AutoSchemaKG: Autonomous Knowledge Graph Construction through Dynamic Schema Induction from Web-Scale Corpora

Chunyang Li; Feng Qin; Gong Zhang; Haoyu Huang; Hong Ting Tsang; Hongyu Luo; Huiwen Yang; Jiaxin Bai; Leijie Wu; Qi Hu

REVIEW 5 major objections 5 minor 3 cited by

AutoSchemaKG: Autonomous Knowledge Graph Construction through Dynamic Schema Induction from Web-Scale Corpora

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A fully automatic pipeline induces schemas from raw text and builds billion-scale knowledge graphs whose retrieval improves multi-hop QA and LLM factuality.

desk verdict A substantial billion-scale KG pipeline that deserves a serious referee, but the headline schema-alignment and QA claims are not supported by the data as reported. read the letter →

arxiv 2505.23628 v3 pith:JTF7WEG2 submitted 2025-05-29 cs.CL cs.AI

Jiaxin Bai , Wei Fan , Qi Hu , Qing Zong , Chunyang Li , Hong Ting Tsang , Hongyu Luo , Yauwai Yim

show 12 more authors

Haoyu Huang Xiao Zhou Feng Qin Tianshi Zheng Xi Peng Xin Yao Huiwen Yang Leijie Wu Yi Ji Gong Zhang Renhai Chen Yangqiu Song

This is my paper · ORCID

classification cs.CLcs.AI

keywords knowledgegraphconstructionschemainductionconceptualizationretrieval-augmentedgenerationlargelanguagemodelsmulti-hopquestionansweringfactualityevent-centric

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

This paper claims that knowledge graphs large enough to complement a language model's parametric memory can be built automatically, with no predefined schema, by having an LLM both extract triples and induce a conceptual hierarchy from raw text. The authors construct ATLAS, a family of billion-scale graphs (over 900 million nodes and 5.9 billion edges) from Wikipedia, Semantic Scholar abstracts, and Common Crawl, and report that the induced schemas reach 92% semantic alignment with human-crafted ontologies. They argue that the resulting event- and concept-rich graphs preserve more passage information than entity-only triples and, when used as a retrieval corpus, improve multi-hop question answering by 12-18% and LLM factuality by up to 9% over text-retrieval baselines. If correct, this would make schema-free, web-scale knowledge graph construction practical and would demonstrate that structured knowledge can complement parametric knowledge even when both derive from the same source corpus.

What carries the argument

The machinery is a three-stage LLM triple-extraction pipeline followed by LLM-driven conceptualization. Stage prompts $P_{EE}$, $P_{EV}$, and $P_{VV}$ respectively produce entity-entity, entity-event, and event-event triples from 1024-token chunks; then schema-induction prompts ask the LLM for at least three abstract phrases per element, with entities augmented by a context string sampled from up to $N_{ctx}$ graph neighbors. The resulting concept set $C$, with mappings $\varphi:V \to \mathcal{P}(C)$ and $\psi:R \to \mathcal{P}(C)$, turns a flat triple store into a hierarchical graph. Event nodes and concept nodes carry the argument: events preserve over 90% of passage content in MCQ tests (vs 70% for entities alone), and concept nodes provide alternative retrieval pathways that bridge disconnected entity subgraphs during PageRank-based retrieval.

What would settle it

Human annotators on a random sample of several hundred passages, comparing their agreed triples and schema types against the LLM output, would settle whether the reported 90%+ F1 and 92% alignment hold under human judgment rather than model self-assessment.

Watch

Extended reading notes

Core claim

The central claim is that the traditional prerequisite of expert-defined schemas can be removed entirely: a single LLM, prompted three ways, extracts entity-entity, entity-event, and event-event triples from text, and a subsequent conceptualization step generalizes each entity (using neighbor context), event, and relation into multi-level abstract phrases. These phrases form an induced schema that maps every node and relation to a non-empty set of concepts, yielding a knowledge graph $G = (V, E, C, \varphi, \psi)$. Applied to Dolma 1.7 subsets, this produces the ATLAS family, and the paper reports triple-extraction F1 above 90%, schema-typing recall above 80% (often 90%) measured with BERTScore, and downstream gains on MuSiQue, 2WikiMultiHopQA, HotpotQA, FELM, and knowledge-intensive MMLU subjects. The authors conclude that billion-scale dynamically-schematized knowledge graphs can effectively complement parametric knowledge in large language models.

Load-bearing premise

The claim that the pipeline produces near-human-quality triples and schemas rests on using DeepSeek-V3 as an automated judge for triple precision and recall and BERTScore for schema alignment, with no human-annotated validation sample in the paper.

Editorial extensions

If this is right

Automatic, schema-free construction can reach web scale: ATLAS contains over 900 million nodes and 5.9 billion edges from more than 50 million documents.
Event-level triples preserve more passage information than entity-only triples (over 90% vs 70% MCQ accuracy), so event-centric graphs are stronger retrieval sources.
Concept nodes created by schema induction supply alternative paths that improve PageRank-based retrieval and multi-hop question answering.
Graph-based retrieval over a billion-scale KG can beat text retrieval by 12-18% on multi-hop QA even when the corpus overlaps LLM pretraining data.
Retrieval from the induced-schema KG improves LLM factuality on FELM and knowledge-intensive MMLU subjects without domain-specific tuning.

Reading between the lines

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

If the 92% alignment is confirmed with human evaluation, auto-induced schemas could serve as living ontologies that track corpus drift, something static expert schemas cannot do.
The BERTScore-based alignment figure likely rewards lexical and semantic paraphrase rather than exact taxonomic structure, so the true structural match with human ontologies may be lower than 92%.
A testable extension is to measure whether the induced schema transfers to new domains: run the pipeline on a held-out corpus and check whether downstream QA gains persist without re-inducing the schema.
The same pipeline could be applied to non-English or domain-specific corpora (e.g., biomedical texts) to see whether conceptualization degrades outside the LLM's strongest language and data regime.

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.

Desk Editor's Note

read the letter

Here's my take. This is a big, serious engineering effort—processing 50M docs into a billion-scale KG with entity, event, and concept layers is not trivial, and the pipeline is coherent. The ATLAS family could be a genuine resource if the authors actually release it. But the headline numbers don't hold up.

The 92% schema-alignment claim is the weakest link. The full-text abstract says 95%; the short abstract says 92. The metric, BS-R, takes the max BERTScore over gold types for each induced schema (Eq. 4), which rewards broad, generic concepts and can stay high even when the induced schema is shallow. No single row in Table 4 gives 92% — the 70B average is 93.05, the 8B average 89.29. And there is no human-annotated evaluation of the induced schemas, nor any comparison of the ATLAS schema layer to a human-crafted ontology. That's the load-bearing claim, and it's currently unverifiable.

Triple extraction has a similar issue: precision/recall are judged by DeepSeek-V3, an LLM judge, with no human sample. Useful as a sanity check, but not the same as ground truth.

The QA results are more modest than the abstract implies. Against the strongest baseline, HippoRAG2, the Full-KG is worse on MuSiQue (31.8 vs 37.2 EM) and HotpotQA (61.8 vs 62.7), and only slightly better on 2Wiki (65.3 vs 65.0), within noise since there are no error bars. The claimed 12–18% gains are relative to BM25 and Contriever, which is a much lower bar.

There's also an internal scale inconsistency: the abstract's 900M nodes/5.9B edges describes ATLAS-CC alone, while the family totals in Table 1 sum to about 1.36B nodes and 8.6B edges. That makes me worry about reporting discipline.

What's genuinely good? Event-centric modeling plus conceptualization is a reasonable direction, and scaling it to this size is an achievement. The limitations section is honest about LLM biases and resource costs. The paper engages the prior literature, including the group's own CAT/CANDLE/ASER line, which is relevant here.

I'd send it to a serious referee. A billion-scale automatically constructed KG with events and concepts is new enough, and the resource, if released, would be valuable to many. But as written, the key claims are not supported. A major revision should reconcile the counts, add human validation on samples of triples and schemas, report error bars and significance tests, and release at least a subset of the ATLAS artifacts. Then it could be a solid resource paper. Right now it's a promising but over-claimed engineering report.

Referee Report

5 major / 5 minor

Summary. The paper presents AutoSchemaKG, an LLM-based pipeline that constructs knowledge graphs without predefined schemas by extracting entity-entity, entity-event, and event-event triples and then inducing a conceptual schema (entity, event, and relation types) directly from text. The pipeline is applied to the Dolma corpus to build the ATLAS family of knowledge graphs (ATLAS-Wiki, ATLAS-Pes2o, ATLAS-CC), reported as containing 900+ million nodes and 5.9 billion edges. The authors evaluate triple extraction quality with an LLM judge, schema quality with BERTScore-based recall/coverage on entity, event, and relation typing benchmarks, information preservation via LLM-generated multiple-choice questions, and downstream performance on multi-hop QA, FELM factuality, and MMLU. The central claims are that schema induction achieves 92% semantic alignment with human-crafted schemas without manual intervention, that the ATLAS graphs outperform state-of-the-art retrieval baselines on multi-hop QA, and that they improve LLM factuality on knowledge-intensive domains.

Significance. If the central claims hold, this is a significant engineering and empirical contribution: a fully automatic, web-scale KG construction pipeline with an induced schema layer, explicit modeling of events and concepts, transparent GPU-hour reporting, and evaluations on multiple external benchmarks. The formal graph definition, detailed prompts, and included algorithms support reproducibility. The strongest evidence in the paper is the breadth of downstream evaluation (multi-hop QA, FELM, MMLU) and the scale of the constructed graphs. However, the paper's headline quantitative claims are not currently backed by the tables and text as written, so the significance cannot be assessed without substantial revision.

major comments (5)

[§5.1, Table 4, Appendix C.2, Eq. (4)] The headline claim that schema induction achieves 92% semantic alignment with human-crafted schemas is not supported by the reported evidence. Table 4 reports BS-R and BS-C on four benchmark typing tasks (FB15kET entity typing, YAGO43kET entity typing, wikiHow event typing, FB15kET relation typing); no row or average equals 92%, and these benchmarks do not evaluate the ATLAS induced schema against a human-crafted ontology. Moreover, BS-R in Eq. (4) takes the maximum BERTScore over all gold types for each induced schema, which rewards over-broad generic concepts and can remain high for shallow schemas. Please report the exact computation of the 92% figure, add a human-annotated evaluation of the induced ATLAS concepts, or substantially weaken the claim.
[Abstract and §1] The full-text abstract reports 95% semantic alignment with human-crafted schemas, while the arXiv abstract and the introduction report 92%. This numerical inconsistency must be resolved, and the final version should state a single, reproducible figure.
[Abstract and Table 1] The claim of '900+ million nodes and 5.9 billion edges' matches ATLAS-CC alone, not the ATLAS family. Summing Table 1 gives approximately 1.36 billion nodes (243.912M + 174.387M + 937.256M) and 8.60 billion edges (1.492B + 1.150B + 5.958B). If the headline refers to the family of graphs, the numbers need correction; if it refers to ATLAS-CC, the text should say so explicitly.
[§5.2, Table 5] The claim that AutoSchemaKG 'outperforms state-of-the-art baselines by 12-18%' is not supported with respect to the strongest graph baseline. On MuSiQue, HippoRAG2 achieves EM 37.2 versus AutoSchemaKG+HippoRAG2 Full-KG EM 31.8; on HotpotQA, HippoRAG2 achieves EM 62.7 versus 61.8. The 12-18% gain appears to be relative to BM25/Contriever, not to state-of-the-art graph RAG. Please restate the comparison precisely and avoid overclaiming the multi-hop results.
[§5.1 and Appendix C] No human-annotated ground truth is used for either triple extraction (DeepSeek-V3 is used as a judge) or schema quality (BERTScore-based BS-R/BS-C is used). The 90%+ triple F1 and the 92% alignment claim both depend on these proxies. Because the load-bearing assumption is that LLM-based and embedding-based metrics faithfully approximate human judgment, the paper would be substantially strengthened by a small human-annotated validation sample, e.g., 100 documents for triple extraction and a few hundred entities/events for schema typing.

minor comments (5)

[§5.1] The sentence 'we show that the Au-toSchemaKG has accurate triplex extraction' contains a typo: 'triplex' should be 'triple'.
[§5.2] The text refers to 'experimental results in Figure 5 and Figure 8'; the relevant results are in Table 5 and Table 8. Please correct the cross-references.
[Table 4] The row label 'Y AGO43kET' has an unintended space; it should read 'YAGO43kET'.
[Appendix G.1, Eq. (6)] The balanced accuracy formula is typeset without the division signs; as written it reads 'TP T P+ F N+ T N T N+ F P' instead of TP/(TP+FN) + TN/(TN+FP).
[References] The in-text citation 'Khandelwal et al.' lacks a year and does not match the reference list formatting used elsewhere.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are evaluated against external benchmarks and gold-type hierarchies; the 92% alignment figure and LLM-as-judge settings are validity concerns, not by-construction reductions.

full rationale

The derivation chain is not circular. Triple-extraction quality is checked against the source text with DeepSeek-V3 as an external judge; schema quality is measured with BERTScore against external gold type hierarchies (FB15kET, YAGO43kET, wikiHow) using Eq. 4-5, not against the induced ATLAS schema itself; QA, factuality, and knowledge-intensive reasoning are tested on MuSiQue/HotpotQA/2Wiki/FELM/MMLU against independent baselines. No parameter is fitted to the reported outcome and then renamed a prediction, and no uniqueness theorem or load-bearing result is imported solely from the authors' prior conceptualization papers—those citations motivate the method but do not establish the empirical numbers. The MCQ information-preservation test does use the same model family to generate and answer questions, and the abstract's 92% vs. the full text's 95% alignment figure is internally inconsistent; however, the lower-bound/no-context condition and the external gold answers prevent these results from being forced by construction. The limitations section also candidly acknowledges inherited LLM biases and weakness in technical domains, which further indicates the claims are empirical rather than definitional. These are correctness and reporting concerns, not circularity in the derivation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claims rest mainly on domain assumptions about the validity of LLM-as-judge and embedding-based evaluations, plus several unspecified generation hyperparameters. There is no formal derivation, so the free-parameter and axiom loads are concentrated in the evaluation protocol rather than in a mathematical core.

free parameters (5)

LLM generation temperature tau and top-p p = not specified
Schema induction and triple extraction use controlled sampling, but the paper never states the values, so the exact outputs are not reproducible from the text.
Context neighborhood size Nctx = e.g., one predecessor and one successor
Entity conceptualization samples neighbors to build a context string; the precise Nctx is left as an example, not a fixed protocol.
Event generation length scaling factor alpha = not specified
Event-event extraction uses an extended generation limit Lext = alpha * Lmax with alpha > 1, but the value is never given.
HippoRAG2 damping factor = 0.9
Set to concentrate PageRank propagation within local subgraphs; chosen heuristically based on graph density.
HippoRAG2 weight adjustment factor = 0.9
Applied in the HippoRAG2 adaptation to balance node and passage scores; value selected by tuning, not by an external principle.

assumptions (5)

domain assumption DeepSeek-V3 judge outputs are treated as ground truth for triple precision and recall.
Section 5.1: the counting-based evaluation uses an LLM judge without human verification; if the judge is biased, the reported F1 scores over 90% are unsupported.
domain assumption BERTScore with RoBERTa embeddings is an adequate measure of semantic alignment with human-crafted schemas.
Section 5.1 and Appendix C.2: BS-R and BS-C use token-embedding recall to compare induced concepts with ground-truth types, which can overestimate agreement through synonymy without verifying the induced type is actually correct.
domain assumption MCQs generated by LLaMA-3-70B from passages are valid probes of information preservation.
Section 5.1 and Figure 8: the lower-upper bound protocol assumes the generated questions test the passage content and that answer accuracy reflects KG information retention.
domain assumption The Dolma 1.7 subsets are representative web-scale corpora and are processed as claimed.
Section 4: ATLAS is built from full Wikipedia, Semantic Scholar abstracts, and 3% slices of Common Crawl; the paper does not audit whether the final graphs contain the claimed content or are internally consistent.
ad hoc to paper Concept nodes are the causal mechanism behind the downstream RAG gains.
Sections 5.2-5.4: the benefit of concept nodes is supported mainly by two case studies in Figures 9 and 10, not by controlled ablations that isolate the contribution of concepts with statistical tests.

invented entities (1)

Concept layer C with mappings phi and psi independent evidence
purpose: Organizes entities, events, and relations into induced semantic categories, intended to enable zero-shot bridging, reduce sparsity, and create alternative retrieval paths.
The typing experiments against FB15kET, YAGO43kET, and wikiHow provide external handles, but the evaluation uses BERTScore similarity rather than exact type matching, and no human inspection of the induced concepts is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AutoSchemaKG: Autonomous Knowledge Graph Construction through Dynamic Schema Induction from Web-Scale Corpora." pith.science (2026). https://pith.science/paper/JTF7WEG2

@misc{pith2026250523628,
  author       = {Pith},
  title        = {Pith review of: AutoSchemaKG: Autonomous Knowledge Graph Construction through Dynamic Schema Induction from Web-Scale Corpora},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JTF7WEG2}},
  note         = {Machine review of arXiv:2505.23628}
}

read the original abstract

We present AutoSchemaKG, a framework for fully autonomous knowledge graph construction that eliminates the need for predefined schemas. Our system leverages large language models to simultaneously extract knowledge triples and induce comprehensive schemas directly from text, modeling both entities and events while employing conceptualization to organize instances into semantic categories. Processing over 50 million documents, we construct ATLAS (Automated Triple Linking And Schema induction), a family of knowledge graphs with 900+ million nodes and 5.9 billion edges. This approach outperforms state-of-the-art baselines on multi-hop QA tasks and enhances LLM factuality. Notably, our schema induction achieves 92\% semantic alignment with human-crafted schemas with zero manual intervention, demonstrating that billion-scale knowledge graphs with dynamically induced schemas can effectively complement parametric knowledge in large language models.

Figures

Figures reproduced from arXiv: 2505.23628 by the authors.

**Figure 3.** The figure demonstrates the prompts we use [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

**Figure 4.** The figure demonstrates the prompts we use [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: This figure shows the prompt used for gener [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

**Figure 7.** Figure 7: , by contrast, rely solely on their textual descriptions without additional context, as their abstraction focuses on inherent semantics rather than graph connectivity. This distinction reflects the differing roles of nodes and edges in the knowledge graph structure. …

**Figure 6.** Figure 6: This figure shows the conceptualization prompts for entities enhanced with context. Events and relations, as shown in [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

**Figure 8.** Figure 8: The prompts for generating and answering [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]

**Figure 9.** Figure 9: Event Node (green) offers enriched context over triplets (blue); dotted line indicates missing edge [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]

**Figure 10.** Figure 10: Concept nodes (orange) provide alternate pathways to access information beyond entities and events. [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

ScalableRAG: High-Quality RAG at Zero Ingestion Cost
cs.AI 2026-07 conditional novelty 6.0 of 10

A set-workspace agentic RAG matches or beats knowledge-graph and schema RAG on six corpora with zero or constant-cost ingestion by doing aggregations over document-aligned sets at inference.
DeCoRAG: Cognitive Decoupling and Semantic-Aware Cropping for Complex Document Understanding
cs.IR 2026-07 conditional novelty 6.0 of 10

Decoupling semantic anchoring from high-res cropping fixes VLM visual attention sinks in document Graph RAG, gaining up to 12.5 SPR points and 40.8% fewer prompt tokens.
Generative Ontology Induction: Domain-Agnostic Schema Discovery from Document Corpora Using Large Language Models
cs.AI 2026-05 reject novelty 4.0 of 10

GOI prompts an LLM to infer a document-class schema, but its headline 'coverage' result mostly measures whether the model echoes the schema it was given.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages · cited by 3 Pith papers

[1]

The ABSTRACT EVENT phrases can well represent the EVENT, and it could be the type of the EVENT or the related concepts of the EVENT

work page
[5]

Sam playing with his dog

Stop immediately if you can’t think of any more phrases, and no explanation is needed. Examples: EVENT: A man retreats to mountains and forests Your answer: retreat, relaxation, escape, nature, soli- tude EVENT: A cat chased a prey into its shelter Your answer: hunting, escape, predation, hidding, stalking EVENT: Sam playing with his dog Your answer: re- ...

work page
[6]

Preprint, arXiv:2009.03300

Measuring massive multitask language under- standing. Preprint, arXiv:2009.03300. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi- hop QA dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th Inter- national Conference on Computational Linguistics, pages 6609–6625, Barcelona, Spain ...

arXiv 2009
[7]

Project Alexandria: Towards Freeing Scientific Knowledge from Copyright Burdens via LLMs

Raptor: Recursive abstractive processing for tree-organized retrieval. In International Conference on Learning Representations (ICLR). Christoph Schuhmann, Gollam Rabby, Ameya Prabhu, Tawsif Ahmed, Andreas Hochlehnert, Huu Nguyen, Nick Akinci Heidrich, Ludwig Schmidt, Robert Kacz- marczyk, Sören Auer, et al. 2025. Project alexandria: Towards freeing scien...

work page Pith review arXiv 2025
[8]

In Findings of the Association for Computational Linguistics: NAACL 2024, pages 51–68, Mexico City, Mexico

DIVKNOWQA: Assessing the reasoning abil- ity of LLMs via open-domain question answering over knowledge base and text. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 51–68, Mexico City, Mexico. Associa- tion for Computational Linguistics. Chang Zong, Yuchen Yan, Weiming Lu, Jian Shao, Yongfeng Huang, Heng Chang, and Yueting Zhuang

work page 2024
[9]

Head": "{a noun}

Triad: A framework leveraging a multi-role LLM-based agent to solve knowledge base question answering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, pages 1698–1710, Miami, Florida, USA. Associ- ation for Computational Linguistics. 13 A Prompts for Triple Extractions Entity Relationship Extraction Given a pas...

work page 2024
[15]

The ABSTRACT ENTITY phrases can well represent the ENTITY , and it could be the type of the ENTITY or the related concepts of the ENTITY

work page
[19]

Stop immediately if you can’t think of any more phrases, and no explanation is needed. Examples: ENTITY: Soul CONTEXT: premiered BFI London Film Festival, became highest-grossing Pixar release Your answer: movie, film ENTITY: Thinkpad X60 CONTEXT: Richard Stall- man announced he is using Trisquel on a Thinkpad X60 Your answer: Thinkpad, laptop, machine, d...

work page

Show all 18 references

[20]

The ABSTRACT RELATION phrases can well represent the RELATION, and it could be the type of the RELATION or the simplest concepts of the RELATION
[21]

Strictly follow the provided format, do not add extra characters or words
[22]

Write at least 3 or more phrases at different abstract level if possible
[23]

Do not repeat the same word and the input in the answer
[24]

question

Stop immediately if you can’t think of any more phrases, and no explanation is needed. Examples: RELATION: participated in Your answer: become part of, attend, take part in, engage in, involve in RELATION: be included in Your answer: join, be a part of, be a member of, be a co...

1993
[2013]

In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013

Translating embeddings for modeling multi- relational data. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Systems 2013. Pro- ceedings of a meeting held December 5-8, 2013, Lake Tahoe, Nevada, United States, pag...

2013
[2021]

In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 10 pages 1084–1099, Online and Punta Cana, Domini- can Republic

ReGen: Reinforcement learning for text and knowledge base generation using pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 10 pages 1084–1099, Online and Punta Cana, Domini- can Republic. Association for Co...

2021 arXiv
[2022]

Nature Communications, 13(1):6736

Systematic tissue annotations of genomics samples by modeling unstructured metadata. Nature Communications, 13(1):6736. Mutian He, Tianqing Fang, Weiqi Wang, and Yangqiu Song. 2024. Acquiring and modeling abstract com- monsense knowledge via conceptualization. Artifi- cial Int...

2024
[2024]

what are you trying to do?

Exploring the potential of ChatGPT on sen- tence level relations: A focus on temporal, causal, and discourse relations. In Findings of the Associ- ation for Computational Linguistics: EACL 2024 , pages 684–721, St. Julian’s, Malta. Association for Computational Linguistics. Ch...

2024 arXiv
[2025]

Great memory, shallow reasoning: Limits of kNN-LMs. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies (Volume 2: Short Papers), pages 471–482, Albuquerque, New Mexico. ...

2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

The ABSTRACT EVENT phrases can well represent the EVENT, and it could be the type of the EVENT or the related concepts of the EVENT

work page

[2] [5]

Sam playing with his dog

Stop immediately if you can’t think of any more phrases, and no explanation is needed. Examples: EVENT: A man retreats to mountains and forests Your answer: retreat, relaxation, escape, nature, soli- tude EVENT: A cat chased a prey into its shelter Your answer: hunting, escape, predation, hidding, stalking EVENT: Sam playing with his dog Your answer: re- ...

work page

[3] [6]

Preprint, arXiv:2009.03300

Measuring massive multitask language under- standing. Preprint, arXiv:2009.03300. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi- hop QA dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th Inter- national Conference on Computational Linguistics, pages 6609–6625, Barcelona, Spain ...

arXiv 2009

[4] [7]

Project Alexandria: Towards Freeing Scientific Knowledge from Copyright Burdens via LLMs

Raptor: Recursive abstractive processing for tree-organized retrieval. In International Conference on Learning Representations (ICLR). Christoph Schuhmann, Gollam Rabby, Ameya Prabhu, Tawsif Ahmed, Andreas Hochlehnert, Huu Nguyen, Nick Akinci Heidrich, Ludwig Schmidt, Robert Kacz- marczyk, Sören Auer, et al. 2025. Project alexandria: Towards freeing scien...

work page Pith review arXiv 2025

[5] [8]

In Findings of the Association for Computational Linguistics: NAACL 2024, pages 51–68, Mexico City, Mexico

DIVKNOWQA: Assessing the reasoning abil- ity of LLMs via open-domain question answering over knowledge base and text. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 51–68, Mexico City, Mexico. Associa- tion for Computational Linguistics. Chang Zong, Yuchen Yan, Weiming Lu, Jian Shao, Yongfeng Huang, Heng Chang, and Yueting Zhuang

work page 2024

[6] [9]

Head": "{a noun}

Triad: A framework leveraging a multi-role LLM-based agent to solve knowledge base question answering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, pages 1698–1710, Miami, Florida, USA. Associ- ation for Computational Linguistics. 13 A Prompts for Triple Extractions Entity Relationship Extraction Given a pas...

work page 2024

[7] [15]

The ABSTRACT ENTITY phrases can well represent the ENTITY , and it could be the type of the ENTITY or the related concepts of the ENTITY

work page

[8] [19]

Stop immediately if you can’t think of any more phrases, and no explanation is needed. Examples: ENTITY: Soul CONTEXT: premiered BFI London Film Festival, became highest-grossing Pixar release Your answer: movie, film ENTITY: Thinkpad X60 CONTEXT: Richard Stall- man announced he is using Trisquel on a Thinkpad X60 Your answer: Thinkpad, laptop, machine, d...

work page