Pith. sign in

REVIEW 4 major objections 5 minor 34 references

FoodTaxo: Generating Food Taxonomies with Large Language Models

T0 review · 4 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 The paper claims that a single open-source large language model, prompted iteratively, can complete existing taxonomies and generate new ones from a bare list of known concepts, with unreliable inner-node placement as the main limitation.

desk verdict Honest, reproducible systems paper whose completion results are solid, but whose generation claims rest on uncalibrated reference-free metrics that the paper itself shows miss obvious errors. read the letter →

arxiv 2505.19838 v1 pith:66ZGQG2K submitted 2025-05-26 cs.CL cs.AI

Pascal Wullschleger , Majid Zarharan , Donnacha Daly , Marc Pouly , Jennifer Foster This is my paper · ORCID
classification cs.CLcs.AI
keywords taxonomygenerationcompletionlargelanguagemodelsfoodtechnologychain-of-thoughtpromptingretrievalaugmentednaturalinferencereference-freeevaluation
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 asks whether a general-purpose open-source language model, prompted in a loop and given only a set of known concepts, can both complete an existing taxonomy and build a new one from scratch. It claims that on taxonomy completion the prompted model is competitive with specialized trained systems on three of five benchmarks and is the best system on the food-industry CookBook taxonomy. For generation, it claims that starting from leaf concepts only and letting the model invent missing parents produces taxonomies whose reference-free quality scores match or beat the human-curated gold standard. The paper's own caveat is decisive: the model places leaves reasonably well but frequently misplaces inner, non-leaf concepts, so generated taxonomies still fall short of human-curated nuance. If the claim holds, taxonomy curation for food, medical, and other domains could scale without a corpus, fine-tuning, or hand-built seed relations.

What carries the argument

The load-bearing mechanism is a bottom-up insertion loop built on parent-query-child triplets. For each query concept the model first proposes parent concepts, drawn from the existing taxonomy in completion mode or invented in generation mode, then selects child concepts from the candidates under those parents; each proposed edge must pass a natural-language-inference check that the concept description entails the is-a relation. Pseudo-root and pseudo-leaf vertices anchor the taxonomy so a concept can be placed as the top or as a leaf. The generation variant primes the model with a taxonomy description and lets newly invented parents enter the insertion queue, which is what lets a bare set of leaf concepts grow upward into a full taxonomy.

What would settle it

Take the generated SemEval-Food taxonomy from the paper's generation experiment and have human annotators label every sampled edge as a genuine is-a relation or not; if the fraction of valid edges is low while the reported NLIV-W and CSC scores remain high, the reference-free metrics are overstating quality and the central claim for generation falls.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that iterative prompting turns a general-purpose open-source LLM into a working taxonomy completion and generation system. The task is reframed as a recursive sequence of parent-query-child triplet placements: a query concept is inserted as a child of one or more parents and as a parent of selected children, so attachments are not treated as independent and newly invented concepts can be queued for later insertion. At each step the model retrieves the most relevant existing edges as context, uses chain-of-thought prompting to propose parents and then to choose children from a candidate list, and filters every proposed relation through a natural-language-inference model that requires the concept description to entail the is-a claim. The paper reports that this approach is competitive with trained state-of-the-art systems on three of the five completion benchmarks and is the best system on the food-industry CookBook taxonomy, while generation starting from leaf concepts alone reaches only near-zero exact agreement with gold-standard positions despite reference-free scores comparable to or better than the gold-standard taxonomies. The authors state plainly that the remaining weakness is non-leaf placement, where frequent erroneous classifications are not fully captured by the reference-free metrics.

Load-bearing premise

The load-bearing premise is that the automatic quality scores used when no gold standard exists actually reflect whether a taxonomy is correct; the paper itself notes these scores can miss invalid relationships, and the near-zero agreement with human-made taxonomies suggests they may be too generous.

Editorial extensions

If this is right

  • Taxonomy completion can be handled without fine-tuning or corpus construction: retrieval plus chain-of-thought prompting on an open-source LLM is enough to be competitive on most tested benchmarks.
  • A taxonomy can be seeded from a list of known leaf concepts alone, with missing intermediate categories invented by the model during the bottom-up loop.
  • Reference-free metrics make it possible to compare generated taxonomies when no gold standard exists, which is the common practical case.
  • Improving non-leaf placement is the critical next step; until it improves, generated taxonomies are drafts rather than final curated structures.
  • The recursive triplet formulation means the same algorithm covers expansion and generation, so gains in one task can transfer to the other.

Reading between the lines

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

  • If the reference-free scores are taken as quality measures, they may reward internally coherent hierarchies more than externally true is-a relations; a practitioner should manually audit a sample of edges before deployment.
  • Because the generation pipeline filters with an NLI model and the NLIV metric also scores with NLI, part of the reported quality may reflect alignment between the filter and the evaluator rather than correctness against a human standard.
  • The non-leaf bottleneck may be sensitive to model scale or instruction tuning; a natural experiment is to rerun the pipeline with a different open-source model and watch whether non-leaf F1 moves.
  • The concept-driven recipe should transfer beyond food and medicine to product catalogs or biological ontologies, provided a clean leaf list with descriptions is available.
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 proposes LLM-based algorithms for taxonomy completion and generation, using Llama-3-70B-Instruct with retrieval-augmented generation, chain-of-thought prompting, NLI-based verification of predicted relations, and backtracking. Completion is evaluated on five taxonomies (SemEval-Food, SemEval-Verb, MeSH, Wikidata, CookBook) against specialized baselines, while generation is evaluated by reconstructing taxonomies from leaf concepts of SemEval-Food and MeSH, using both gold-standard F1 and the reference-free metrics CSC and NLIV introduced in the authors' prior work. The central claims are that completion is competitive with state-of-the-art methods on three of five datasets, and that generation from a set of known concepts is 'promising' despite difficulty in placing inner nodes (Abstract, Section 4.3, Section 5.3).

Significance. If the completion result holds, it is a useful demonstration that a single open-source LLM with retrieval and prompting can approach or match specialized taxonomy completion models without fine-tuning. The paper has notable strengths: it releases code and datasets, uses paired randomization tests for completion comparisons, includes ablation studies for backtracking and NLI verification, and is candid about the weakness of inner-node placement in generation. The generation claim, however, rests on uncalibrated reference-free metrics that are partially aligned with the method's own NLI filtering, and the reported gold-standard F1 is near zero. The paper's contribution is therefore a solid empirical study of LLM-based completion, but the generation result is not established as stated. The authors' explicit caveats and public resources make the work valuable to the community even in its current form.

major comments (4)
  1. [§5.2, §5.3, Table 5] The generation claim is supported only by the reference-free metrics CSC and NLIV, but these metrics are not calibrated: no baseline or control establishes that they can distinguish a valid hierarchy from a nonsense one. The paper itself states that CSC 'does not respect that there might be invalid relationships in the taxonomy' and that qualitative inspection reveals 'frequent erroneous classifications...not well captured by the metrics.' With Position-F1 of only 0.0234 on SemEval-Food and 0.0094 on MeSH, these metrics are the sole support for the 'promising' generation result. A control condition, such as scoring a randomly rewired version of the same node set or a deliberately nonsensical taxonomy, is needed to show that the metrics have discriminative validity in this setting.
  2. [§3.5, §5.2] The generation pipeline is method-aligned with the evaluation metric: Section 3.5 filters predicted children through an NLI entailment model, while Section 5.2 defines NLIV-S by NLI entailment of the same kind of relation. Even though different NLI models are used for inference and evaluation, the generation procedure is explicitly optimized for the property that NLIV measures, so high NLIV scores are partly a consequence of the pipeline design rather than independent evidence of taxonomy quality. The paper acknowledges this bias but does not quantify it or provide an independent evaluation, such as a human judgment study or a held-out gold standard that allows for multiple valid taxonomies.
  3. [§4.3, Table 3] No contamination analysis is provided for the benchmark taxonomies. SemEval-Food, SemEval-Verb, and MeSH are widely used, public taxonomies, and Llama-3's pretraining data may well contain them or near-equivalent hypernymy relations from WordNet and other sources. The competitive completion results could therefore reflect memorization rather than the proposed method's generalization ability. The manuscript should either present a contamination check (for example, testing whether the model reproduces taxonomy edges verbatim) or discuss this threat explicitly as a limitation that affects the interpretation of the headline 'competitive with state-of-the-art' claim.
  4. [Appendix B.1, Table 3] The comparison with baselines may not be apples-to-apples. The paper states that baseline implementations were reused and 'adjusted...by adding the functionality to output the best placements (triplets) for a query instead of a ranked list.' The details of this conversion, and whether the selected triplets correspond to the top-ranked positions under each baseline's scoring function, are not given. If the conversion pipeline is not carefully matched across methods, the reported F1 differences could reflect implementation choices rather than method quality. The authors should describe the triplet extraction procedure for baselines and ideally validate it by reproducing at least one published baseline score.
minor comments (5)
  1. [§3.2, Appendix A.4] The text invokes the 'demonstrate-search-predict (DSP) paradigm' but the implementation details concern the DSPy library; please clarify the relationship between the DSP paradigm and the DSPy framework, since the terminology may confuse readers.
  2. [§5.2] The definition of CSC should state the exact correlation coefficient (e.g., Spearman or Pearson) and the embedding model used for semantic similarity scores, so that the metric is fully reproducible.
  3. [Figure 2(c)] The figure caption and the surrounding text refer to 'alpha tocopheral'; this appears to be a typo for 'alpha tocopherol' and should be corrected.
  4. [Tables 6 and 9] Tables 6 and 9 contain the same generated-taxonomy statistics; duplicate presentation is unnecessary, and Table 9 contains the typo 'brenchmark' in its caption.
  5. [§3.4, Listings 1 and 2] The prompt listings contain inconsistent spacing within words such as 'su pe rt yp e-s ub ty pe'; please ensure the rendered prompts match the exact strings used in the experiments.

Circularity Check

2 steps flagged · score 6.0 of 10

Generation evaluation is partially circular: the NLIV metric re-checks the same NLI-entailment property used to filter edges during generation, and the only positive support for the generation claim comes from the authors' own reference-free metrics; completion results are independently benchmarked.

  1. self definitional [Section 3.5 (Verifying Placements), Section 5.2 (Evaluation), and Algorithm 2 lines 11-12]
    "To mitigate this, we require that the description of the concept (premise) entails the relation (hypothesis) when passed through an natural language inference (NLI) model for all predicted children. ... Note that due to our model-internal NLI-verification (see Section 3.5), results might be biased towards our model."

    Algorithm 2 (lines 11-12) keeps only child edges for which entails(dq, 'lemma(c) is a lemma(q)') holds, so every accepted child relation in the generated taxonomy is, by construction, one that passes NLI entailment. NLIV, defined in Section 5.2, 'estimates classification probabilities with NLI and normalizes them by walk length', i.e. it scores exactly whether edges pass NLI entailment. Although the inference and evaluation use different NLI models, the property measured is the same property enforced by the method's own filter; a high NLIV-S on generated taxonomies is therefore partly a re-statement of the filter rather than independent evidence of quality. The paper's own caveat that 'results might be biased towards our model' is a direct admission of this reduction.

  2. self citation load bearing [Section 5.2 and Section 5.3 (Generation Experiments)]
    "In addition to gold-standard comparisons, we rely on recently introduced reference-free metrics which evaluate the robustness and logical adequacy of generated taxonomies (Wullschleger et al., 2025). ... our reference-free scores are competitive with the gold standard and according to CSC even better on both datasets."

    The only positive evidence offered for the generation claim ('promising') is the CSC/NLIV scores, which come from the authors' own prior paper (Wullschleger et al., 2025). The gold-standard Position-F1 for generated taxonomies is 0.0234 on SemEval-Food and 0.0094 on MeSH (Table 5), and the authors concede 'frequent erroneous classifications...not well captured by the metrics.' Since CSC/NLIV are uncalibrated and NLIV is aligned with the method's own NLI filter, the conclusion that generation is 'promising' rests on a self-citation chain rather than on independent, externally validated evidence. This self-citation is load-bearing, not peripheral.

full rationale

The completion experiments (Section 4) are not circular: they are evaluated against gold-standard triples and compared with external baselines such as TacoPrompt, QEN, and TEMP, so those results stand independently. The circularity concerns the generation experiments (Section 5). There, the paper argues that generated taxonomies are 'promising' based primarily on reference-free metrics CSC and NLIV, both introduced in the authors' own prior work (Wullschleger et al., 2025). More importantly, the generation algorithm is explicitly filtered by NLI entailment (Section 3.5, Algorithm 2), while NLIV measures NLI entailment of edges (Section 5.2). Thus a strong NLIV score is substantially built into the method, as the paper itself notes: 'due to our model-internal NLI-verification, results might be biased towards our model.' The use of two different NLI models mitigates but does not remove the structural alignment. The paper also concedes that CSC 'does not respect that there might be invalid relationships' and that 'frequent erroneous classifications...are not well captured by the metrics,' further undermining the reference-free support. Because the central generation claim is only as strong as these method-aligned, self-cited metrics, while gold-standard F1 is near zero, a partial circularity score of 6 is appropriate. The score is not higher because the completion results are independently validated and because the generation pipeline is not fully determined by the metrics alone.

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

The central claim rests on a chain of assumptions: the LLM can produce valid taxonomic relations from context, FastText retrieval surfaces relevant edges, NLI reliably checks is-a entailment, and the authors' reference-free metrics measure taxonomy quality. The last is the most fragile because the method is filtered with NLI and evaluated with NLI-based metrics, and the paper's own examples show the metrics miss errors. Prompt and retry hyperparameters are manually tuned on validation data.

free parameters (5)
  • Manual prompt templates = Hand-tuned prompt text (Listings 1-2, App. B.2)
    Optimized by manual trial and error on validation data (Section 3.4); performance is tied to these exact instructions.
  • Taxonomy description sample size = 100
    Algorithm 2 step 2 samples 100 nodes for the LLM to write a taxonomy description; chosen without reported sensitivity analysis.
  • Backtracking retry limit = 3
    Section 3.6: re-prompts at most three times; chosen by hand.
  • Concept length limit = six or more words
    Section 3.6 assertion 6 rejects concepts of six or more words; ad hoc rule.
  • Retrieval top-k k
    Number of edges retrieved by FastText in Algorithm 1/2 is not specified in the paper; set in code, affecting context quality.
assumptions (5)
  • domain assumption Taxonomies are DAGs with hypernym-hyponym edges
    Problem definition in Section 3.1; standard for the field, not independently proven.
  • domain assumption FastText embeddings are suitable for retrieving the most relevant taxonomy edges for a query
    Section 3.2 step 1 and Algorithm 1 line 2; if retrieval misses the true parent neighborhood, the LLM cannot place the concept correctly.
  • domain assumption NLI models can reliably detect whether concept descriptions entail is-a relations
    Sections 3.5 and 5.2 use NLI both to filter edges and to score taxonomies; the paper reports that the LLM sometimes confuses similarity with hypernymy, so this assumption is fragile.
  • domain assumption The LLM's parametric knowledge plus the retrieved context is sufficient to propose valid taxonomic parents and children
    Core to both algorithms; the paper provides no contamination analysis for benchmark taxonomies, so part of the model's success may come from memorization rather than reasoning from context.
  • ad hoc to paper Reference-free metrics CSC and NLIV are valid proxies for taxonomy quality
    These metrics are introduced in the authors' own prior preprint (Wullschleger et al., 2025) and are the primary evidence for generation quality; Section 5.3 admits they miss invalid is-a relations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FoodTaxo: Generating Food Taxonomies with Large Language Models." pith.science (2026). https://pith.science/paper/66ZGQG2K

@misc{pith2026250519838,
  author       = {Pith},
  title        = {Pith review of: FoodTaxo: Generating Food Taxonomies with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66ZGQG2K}},
  note         = {Machine review of arXiv:2505.19838}
}
read the original abstract

We investigate the utility of Large Language Models for automated taxonomy generation and completion specifically applied to taxonomies from the food technology industry. We explore the extent to which taxonomies can be completed from a seed taxonomy or generated without a seed from a set of known concepts, in an iterative fashion using recent prompting techniques. Experiments on five taxonomies using an open-source LLM (Llama-3), while promising, point to the difficulty of correctly placing inner nodes.

Figures

Figures reproduced from arXiv: 2505.19838 by the authors.

Figure 1
Figure 1. The basic intuition behind the generation [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Examples of generated sub-graphs of the tax [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 7 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  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]

    Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. https://doi.org/10.1162/tacl_a_00051 Enriching word vectors with subword information . Transactions of the Association for Computational Linguistics, 5:135--146

    doi:10.1162/tacl_a_00051 2017

  4. [4]

    Georgeta Bordea, Els Lefever, and Paul Buitelaar. 2016. https://doi.org/10.18653/v1/S16-1168 S em E val-2016 task 13: Taxonomy extraction evaluation ( TE x E val-2) . In Proceedings of the 10th International Workshop on Semantic Evaluation ( S em E val-2016) , pages 1081--1091, San Diego, California. Association for Computational Linguistics

    doi:10.18653/v1/s16-1168 2016

  5. [5]

    Boqi Chen, Fandi Yi, and Dániel Varró. 2023. https://doi.org/10.1109/MODELS-C59198.2023.00097 Prompting or fine-tuning? a comparative study of large language models for taxonomy construction . In 2023 ACM/IEEE International Conference on Model Driven Engineering Languages and Systems Companion (MODELS-C), pages 588--596

    arXiv 2023

  6. [6]

    Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. https://doi.org/10.18653/v1/P18-1128 The hitchhiker`s guide to testing statistical significance in natural language processing . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1383--1392, Melbourne, Australia. Associat...

    doi:10.18653/v1/p18-1128 2018

  7. [7]

    Faisal Mahbub Chowdhury, and Nandana Mihindukulasooriya

    Nicolas Rodolfo Fauceglia, Alfio Gliozzo, Sarthak Dash, Md. Faisal Mahbub Chowdhury, and Nandana Mihindukulasooriya. 2019. https://doi.org/10.18653/v1/D19-3005 Automatic taxonomy induction and expansion . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language...

    work page doi:10.18653/v1/d19-3005 2019

  8. [8]

    Christiane Fellbaum. 2010. https://doi.org/10.1007/978-90-481-8847-5_10 WordNet , pages 231--243. Springer Netherlands, Dordrecht

    work page doi:10.1007/978-90-481-8847-5_10 2010

Show all 34 references
  1. [9]

    David Jurgens and Mohammad Taher Pilehvar. 2016. https://doi.org/10.18653/v1/S16-1169 S em E val-2016 task 14: Semantic taxonomy enrichment . In Proceedings of the 10th International Workshop on Semantic Evaluation ( S em E val-2016) , pages 1092--1102, San Diego, California. ...

    2016 doi

  2. [10]

    Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang, Christopher Potts, and Matei Zaharia. 2023 a . https://arxiv.org/abs/2212.14024 Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive nlp . Preprint, arXiv:2212.14024

    2023 arXiv

  3. [11]

    Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2023 b . https://arxiv.org/abs/2310.03714 Dspy: Compiling declar...

    2023 arXiv

  4. [12]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. 2020. https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205f78...

    2020

  5. [13]

    C E Lipscomb. 2000. Medical subject headings ( MeSH ). Bulletin of the Medical Library Association, 88(3):265--266

    2000

  6. [14]

    Zichen Liu, Hongyuan Xu, Yanlong Wen, Ning Jiang, HaiYing Wu, and Xiaojie Yuan. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.313 TEMP : Taxonomy expansion with dynamic margin loss through taxonomy-paths . In Proceedings of the 2021 Conference on Empirical Methods in Natur...

    2021 doi

  7. [15]

    Mingyu Derek Ma, Muhao Chen, Te-Lin Wu, and Nanyun Peng. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.353 H yper E xpan: Taxonomy expansion with hyperbolic representation learning . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4182--...

    2021 doi

  8. [16]

    Emaad Manzoor, Rui Li, Dhananjay Shrouty, and Jure Leskovec. 2020. https://doi.org/10.1145/3366423.3380271 Expanding taxonomies with implicit edge semantics . In Proceedings of The Web Conference 2020, WWW '20, page 2044–2054, New York, NY, USA. Association for Computing Machinery

    2020

  9. [17]

    Daniele Margiotta, Danilo Croce, and Roberto Basili. 2023. Taxosbert: Unsupervised taxonomy expansion through expressive semantic similarity. In Deep Learning Theory and Applications, pages 295--307, Cham. Springer Nature Switzerland

    2023

  10. [18]

    Aristides Milios, Siva Reddy, and Dzmitry Bahdanau. 2023. https://doi.org/10.18653/v1/2023.genbench-1.14 In-context learning for text classification with many labels . In Proceedings of the 1st GenBench Workshop on (Benchmarking) Generalisation in NLP, pages 173--184, Singapor...

    2023 doi

  11. [19]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference...

    2019 doi

  12. [20]

    Jiaming Shen, Zhihong Shen, Chenyan Xiong, Chi Wang, Kuansan Wang, and Jiawei Han. 2020. https://doi.org/10.1145/3366423.3380132 Taxoexpan: Self-supervised taxonomy expansion with position-enhanced graph neural network . In Proceedings of The Web Conference 2020, WWW '20, page...

    2020

  13. [21]

    Vanni, Brian M

    Jiaming Shen, Zeqiu Wu, Dongming Lei, Chao Zhang, Xiang Ren, Michelle T. Vanni, Brian M. Sadler, and Jiawei Han. 2018. https://doi.org/10.1145/3219819.3220115 Hiexpan: Task-guided taxonomy construction by hierarchical tree expansion . In Proceedings of the 24th ACM SIGKDD Inte...

    2018

  14. [22]

    Arnav Singhvi, Manish Shetty, Shangyin Tan, Christopher Potts, Koushik Sen, Matei Zaharia, and Omar Khattab. 2024. https://arxiv.org/abs/2312.13382 Dspy assertions: Computational constraints for self-refining language model pipelines . Preprint, arXiv:2312.13382

    2024 arXiv

  15. [23]

    Suyuchen Wang, Ruihui Zhao, Yefeng Zheng, and Bang Liu. 2022. https://doi.org/10.1145/3485447.3511943 Qen: Applicable taxonomy completion via evaluating full taxonomic relations . In Proceedings of the ACM Web Conference 2022, WWW '22, page 1008–1017, New York, NY, USA. Associ...

    2022

  16. [24]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf Chain-of-thought prompting elicits reasoning...

    2022

  17. [25]

    Zhibiao Wu and Martha Palmer. 1994. https://doi.org/10.3115/981732.981751 Verb semantics and lexical selection . In 32nd Annual Meeting of the Association for Computational Linguistics, pages 133--138, Las Cruces, New Mexico, USA. Association for Computational Linguistics

    1994

  18. [26]

    Pascal Wullschleger, Majid Zarharan, Donnacha Daly, Marc Pouly, and Jennifer Foster. 2025. https://arxiv.org/abs/2505.11470 No gold standard, no problem: Reference-free evaluation of taxonomies . Preprint, arXiv:2505.11470

    2025 arXiv

  19. [27]

    Hongyuan Xu, Ciyi Liu, Yuhang Niu, Yunong Chen, Xiangrui Cai, Yanlong Wen, and Xiaojie Yuan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.979 T aco P rompt: A collaborative multi-task prompt learning method for self-supervised taxonomy completion . In Proceedings of the 2...

    2023 doi

  20. [28]

    Yue Yu, Yinghao Li, Jiaming Shen, Hao Feng, Jimeng Sun, and Chao Zhang. 2020. https://doi.org/10.1145/3394486.3403145 Steam: Self-supervised taxonomy expansion with mini-paths . In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining...

    2020

  21. [29]

    Qingkai Zeng, Jinfeng Lin, Wenhao Yu, Jane Cleland-Huang, and Meng Jiang. 2021. https://doi.org/10.1145/3447548.3467308 Enhancing taxonomy completion with concept generation via fusing relational representations . In Proceedings of the 27th ACM SIGKDD Conference on Knowledge D...

    2021

  22. [30]

    Biao Zhang, Barry Haddow, and Alexandra Birch. 2023. https://proceedings.mlr.press/v202/zhang23m.html Prompting large language model for machine translation: A case study . In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Ma...

    2023

  23. [31]

    Jieyu Zhang, Xiangchen Song, Ying Zeng, Jiaze Chen, Jiaming Shen, Yuning Mao, and Lei Li. 2021. https://doi.org/10.1609/aaai.v35i5.16596 Taxonomy completion via triplet matching network . Proceedings of the AAAI Conference on Artificial Intelligence, 35(5):4662--4670

    2021 doi

  24. [32]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  25. [33]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  26. [34]

    lemma (q) is a lemma (p)

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2017

Pith tools

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