, Y.; Yang, Z.; and Yu, H. 2024. Rare: Retrieval-augmented reasoning enhancement for large language models. arXiv preprint arXiv:2412.02830

    2024 arXiv

  17. [25]

    Wang, L.; Zhao, W.; Wei, Z.; and Liu, J. 2022. Simkgc: Simple contrastive knowledge graph completion with pre-trained language models. arXiv preprint arXiv:2203.02167

    2022 arXiv

  18. [26]

    Wang, X.; Gao, T.; Zhu, Z.; Zhang, Z.; Liu, Z.; Li, J.; and Tang, J. 2021. KEPLER : A Unified Model for Knowledge Embedding and Pre-trained Language Representation. Transactions of the Association for Computational Linguistics, 9: 176--194

    2021

  19. [27]

    Wang, X.; He, X.; Cao, Y.; Liu, M.; and Chua, T.-S. 2019. Kgat: Knowledge graph attention network for recommendation. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 950--958

    2019

  20. [28]

    Xiong, C.; Power, R.; and Callan, J. 2017. Explicit semantic ranking for academic search via knowledge graph embedding. In Proceedings of the 26th international conference on world wide web, 1271--1279

    2017

  21. [29]

    Xu, C.; Bai, Y.; Bian, J.; Gao, B.; Wang, G.; Liu, X.; and Liu, T.-Y. 2014. Rc-net: A general framework for incorporating knowledge into word representations. In Proceedings of the 23rd ACM international conference on conference on information and knowledge management, 1219--1228

    2014

  22. [30]

    Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K.; and Cao, Y. 2022. ReAct: Synergizing Reasoning and Acting in Language Models. (arXiv:2210.03629). ArXiv:2210.03629 [cs]

    2022 arXiv

  23. [31]

    Zhou, T.; Chen, Y.; Liu, K.; and Zhao, J. 2024. C og MG : Collaborative Augmentation Between Large Language Model and Knowledge Graph. In Cao, Y.; Feng, Y.; and Xiong, D., eds., Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: ...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.