Pith. sign in

REVIEW 3 major objections 3 minor 34 references

Mind the Language Gap in Digital Humanities: LLM-Aided Translation of SKOS Thesauri

T0 review · 3 major / 3 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 The paper claims WOKIE makes automatic translation of SKOS thesauri accurate and scalable, with English pre-translation lifting ontology-matching F1 from 0 to 0.5.

desk verdict WOKIE is a genuinely useful open-source translation pipeline for SKOS thesauri, backed by a strong ontology-matching result; its translation-quality claims rest on a back-translation similarity proxy and need human validation to be taken at face value. read the letter →

arxiv 2507.19537 v2 pith:7K2ACCBT submitted 2025-07-22 cs.CL

Felix Kraus , Nicolas Blumenröhr , Danah Tonne , Achim Streit This is my paper
classification cs.CL
keywords SKOSthesaurimachinetranslationlargelanguagemodelsontologymatchingdigitalhumanitiesmultilingualvocabulariesLLMrefinementpipeline
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

Large parts of the Digital Humanities publish knowledge in English-only controlled vocabularies, and the paper argues this excludes communities and makes cross-lingual matching fail. To close that gap, the authors build WOKIE, a pipeline that translates each term with several lightweight web translation services and invokes a large language model only when those services disagree. The central claim is that this selective refinement produces accurate and scalable translations of SKOS thesauri without specialised infrastructure, and that pre-translating labels to English substantially improves ontology matching. The strongest quantitative evidence is one benchmark pair where the matching F1 score rose from 0 to 0.5 after translation.

What carries the argument

The mechanism that carries the argument is WOKIE, a two-stage pipeline: one or more external translation services produce candidates for each skos:prefLabel, and a frequency-based confidence score accepts the majority candidate when it reaches a user-set threshold. Otherwise an LLM receives the term plus its description or context, returns a translation, and if that translation is not among the primary candidates, a second LLM prompt selects the best candidate from the full set. This design keeps cost and latency low while reserving LLM judgment for ambiguous terms. Evaluation rests on back-translation, removing all labels in a target language, translating them back, and comparing to the removed originals with four similarity measures, plus an ontology-matching benchmark that compares F1 scores with and without English pre-translation.

What would settle it

Take a random sample of WOKIE's back-translations and have two human experts rate their correctness against the source concepts, then compare those ratings with WOKIE's similarity scores; if human-rated quality stays high but similarity scores fall whenever reference labels are defective, the quantitative quality claims are an artefact of the ground truth. Alternatively, rerun the ontology-matching experiment on a set of non-English thesauri where matchers already work, to see whether pre-translation still helps or only rescues the specific benchmark pairs.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that combining external translation services with targeted LLM refinement makes automated translation of Digital Humanities thesauri accurate, automatic, and scalable. The paper demonstrates that LLM refinement and selection in ambiguous cases improves translation quality relative to any single service, that a confidence threshold of 0.6 with five primary candidates is the resource-quality sweet spot, and that translating all labels into English before ontology matching makes matching systems work on language pairs where they previously failed or scored zero. The headline quantitative result is the rise in F1 from 0 to 0.5 on an English-Italian pair after pre-translation.

Load-bearing premise

The load-bearing assumption is that the existing multilingual labels in the thesauri are correct enough to serve as ground truth, so that back-translation similarity scores reflect real translation quality.

Editorial extensions

If this is right

  • SKOS thesauri with only one or two languages can be enriched with additional languages automatically, lowering the barrier for non-English-speaking researchers.
  • Ontology matching systems that assume English input can be applied to non-English thesauri by running WOKIE as a preprocessing step, without changing the matcher itself.
  • The reported optimal configuration, five primary candidates and a 0.6 confidence threshold, gives a concrete default that keeps LLM calls and cost modest.
  • LLM refinement helps most on low-resource or historically relevant languages such as Latin, where the best single service alone scores near zero.
  • Small-to-medium thesauri of up to about 10,000 terms can be translated in a few hours on ordinary hardware.

Reading between the lines

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

  • Beyond the paper, the same disagreement-then-refine design could be applied to other structured vocabularies, such as Wikidata labels or domain ontologies, whenever multiple translation sources and a context field are available.
  • A consequence of the paper's own admission that reference labels can be defective is that the reported similarity scores are only as trustworthy as those labels; a human-verified gold standard would be needed to confirm the quantitative quality claims.
  • The recommendation of specific LLMs will age quickly; the lasting contribution may be the modular interface and the threshold heuristic rather than the particular model ranked best at the time of writing.
  • A testable extension would be to measure whether machine-generated labels are actually used and understood by the communities the paper wants to include, rather than only comparing strings to existing labels.
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

3 major / 3 minor

Summary. The paper introduces WOKIE, an open-source, modular pipeline for automatically translating SKOS thesaurus labels. The pipeline combines primary external translation services with an LLM-based refinement step that is triggered when the primary translators disagree. The evaluation is structured around five research questions: comparing primary translation services, assessing LLM refinement, tuning pipeline parameters (confidence threshold, minimum number of translations, temperature), comparing LLMs, and measuring the impact of English pre-translation on ontology matching. The main quantitative findings are that combining primary services with LLM refinement yields the best back-translation similarity scores, that a threshold of 0.6 and five minimum translations are optimal, that Gemini 2.0 Flash performs best among the tested LLMs, and that pre-translating to English raises ontology matching F1 scores substantially, including from 0 to 0.5 on one benchmark pair.

Significance. If the translation-quality claims are validated, WOKIE would be a useful, accessible tool for the Digital Humanities community, where multilingual thesauri are scarce and commercial MT systems often handle domain-specific and low-resource languages poorly. The paper's strengths include a fully open-source implementation, a reproducible evaluation with all scripts and data deposited on Zenodo, and a concrete ontology matching experiment grounded in an external benchmark with ground-truth alignments. The OM improvement is the most convincing evidence of practical value. However, the primary translation-quality evaluation (RQ1-RQ4) rests on a back-translation similarity metric that measures consistency with original labels rather than translation adequacy, and the same metric is used to select the optimal configuration and representative LLM. This weakens the central claims that WOKIE enables 'accurate' translation and that LLM refinement 'substantially improved' translation quality. The authors themselves acknowledge in Section 6 that the reference labels are sometimes wrong, which further undermines the proxy metric.

major comments (3)
  1. [§4.2, §5.3–§5.5, §7] The evaluation of RQ1–RQ4 uses back-translation similarity to the original labels as the measure of translation quality. This measures round-trip consistency, not translation adequacy: a correct translation can legitimately differ from the original through synonymy, morphology, or register, and the similarity metrics (exact match, Levenshtein, Jaro-Winkler, BPEmb cosine) systematically penalize such valid alternatives. The paper's own example in Section 6—where WOKIE's "rot auf rot (a3 γ)" is arguably a better translation than the defective reference "Rot on red (a3c)" and is nevertheless penalized—demonstrates that the reference labels are not always correct. Consequently, the reported scores can misstate the relative quality of services and the benefit of LLM refinement, and the Section 7 conclusion that WOKIE enables "accurate, automatic and scalable translation" is not supported by the presented evidence. I recommend adding a human expert evaluation on a representative sample of translated terms, or using an external gold standard with verified reference labels, before the translation-quality claims are accepted.
  2. [§5.3–§5.5] The optimal threshold, minimum number of primary translations, temperature, and the representative LLM (Gemini 2.0 Flash) are all selected using the same back-translation similarity scores that are later cited as evidence of translation quality. This creates an empirical circularity: the configuration is tuned to maximize agreement with the original labels on the test set, so the reported improvements of the tuned configuration and the LLM ranking in Figure 5 are partly an artifact of the selection procedure. The conclusion that LLM-based refinement improves translation quality would be considerably strengthened by an out-of-sample evaluation, for example by holding out a subset of thesauri during parameter selection or by using independent human judgments on a held-out test set.
  3. [§6, §7] The paper's concluding claims conflate the ontology matching result with translation quality. Section 5.6 demonstrates a concrete and reproducible improvement in OM F1 scores after English pre-translation, with an external ground truth. However, this experiment only addresses RQ5 and does not validate RQ1–RQ4, because an OM system can benefit from approximate or imperfect translations. The Section 7 statements that "LLM-based refinement substantially improved translation quality" and that WOKIE provides "accurate" translation should be separated from the OM evidence; the OM result is strong on its own and does not need to be supplemented by the weaker back-translation evidence. I suggest reframing the contributions to distinguish the validated OM preprocessing benefit from the translation-quality claims that still require additional evaluation.
minor comments (3)
  1. [§4.2] The text states that the four similarity measures "provide relative comparisons rather than absolute similarity," but Section 7 uses the unqualified term "accurate" to describe the translation quality. Please add a qualifier such as "consistent with the original labels" or "as measured by back-translation similarity" to avoid overstatement.
  2. [§5.6, Figure 6] The caption and text name "Matcha" as one of the matching systems, but the two systems used are not both identified in the main text. Please clarify which benchmark systems were run and label them explicitly in Figure 6.
  3. [§3.2.4] The prompt example in Section 3.2.4 contains an inline parenthetical "(first part identical to translation prompt)" that appears to be a meta-instruction rather than part of the actual prompt. Please move this explanation outside the prompt text or mark it as an editorial note.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: WOKIE's central evaluation is an external back-translation benchmark, not a derivation that reduces to its own inputs; the self-citation used for dataset selection is not load-bearing.

full rationale

The paper contains no derivation chain in which a claimed result is equivalent to its inputs by construction. The pipeline's translation quality is assessed by removing one language's labels, back-translating the remaining labels, and comparing to the originals with string and embedding similarities (Section 4.2). This is an evaluation proxy, not a self-definitional loop: the original labels are human-created thesaurus labels, and the pipeline is not defined in terms of those labels. The selection of a confidence threshold (Section 5.4) and the representative LLM Gemini 2.0 Flash (Section 5.3) does use the same similarity measure as the later LLM comparison (Section 5.5), which is a real methodological limitation and a potential source of selection bias, but it is hyperparameter selection rather than a fitted parameter being renamed as a prediction. Section 6 explicitly concedes that the ground-truth labels are not always reliable, giving the 'rot auf rot' example where a correct translation is penalized; this weakens the validity of RQs 1-4 but is not circularity. The ontology matching result (Section 5.6, F1 rising from zero to 0.5) uses an external OAEI benchmark and provides independent evidence for the preprocessing value of WOKIE. The only self-citation is reference [16], a prior benchmark by the same first author, used in Section 4.1 to motivate thesauri selection; it is not load-bearing for the central translation-quality claims. Overall, no equation-level or definition-level circularity is present; the paper's limitations concern measurement validity and evaluation leakage, not circular reduction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The pipeline introduces no free theoretical parameters or invented entities. The load-bearing assumptions are empirical: similarity-to-reference labels is treated as translation quality; the provenance of the removed language is taken from project history and developer contact; and the eight selected thesauri are treated as representative of DH practice. The numerical settings (threshold 0.6, minimum 5 translations, temperature 0) are hand-tuned on the evaluation set rather than derived.

free parameters (4)
  • confidence threshold = 0.6
    Tuned in Section 5.4 by comparing 0.4, 0.6, 0.8 on test thesauri; 0.6 chosen as the cost-quality sweet spot.
  • minimum number of primary translations = 5
    Selected in Section 5.4 from tests with values 3, 5, and 8; the paper also recommends the product of threshold and minimum be at least 3.
  • LLM temperature = 0
    Chosen in Section 5.4 after finding no notable performance differences for temperatures 0, 0.5, 1; 0 is deterministic.
  • representative LLM (Gemini 2.0 Flash) = Gemini 2.0 Flash
    Selected after preliminary tests as best-performing and low-cost in Section 5.3; used to claim refinement benefits in Figure 3.
assumptions (3)
  • domain assumption Back-translation string/semantic similarity is a sufficient proxy for translation quality.
    All translation quality claims (RQs 1-4) are evaluated by translating removed labels and comparing to originals via Levenshtein, Jaro-Winkler, and BPEmb cosine similarity (Section 4.2). No human evaluation is performed in this paper; the authors plan manual review in future work (Section 7).
  • domain assumption The language identified as the thesaurus's original development language is correct, so remaining labels are not MT-generated.
    Section 4.1: 'We determined this by examining early versions, related projects, or contacting developers.' If this provenance is wrong, back-translating MT-generated labels would bias scores.
  • domain assumption The eight selected thesauri are representative of DH SKOS thesauri.
    Section 6: 'we do not claim to cover the wide range of domains belonging to or using elements of the DH.' The generalization to untested languages is asserted without direct evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mind the Language Gap in Digital Humanities: LLM-Aided Translation of SKOS Thesauri." pith.science (2026). https://pith.science/paper/7K2ACCBT

@misc{pith2026250719537,
  author       = {Pith},
  title        = {Pith review of: Mind the Language Gap in Digital Humanities: LLM-Aided Translation of SKOS Thesauri},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7K2ACCBT}},
  note         = {Machine review of arXiv:2507.19537}
}
read the original abstract

We introduce WOKIE, an open-source, modular, and ready-to-use pipeline for the automated translation of SKOS thesauri. This work addresses a critical need in the Digital Humanities (DH), where language diversity can limit access, reuse, and semantic interoperability of knowledge resources. WOKIE combines external translation services with targeted refinement using Large Language Models (LLMs), balancing translation quality, scalability, and cost. Designed to run on everyday hardware and be easily extended, the application requires no prior expertise in machine translation or LLMs. We evaluate WOKIE across several DH thesauri in 15 languages with different parameters, translation services and LLMs, systematically analysing translation quality, performance, and ontology matching improvements. Our results show that WOKIE is suitable to enhance the accessibility, reuse, and cross-lingual interoperability of thesauri by hurdle-free automated translation and improved ontology matching performance, supporting more inclusive and multilingual research infrastructures.

Figures

Figures reproduced from arXiv: 2507.19537 by the authors.

Figure 1
Figure 1. For each term and selected property (default is [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 1
Figure 1. Simplified flowchart diagram of WOKIE’s translation steps using "marginal gloss" as an example. The involvement of external translation services is marked in yellow, the use of LLMs blue. Positive decisions are on the right-hand side of the diamond (green arrow), negative decisions on the left (red arrow). 5 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the performance (left) and execution time (right) of different primary services, averaged over test cases. None of the prompt strategies outperformed the others. The hierarchical method increased token usage without benefits because none of the tested representations of hierarchy relations could be correctly processed by the LLMs and quality did therefore not improve. The batch prompts provided only on… view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Similarity measures for the four test setups and multiple test cases. The removed and back-translated language is indicated in parentheses. cost, and unclear training language coverage, the optimal configuration uses external translation services with targeted LLM refi…
Figure 4
Figure 4. Figure 4: Similarity measures for different threshold values using OeAI Thesaurus back-translated to German. corresponds to accepting a translation directly if at least two, three, or four out of five primary translation candidates agree. As shown in [PITH_FULL_IMAGE:figures/fu…
Figure 5
Figure 5. Figure 5: Similarity measures for different LLMs averaged over test cases. Green indicates the best values, transitioning over yellow to red for lower scores. Additional details on pricing and availability can be found in the GitHub repository9 . The similarity scores indicate a…
Figure 6
Figure 6. Figure 6: Comparison of the F1 scores from the original benchmark "(original)" with the results obtained after translating all thesauri to English first "(translated)". F1 scores of zero are denoted with a "0", errors on execution of the matching system with "X". As shown in [P…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Ontology Label Translation

    Arcan, Mihael and Buitelaar, Paul. “Ontology Label Translation.” In: Proceedings of the 2013 NAACL HLT Student Research Workshop, ed. by Annie Louis et al. Atlanta, Georgia: Association for Computational Linguistics, June 2013, pp. 40–46

    work page 2013

  2. [2]

    Translating Terminological Expressions in Knowledge Bases with Neural Machine Translation

    Arcan, Mihael, Torregrosa, Daniel, and Buitelaar, Paul. “Translating Terminological Expres- sions in Knowledge Bases with Neural Machine Translation.” July 2019. DOI: 10.48550/ arXiv.1709.02184. arXiv: 1709.02184 [cs]

    work page Pith review arXiv doi:10.48550/arxiv.1709.02184 2019

  3. [3]

    Transfer Learning for Digital Heritage Collections: Comparing Neural Machine Translation at the Subword-level and Character-level:

    Banar, Nikolay et al. “Transfer Learning for Digital Heritage Collections: Comparing Neural Machine Translation at the Subword-level and Character-level:” in: Proceedings of the 12th International Conference on Agents and Artificial Intelligence. V alletta, Malta: SCITEPRESS - Science and Technology Publications, 2020, pp. 522–529. ISBN: 978-989-758-395-7....

    work page doi:10.5220/0009167205220529 2020

  4. [4]

    Enriching Word V ectors with Subword Information

    Bojanowski, Piotr et al. “Enriching Word V ectors with Subword Information.” In: Transac- tions of the Association for Computational Linguistics 5 (2017), ed. by Lillian Lee, Mark Johnson, and Kristina Toutanova, pp. 135–146. DOI: 10.1162/tacl_a_00051

    doi:10.1162/tacl_a_00051 2017

  5. [5]

    Language Models Are Few-Shot Learners

    Brown, Tom B. et al. “Language Models Are Few-Shot Learners.” In: Proceedings of the 34th International Conference on Neural Information Processing Systems . NIPS ’20. Red Hook, NY , USA: Curran Associates Inc., Dec. 2020, pp. 1877–1901. ISBN: 978-1-7138- 2954-6

    work page 2020

  6. [6]

    A Comparison of Personal Name Matching: Techniques and Practical Is- sues

    Christen, Peter. “A Comparison of Personal Name Matching: Techniques and Practical Is- sues.” In: Sixth IEEE International Conference on Data Mining - Workshops (ICDMW’06) . Hong Kong, China: IEEE, 2006, pp. 290–294. ISBN: 978-0-7695-2702-4. DOI: 10.1109/ ICDMW.2006.2

    work page 2006

  7. [7]

    LabelTranslator - A Tool to Automatically Localize an Ontology: 5th European Semantic Web Conference, ESWC 2008

    Espinoza, Mauricio, Gómez-Pérez, Asunción, and Mena, Eduardo. “LabelTranslator - A Tool to Automatically Localize an Ontology: 5th European Semantic Web Conference, ESWC 2008.” In: The Semantic Web. Lecture Notes in Computer Science (Including Sub- series Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (2008), pp. 792–796. ISS...

    work page doi:10.1007/978-3-540-68234-9_60 2008

  8. [8]

    Ontology Matching

    Euzenat, Jérôme and Shvaiko, Pavel. Ontology Matching . second edition. Berlin, Heidel- berg: Springer, 2013. ISBN: 978-3-642-38721-0. DOI: 10.1007/978-3-642-38721-0

    doi:10.1007/978-3-642-38721-0 2013

Show all 34 references
  1. [9]

    English-Chinese Knowledge Base Translation with Neural Net- work

    Feng, Xiaocheng et al. “English-Chinese Knowledge Base Translation with Neural Net- work.” In: Proceedings of COLING 2016, the 26th International Conference on Compu- tational Linguistics: Technical Papers , ed. by Y uji Matsumoto and Rashmi Prasad. Osaka, Japan: The COLING 20...

    2016

  2. [10]

    MLGrafViz: Multilingual Ontology Visualization Plug-in for Protégé

    Florrence, Merlin. “MLGrafViz: Multilingual Ontology Visualization Plug-in for Protégé.” In: Computer Science and Information Technologies 2., no. 1 (Mar. 2021), pp. 43–48. ISSN: 2722-3221. DOI: 10.11591/csit.v2i1.p43-48

    2021 doi

  3. [11]

    Knowledge Graphs in the Li- braries and Digital Humanities Domain

    Haslhofer, Bernhard, Isaac, Antoine, and Simon, Rainer. “Knowledge Graphs in the Li- braries and Digital Humanities Domain.” In: Encyclopedia of Big Data Technologies , ed. by Sherif Sakr and Albert Zomaya. Cham: Springer International Publishing, 2018, pp. 1–8. ISBN: 978-3-31...

    2018 doi

  4. [12]

    BPEmb: Tokenization-free Pre-trained Sub- word Embeddings in 275 Languages

    Heinzerling, Benjamin and Strube, Michael. “BPEmb: Tokenization-free Pre-trained Sub- word Embeddings in 275 Languages.” In: Proceedings of the Eleventh International Con- ference on Language Resources and Evaluation (LREC 2018) , ed. by Nicoletta Calzolari et al. Miyazaki, Ja...

    2018

  5. [13]

    Advances in Record-Linkage Methodology as Applied to Matching the 1985 Census of Tampa, Florida

    Jaro, Matthew A. “Advances in Record-Linkage Methodology as Applied to Matching the 1985 Census of Tampa, Florida.” In: Journal of the American Statistical Association 84 ., no. 406 (June 1989), pp. 414–420. ISSN: 0162-1459. DOI: 10.1080/01621459.1989. 10478785

    1985

  6. [14]

    Is ChatGPT A Good Translator? Y es With GPT-4 As The Engine

    Jiao, Wenxiang et al. “Is ChatGPT A Good Translator? Y es With GPT-4 As The Engine.” Nov. 2023. DOI: 10.48550/arXiv.2301.08745. arXiv: 2301.08745 [cs]

    2023 doi arXiv

  7. [15]

    When and Why Is Unsupervised Neural Machine Translation Useless?

    Kim, Y unsu, Graça, Miguel, and Ney, Hermann. “When and Why Is Unsupervised Neural Machine Translation Useless?” In: Proceedings of the 22nd Annual Conference of the Eu- ropean Association for Machine Translation , ed. by André Martins et al. Lisboa, Portugal: European Associa...

    2020

  8. [16]

    A Gold Standard Benchmark Dataset for Digital Humanities

    Kraus, Felix et al. “A Gold Standard Benchmark Dataset for Digital Humanities.” In: Pro- ceedings of the 19th International Workshop on Ontology Matching, ed. by Ernesto Jiménez- Ruiz et al. V ol. 3897. CEUR Workshop Proceedings. Baltimore, MD, USA: CEUR, Nov. 2024, pp. 1–17. ...

    2024

  9. [17]

    Gemini Embedding: Generalizable Embeddings from Gemini

    Lee, Jinhyuk et al. “Gemini Embedding: Generalizable Embeddings from Gemini.” Mar

  10. [18]

    Binary Codes Capable of Correcting Deletions, Insertions, and Reversals

    Levenshtein, Vladimir I. “Binary Codes Capable of Correcting Deletions, Insertions, and Reversals.” In: Soviet Physics Doklady 10 ., no. 8 (1966), pp. 707–710

    1966

  11. [19]

    Multilingual Denoising Pre-training for Neural Machine Translation

    Liu, Yinhan et al. “Multilingual Denoising Pre-training for Neural Machine Translation.” In: Transactions of the Association for Computational Linguistics 8 (Dec. 2020), pp. 726–742. ISSN: 2307-387X. DOI: 10.1162/tacl_a_00343

    2020 doi

  12. [20]

    Linguistically Motivated Evaluation of the 2023 State-of- the-art Machine Translation: Can ChatGPT Outperform NMT?

    Manakhimova, Shushen et al. “Linguistically Motivated Evaluation of the 2023 State-of- the-art Machine Translation: Can ChatGPT Outperform NMT?” In:Proceedings of the Eighth Conference on Machine Translation. Singapore: Association for Computational Linguistics, 2023, pp. 224–...

    2023 doi

  13. [21]

    Extending Statistical Machine Translation with Discriminative and Trigger-Based Lexicon Models

    Mauser, Arne, Hasan, Saša, and Ney, Hermann. “Extending Statistical Machine Translation with Discriminative and Trigger-Based Lexicon Models.” In: Proceedings of the 2009 Con- ference on Empirical Methods in Natural Language Processing , ed. by Philipp Koehn and Rada Mihalcea....

    2009

  14. [22]

    SKOS Simple Knowledge Organization System Refer- ence

    Miles, Alistair and Bechhofer, Sean. SKOS Simple Knowledge Organization System Refer- ence. W3C Recommendation. United States: World Wide Web Consortium, Aug. 2009

    2009

  15. [23]

    Integrating Multiple Knowledge Graphs in Dig- ital Humanities

    Morvillo, Alberto and Mecella, Massimo. “Integrating Multiple Knowledge Graphs in Dig- ital Humanities.” In: ST4DM 2024: Semantic Technologies for Data Management . Twente, Italy, July 2024

    2024

  16. [24]

    THOTH: Neu- ral Translation and Enrichment of Knowledge Graphs

    Moussallem, Diego, Soru, Tommaso, and Ngonga Ngomo, Axel-Cyrille. “THOTH: Neu- ral Translation and Enrichment of Knowledge Graphs.” In: The Semantic Web – ISWC 2019: 18th International Semantic Web Conference, Auckland, New Zealand, October 26– 30, 2019, Proceedings, Part I. B...

    2019 doi

  17. [25]

    Machine Transla- tion Using Semantic Web Technologies: A Survey

    Moussallem, Diego, Wauer, Matthias, and Ngomo, Axel-Cyrille Ngonga. “Machine Transla- tion Using Semantic Web Technologies: A Survey.” In: Journal of Web Semantics 51 (Aug. 2018), pp. 1–19. ISSN: 15708268. DOI: 10 . 1016 / j . websem . 2018 . 07 . 001. arXiv: 1711.09476 [cs]

    2018 arXiv

  18. [26]

    Results of the Ontology Alignment Evaluation Initiative 2024

    Pour, Mina Abd Nikooie et al. “Results of the Ontology Alignment Evaluation Initiative 2024.” In: Proceedings of the 19th International Workshop on Ontology Matching (OM 2024), Baltimore, USA, November 11, 2024. Ed.: E. Jiménez-Ruiz, O. Hassanzadeh, C. Tro- jahn, S. Hertling, ...

    2024

  19. [27]

    The Effect of Sampling Temperature on Problem Solving in Large Lan- guage Models

    Renze, Matthew. “The Effect of Sampling Temperature on Problem Solving in Large Lan- guage Models.” In: Findings of the Association for Computational Linguistics: EMNLP 2024, ed. by Y aser Al-Onaizan, Mohit Bansal, and Y un-Nung Chen. Miami, Florida, USA: Association for Compu...

    2024

  20. [28]

    Text Analysis Using Deep Neural Networks in Digital Humanities and Information Science

    Suissa, Omri, Elmalech, Avshalom, and Zhitomirsky-Geffet, Maayan. “Text Analysis Using Deep Neural Networks in Digital Humanities and Information Science.” In: Journal of the Association for Information Science and Technology 73 ., no. 2 (2022), pp. 268–287. ISSN: 2330-1643. D...

    2022 doi

  21. [29]

    Comparing Rule-Based and Statistical MT Output

    Thurmair, Gregor. “Comparing Rule-Based and Statistical MT Output.” In: Proceedings of the LREC 2004 Workshop on The Amazing Utility of Parallel and Comparable Corpora . Lissabon, 2004, pp. 5–9

    2004

  22. [30]

    Editorial: Data and Workflows for Multilingual Digital Humanities

    Viola, Lorella. “Editorial: Data and Workflows for Multilingual Digital Humanities.” In: Journal of Open Humanities Data 10 (June 2024), p. 37. ISSN: 2059-481X. DOI: 10 . 5334/johd.220

    2024

  23. [31]

    String Comparator Metrics and Enhanced Decision Rules in the Fellegi- Sunter Model of Record Linkage

    Winkler, William E. “String Comparator Metrics and Enhanced Decision Rules in the Fellegi- Sunter Model of Record Linkage.” In: Proceedings of the Survey Research Methods Section, ASA (1990) (Jan. 1990), pp. 354–359

    1990

  24. [32]

    mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer

    Xue, Linting et al. “mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer.” In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies . Online: Association for Com- putationa...

    2021 doi

  25. [33]

    Multilingual MMLU Benchmark Leaderboard

    Zhou, Yi et al. “Multilingual MMLU Benchmark Leaderboard.” 2024. 17 Acronyms DH Digital Humanities LLM Large Language Model MT Machine Translation OAEI Ontology Alignment Evaluation Initiative OM Ontology Matching RQ research question SKOS Simple Knowledge Organization System ...

    2024

  26. [2025]

    arXiv: 2503.07891 [cs]

    DOI: 10.48550/arXiv.2503.07891. arXiv: 2503.07891 [cs]

    doi arXiv

Pith tools

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