Pith. sign in

REVIEW 2 major objections 6 minor 69 references

ERU-KG: Efficient Reference-aligned Unsupervised Keyphrase Generation

T0 review · 2 major / 6 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 Keyphrases without labels reach 89% of supervised performance

desk verdict Genuinely novel unsupervised KG pipeline with shipped code, but the KP20K 'test' set appears to be the training split, which puts the headline comparison with a supervised baseline on shaky ground. read the letter →

arxiv 2505.24219 v1 pith:7J7WLF36 submitted 2025-05-30 cs.CL

Lam Thanh Do , Aaditya Bodke , Pritom Saha Akash , Kevin Chen-Chuan Chang This is my paper · ORCID
classification cs.CL
keywords unsupervisedkeyphrasegenerationextractiontermimportanceSPLADEreference-basedtrainingscientificdocumentretrievalqueryexpansion
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

ERU-KG is a label-free keyphrase generator that replaces hand-designed importance scores with a learned sense of what a community treats as key. It trains SPLADE, a sparse term-importance model, on triples in which real user queries, citation contexts, and titles act as pseudo-queries for the papers they reference, and then scores each candidate phrase by summing the importance of its component terms. On five scientific benchmarks, the paper reports that this approach outperforms unsupervised keyphrase extraction and generation baselines and reaches about 89 percent of the supervised CopyRNN's top-10 performance, with CopyRNN ahead by only 1.1 percentage point in overall top-10 F1. The same model switches between generation and extraction by changing two interpolation weights, and its inference throughput is the highest among comparable baselines, which matters for processing large document collections.

What carries the argument

The load-bearing machinery is SPLADE, a sparse lexical expansion model that outputs a vocabulary-wide term-importance vector for an input document. ERU-KG trains SPLADE with a ranking loss on triplets where the query slot is filled by a reference, namely a real search query, a citation context, or a title, and the positive document is the paper it refers to. At inference, each candidate phrase's informativeness is the average of its words' interpolated importances, where the interpolation blends the input document's own term weights with weights from BM25-retrieved related documents, following a pseudo-relevance-feedback intuition that compensates for underestimated absent terms. A separate phraseness module generates candidates by extracting noun phrases from the document and its neighbors, prunes candidates by informativeness and cross-document reliability, and combines phraseness and informativeness as a product of experts with a position penalty.

What would settle it

Train ERU-KG twice on the same corpus, once with genuine references and once with matched but unrelated references (same length and topic vocabulary, drawn from different papers), and compare keyphrase F1 on a held-out set; if the mismatched-reference model does not lose substantially, the reference-alignment claim is not doing the work. A second check is to score documents that have no citations or search traffic in the training collection and see whether performance falls toward the heuristic baselines.

Watch

Extended reading notes

Core claim

The central claim is that informativeness can be learned from references rather than assumed by a heuristic. The paper argues that how a document is searched for, cited, and titled encodes what readers and experts consider its key concepts, so a term-importance model trained to match documents to their references can serve as an accurate informativeness estimator. Because the estimator works at the term level, phrase scores are cheap to compute, and because the underlying model can assign importance to words that do not appear in the document, absent keyphrases can be scored as well as present ones. In the paper's experiments this design beats existing unsupervised methods, brings the model close to a supervised sequence-to-sequence baseline, and produces keyphrases that improve retrieval when used as query and document expansion.

Load-bearing premise

The assumption that carries the method is that references, meaning queries, citation contexts, and titles, are honest mirrors of a document's key concepts; if those references systematically talk around the core content rather than about it, the learned term importances will rank the wrong phrases.

Editorial extensions

If this is right

  • Unsupervised keyphrase generation can be nearly as accurate as supervised generation: on the five benchmarks, ERU-KG's present-keyphrase F1 is best or second-best among unsupervised methods on all datasets except one, and its absent-keyphrase recall is the best on every dataset.
  • The speed numbers are practical: in the paper's throughput tests, ERU-KG-base reaches 15.5 documents per second in generation mode with a small neighbor set and 72.9 documents per second in extraction mode, faster than the comparable baselines tested.
  • Generated keyphrases can serve as human-readable expansions for first-stage retrieval: on average they raise recall at rank 1000 from 62.1 to 68.4 over BM25 alone, outperform other keyphrase-generation expansions, and match RM3 and DocT5Query while remaining concise enough to display.
  • One trained model covers two deployment modes: with the two interpolation weights set to one, neighbor documents are disabled and ERU-KG behaves as a pure extractor, so a single checkpoint can serve extraction and generation use cases.

Reading between the lines

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

  • Beyond the paper: the reference-alignment principle should transfer to any domain where documents accumulate mention contexts, such as news articles with search traffic, Wikipedia entries with inbound link anchors, or software issues with user queries, though the paper validates it only on scientific text.
  • Beyond the paper: because keyness is learned from external references, ERU-KG's rankings are community-relative; documents that are seldom queried or cited may have accurate keyphrases that the model systematically under-ranks, a failure mode not covered by the scientific benchmarks.
  • Beyond the paper: ERU-KG's term-level scores could be used as silver labels to bootstrap a seq2seq keyphrase generator, potentially combining the coverage of learned generation with the label-free ranking signal; the paper does not test this.
  • Beyond the paper: the ablations show each reference type contributes, but they do not explore weighting references by type per document; an adaptive mixture that trusts titles for extraction-oriented documents and citation contexts for absent-keyphrase documents is a natural extension.
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

2 major / 6 minor

Summary. The paper proposes ERU-KG, an unsupervised keyphrase generation model with two modules. The informativeness module trains a SPLADE term-importance predictor on reference-document pairs (queries, citation contexts, and titles) and scores candidates by aggregating term importances with a length penalty and pseudo-relevance feedback from related documents. The phraseness module extracts noun phrases from the input and from related documents, prunes them with informativeness and reliability heuristics, and combines phraseness and informativeness through a product-of-experts score with a position penalty. The model is evaluated on five keyphrase benchmarks against unsupervised and supervised baselines, on six retrieval datasets for query/document expansion, and on inference speed. The paper includes ablations for reference types, a sensitivity analysis for the interpolation weights, and a limitations section.

Significance. If the results hold, the paper makes a useful contribution: it shows that a label-free, term-level scoring approach can approach a supervised seq2seq baseline on keyphrase generation while being fast at inference, and it demonstrates retrieval-side utility of the generated keyphrases. A notable strength is that the training signal is not circular with respect to the keyphrase labels: SPLADE is trained on queries, citation contexts, and titles rather than on the benchmark keyphrases, and the authors provide ablations isolating the contribution of each reference type. The paper also ships code and includes a candid limitations section. However, the evaluation of the largest benchmark (KP20K) appears to be performed on a tuning/training split rather than the standard held-out test split, and this directly affects the headline comparison with the supervised baseline and all average results.

major comments (2)
  1. [§3.1, Table 5, Table 6, Table 1] The KP20K evaluation split is not the standard held-out test split. Table 5 reports 19,987 test documents for KP20K, whereas the standard KP20K test split contains 2,000 documents and 19,987 matches the training split after filtering. More importantly, the KP20K numbers reported for ERU-KG-base in Table 1 (F@5=22, F@10=19.4, R@5=6, R@10=8.1) are identical to the alpha=beta=0.8 row of the sensitivity analysis in Table 6, which is explicitly conducted on the KP20K dataset to select alpha and beta. This indicates that the KP20K row in Table 1 is a tuning-set result, not an evaluation on held-out data. Because CopyRNN is a supervised baseline trained on the KP20K training split, the comparison on this dataset is confounded in two ways: the supervised model is scored on data it was trained on, and ERU-KG's hyperparameters were selected on the same documents. The paper's headline claim that CopyRNN outperforms ERU-KG by only 1.1 percentage points in overall F1@10 and all average results include KP20K, so this issue is load-bearing. The authors should clarify which split is used and rerun the KP20K row on the official 2,000-document test split, with hyperparameters selected on the validation split.
  2. [Abstract and §3.3.1, Table 1] The abstract states that ERU-KG achieves 'on average 89% of the performance of a supervised model for top 10 predictions,' but the body reports that CopyRNN outperforms ERU-KG by 1.1 percentage points in overall F1@10, and Table 1's averages (26.5 vs. 27.6) imply a ratio of about 96%, not 89%. No calculation presented in the paper yields the 89% figure, so the abstract and the body are mutually inconsistent. Please either correct the abstract to reflect the reported numbers or state explicitly which metric combination produces 89%.
minor comments (6)
  1. [§3.2.2 and Table 1] The text says that macro-average F1-score and Recall are utilized for evaluation of present and absent keyphrases, but Table 1 reports only Recall for absent keyphrases; please clarify whether F1 for absent keyphrases is omitted or whether the text should read that F1 is used for present and Recall for absent keyphrases.
  2. [§3.3.1 and Table 1] The claim that 'for absent keyphrase generation, our model achieves the best performance across all benchmark datasets' is contradicted by Table 1, where UOKG has R@10=6.9 on Krapivin versus ERU-KG's 6.2 and CopyRNN has higher R@10 on four of the five datasets; the sentence should be restricted to average performance or to comparisons with unsupervised baselines, with the Krapivin exception noted.
  3. [Table 1 and §3.2.1] The asterisk definition in Table 1 says significance over AutoKeyGen, UOKG, and TPG with p<0.05, but no statistical test or multiple-comparison procedure is described; please state the test used and whether any correction was applied.
  4. [§3.4 and Table 3] The text says ERU-KG achieves the best throughput in both scenarios, but in the keyphrase generation scenario the default setting |N(x)|=100 has 10.9 doc/s, which is not significantly better than CopyRNN (11.0) or PromptKP (10.4), and only the |N(x)|=50 and |N(x)|=10 variants carry significance asterisks; the default-configuration result should be described as comparable rather than best.
  5. [§2.3 and Algorithm 1] Equation (11) writes P_kp as a probability, but Algorithm 1 explicitly treats the score as unnormalized and comments that normalization is skipped; the equation should be marked as proportional to make this consistent.
  6. [Introduction and Figure 3] The Introduction contains a typo ('provide insights into of what the community considers') and Figure 3 reports performance changes without error bars or significance information; please fix the typo and indicate whether the reported differences are stable across datasets and runs.

Circularity Check

1 steps flagged · score 6.0 of 10

KP20K benchmark row equals the alpha/beta tuning run, so part of the headline result is an in-sample score.

  1. fitted input called prediction [Section D.1 / Table 6 versus Table 1; Table 5]
    "Sensitivity analysis of interpolation hyperparameters α and β. Experiments are conducted on the KP20K dataset, using ERU-KG-base."

    Section D.1 and Table 6 select α and β on 'the KP20K dataset' and report the α=β=0.8 configuration. That exact row gives present F1@10=19.4 and absent R@10=8.1. Table 1 reports ERU-KG-base on KP20K as present F1@10=19.4 and absent R@10=8.1. Thus the KP20K row in the main benchmark is not a held-out prediction; it is the very run used to set the interpolation weights. Since the paper's headline '1.1 percentage point' F1@10 gap is the average over five datasets, and KP20K (19,987 docs per Table 5) dominates that average, the comparison to CopyRNN is partially confounded by in-sample tuning. Table 5's 'test' count also matches the original training-split size rather than the official 2,000-document test set, further indicating a benchmark-split problem rather than an independent evaluation.

full rationale

ERU-KG's central derivation is not definitionally circular: the term-importance model (SPLADE) is trained on reference-document triplets built from SciRepEval queries, unarXive citation contexts, and titles, not on the human keyphrase labels used in evaluation. The SPLADE ranking loss and the aggregation in Eqs. 4-5 are independent of the benchmark annotations. The phraseness module builds on the authors' earlier UOKG (same group), but that prior work is cited for the noun-phrase and related-document intuition, not as a uniqueness theorem, so the self-citation is not load-bearing circularity. The concrete circular step is in the evaluation loop. Section D.1 and Table 6 tune α and β on 'the KP20K dataset'; the chosen α=β=0.8 row reports present F1@10=19.4 and absent R@10=8.1, and Table 1 gives exactly the same values for ERU-KG-base on KP20K. So the KP20K column is the tuning run presented as a benchmark result. Table 5's 19,987-document 'test' count makes the split problem worse: this is the size of the original KP20K training split, not the standard 2,000-document test split, so the supervised CopyRNN baseline (trained on that split per Section B.2) is being scored in-sample as well. The headline 1.1-point F1@10 gap is a five-dataset average dominated by this 19,987-document row. Because the model is trained without keyphrase labels and its other four ground-truth datasets (SemEval, Inspec, NUS, Krapivin) plus the retrieval evaluations are held-out, the circularity is partial rather than total. The central claim retains independent content, but the KP20K component of the headline comparison reduces to a tuning-set result, so the score is 6.

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

The central claim rests on the reference-as-proxy assumption and on a stack of tuned hyperparameters and pruning thresholds. No new physical entities are invented, but the method introduces a learned term-importance model, a BM25 neighbor set, and a precomputed glossary, all of which carry their own assumptions. The ledger shows the paper is more of an engineered pipeline with tuned components than a parameter-free derivation.

free parameters (5)
  • alpha (interpolation weight for related-document term importances) = 0.8
    Set using the KP20K sensitivity analysis in Table 6 and Section D.1, then used in the main reported results including KP20K.
  • beta (interpolation weight for related-document phraseness) = 0.8
    Tuned jointly with alpha on KP20K in Table 6; it controls the balance between the input document and BM25-related documents in Eq. 7.
  • gamma (length penalty) = -0.25
    Set in Section B.1; controls preference for longer versus shorter candidates in Eq. 3 and 4, with no sensitivity analysis reported.
  • lambda (phraseness weight in product-of-experts) = 1.5
    Set in Section B.1; controls the relative importance of phraseness versus informativeness in Eq. 11, with no sensitivity analysis reported.
  • Pruning thresholds: top-10 candidates per related document, top-100 from CN(x), glossary frequency at least 3… = 10/100/3/100
    Hand-chosen in Section 2.2 to balance speed and accuracy; the ablation in Table 4 only varies |N(x)|, so the other thresholds are not swept.
assumptions (5)
  • domain assumption References (queries, citation contexts, titles) encode community perception of key concepts and are valid proxies for keyphrase informativeness.
    Introduced in Section 1 and operationalized in Section 2.1.2 where reference-document triplets are used as SPLADE training data. If this proxy is weak, the learned term importances will not align with human keyphrases.
  • domain assumption Noun phrases extracted by the specified NLTK grammar cover the true keyphrase space.
    Section 2.2 assumes keyphrases are noun phrases, citing Chuang et al. and Do et al.; the grammar in Section B.1 defines the candidate space.
  • domain assumption SPLADE max-pooled MLM logits provide term-level importances that transfer from retrieval relevance to keyphrase informativeness.
    Eq. 1 and Eq. 3 define informativeness as an average of SPLADE importances; the paper provides no direct term-level validation that these weights correlate with keyphrase annotations.
  • domain assumption BM25-retrieved related documents are a sufficient source of absent keyphrases.
    Eq. 6 and Algorithm 1 source all absent candidates from N(x); if related documents rarely contain true absent keyphrases, absent-keyphrase recall would collapse.
  • ad hoc to paper Product-of-experts combination with a position penalty is a valid ranking model for keyphrases.
    Eq. 11 and Eq. 12 are heuristic combination rules; the paper adopts them from prior work but does not test alternative ranking functions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ERU-KG: Efficient Reference-aligned Unsupervised Keyphrase Generation." pith.science (2026). https://pith.science/paper/7J7WLF36

@misc{pith2026250524219,
  author       = {Pith},
  title        = {Pith review of: ERU-KG: Efficient Reference-aligned Unsupervised Keyphrase Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7J7WLF36}},
  note         = {Machine review of arXiv:2505.24219}
}
read the original abstract

Unsupervised keyphrase prediction has gained growing interest in recent years. However, existing methods typically rely on heuristically defined importance scores, which may lead to inaccurate informativeness estimation. In addition, they lack consideration for time efficiency. To solve these problems, we propose ERU-KG, an unsupervised keyphrase generation (UKG) model that consists of an informativeness and a phraseness module. The former estimates the relevance of keyphrase candidates, while the latter generate those candidates. The informativeness module innovates by learning to model informativeness through references (e.g., queries, citation contexts, and titles) and at the term-level, thereby 1) capturing how the key concepts of documents are perceived in different contexts and 2) estimating informativeness of phrases more efficiently by aggregating term informativeness, removing the need for explicit modeling of the candidates. ERU-KG demonstrates its effectiveness on keyphrase generation benchmarks by outperforming unsupervised baselines and achieving on average 89\% of the performance of a supervised model for top 10 predictions. Additionally, to highlight its practical utility, we evaluate the model on text retrieval tasks and show that keyphrases generated by ERU-KG are effective when employed as query and document expansions. Furthermore, inference speed tests reveal that ERU-KG is the fastest among baselines of similar model sizes. Finally, our proposed model can switch between keyphrase generation and extraction by adjusting hyperparameters, catering to diverse application requirements.

Figures

Figures reproduced from arXiv: 2505.24219 by the authors.

Figure 1
Figure 1. An example of the different type of references. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of ERU-KG. Further details of the inference process are provided in Algorithm 1 generating those candidates. 2.1 Informativeness Module The informativeness module is responsible for rank￾ing candidate phrases. As mentioned above, it in￾corporates our key ideas to addressing the chal￾lenges of accurate and efficient informativeness estimation: modeling informativeness through ref￾erences and at the term-leve… view at source ↗
Figure 3
Figure 3. The performance change (in percentage) when [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

69 extracted references · 37 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]

    Nasreen Abdul-Jaleel, James Allan, W Bruce Croft, Fernando Diaz, Leah Larkey, Xiaoyan Li, Mark D Smucker, and Courtney Wade. 2004. Umass at trec 2004: Novelty and hard. Computer Science Department Faculty Publication Series, page 189

    2004

  4. [4]

    Yang Bai, Xiaoguang Li, Gang Wang, Chaoliang Zhang, Lifeng Shang, Jun Xu, Zhaowei Wang, Fangshan Wang, and Qun Liu. 2020. Sparterm: Learning term-based sparse representation for fast text retrieval. arXiv preprint arXiv:2010.00768

    arXiv 2020

  5. [5]

    Kamil Bennani-Smires, Claudiu Musat, Andreea Hossmann, Michael Baeriswyl, and Martin Jaggi. 2018. https://doi.org/10.18653/v1/K18-1022 Simple unsupervised keyphrase extraction using sentence embeddings . In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 221--229, Brussels, Belgium. Association for Computational Linguistics

    work page doi:10.18653/v1/k18-1022 2018

  6. [6]

    Steven Bird and Edward Loper. 2004. https://aclanthology.org/P04-3031/ NLTK : The natural language toolkit . In Proceedings of the ACL Interactive Poster and Demonstration Sessions , pages 214--217, Barcelona, Spain. Association for Computational Linguistics

    2004

  7. [7]

    Vera Boteva, Demian Gholipour, Artem Sokolov, and Stefan Riezler. 2016. A full-text learning to rank dataset for medical information retrieval. In Advances in Information Retrieval: 38th European Conference on IR Research, ECIR 2016, Padua, Italy, March 20--23, 2016. Proceedings 38, pages 716--722. Springer

    work page 2016

  8. [8]

    Florian Boudin. 2016. https://aclanthology.org/C16-2015/ pke: an open source python-based keyphrase extraction toolkit . In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: System Demonstrations , pages 69--73, Osaka, Japan. The COLING 2016 Organizing Committee

    work page 2016

Show all 69 references
  1. [9]

    Florian Boudin. 2018. https://doi.org/10.18653/v1/N18-2105 Unsupervised keyphrase extraction with multipartite graphs . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 ...

    2018 doi

  2. [10]

    Florian Boudin. 2021. Acm-cr: A manually annotated test collection for citation recommendation. In 2021 ACM/IEEE Joint Conference on Digital Libraries (JCDL), pages 280--281. IEEE

    2021

  3. [11]

    Florian Boudin and Akiko Aizawa. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.33 Unsupervised domain adaptation for keyphrase generation using citation contexts . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 598--614, Miami, Florida,...

    2024 doi

  4. [12]

    Florian Boudin, Ygor Gallina, and Akiko Aizawa. 2020. https://doi.org/10.18653/v1/2020.acl-main.105 Keyphrase generation for scientific document retrieval . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1118--1126, Online. As...

    2020 doi

  5. [13]

    Adrien Bougouin, Florian Boudin, and B \'e atrice Daille. 2013. Topicrank: Graph-based topic ranking for keyphrase extraction. In International joint conference on natural language processing (IJCNLP), pages 543--551

    2013

  6. [14]

    Ricardo Campos, V \' tor Mangaravite, Arian Pasquali, Al \' pio M \'a rio Jorge, C \'e lia Nunes, and Adam Jatowt. 2018. Yake! collection-independent automatic keyword extractor. In Advances in Information Retrieval: 40th European Conference on IR Research, ECIR 2018, Grenoble...

    2018

  7. [15]

    Guihong Cao, Jian-Yun Nie, Jianfeng Gao, and Stephen Robertson. 2008. Selecting good expansion terms for pseudo-relevance feedback. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval, pages 243--250

    2008

  8. [16]

    Cornelia Caragea, Florin Adrian Bulgarov, Andreea Godea, and Sujatha Das Gollapalli. 2014. https://doi.org/10.3115/v1/D14-1150 Citation-enhanced keyphrase extraction from research papers: A supervised approach . In Proceedings of the 2014 Conference on Empirical Methods in Nat...

    2014 doi

  9. [17]

    Hou Pong Chan, Wang Chen, Lu Wang, and Irwin King. 2019. https://doi.org/10.18653/v1/P19-1208 Neural keyphrase generation via reinforcement learning with adaptive rewards . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2163--...

    2019 doi

  10. [18]

    Wang Chen, Yifan Gao, Jiani Zhang, Irwin King, and Michael R Lyu. 2019. -guided encoding for keyphrase generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6268--6275

    2019

  11. [19]

    without the clutter of unimportant words

    Jason Chuang, Christopher D Manning, and Jeffrey Heer. 2012. “without the clutter of unimportant words” descriptive keyphrases for text visualization. ACM Transactions on Computer-Human Interaction (TOCHI), 19(3):1--29

    2012

  12. [20]

    Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel Weld. 2020. https://doi.org/10.18653/v1/2020.acl-main.207 SPECTER : Document-level representation learning using citation-informed transformers . In Proceedings of the 58th Annual Meeting of the Association for C...

    2020 doi

  13. [21]

    Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. arXiv preprint arXiv:1910.10687

    2019 arXiv

  14. [22]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  15. [23]

    Haoran Ding and Xiao Luo. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.146 A ttention R ank: Unsupervised keyphrase extraction using self and cross attentions . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1919--1928, On...

    2021 doi

  16. [24]

    Lam Do, Pritom Saha Akash, and Kevin Chen-Chuan Chang. 2023. https://doi.org/10.18653/v1/2023.acl-long.592 Unsupervised open-domain keyphrase generation . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages ...

    2023 doi

  17. [25]

    Joel L Fagan. 2017. Automatic p hrase indexing for document retrieval: an examination of syntactic and non-syntactic methods. In ACM SIGIR Forum, volume 51, pages 51--61. ACM New York, NY, USA

    2017

  18. [26]

    Corina Florescu and Cornelia Caragea. 2017. https://doi.org/10.18653/v1/P17-1102 P osition R ank: An unsupervised approach to keyphrase extraction from scholarly documents . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

    2017 doi

  19. [27]

    Thibault Formal, Carlos Lassance, Benjamin Piwowarski, and Stéphane Clinchant. 2021 a . https://doi.org/10.48550/ARXIV.2109.10086 Splade v2: Sparse lexical and expansion model for information retrieval . arXiv preprint

    2021 doi arXiv

  20. [28]

    Thibault Formal, Benjamin Piwowarski, and St\' e phane Clinchant. 2021 b . https://doi.org/10.1145/3404835.3463098 Splade: Sparse lexical and expansion model for first stage ranking . In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in ...

    2021

  21. [29]

    Ygor Gallina, Florian Boudin, and B \'e atrice Daille. 2020. Large-scale evaluation of keyphrase extraction models. In Proceedings of the ACM/IEEE Joint Conference on Digital Libraries in 2020, pages 271--278

    2020

  22. [30]

    Krishna Garg, Jishnu Ray Chowdhury, and Cornelia Caragea. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.427 Keyphrase generation beyond the boundaries of title and abstract . In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5809--5821, Ab...

    2022 doi

  23. [31]

    Sujatha Das Gollapalli and Cornelia Caragea. 2014. Extracting keyphrases from research papers using citation networks. In Proceedings of the AAAI conference on artificial intelligence, volume 28

    2014

  24. [32]

    Carl Gutwin, Gordon Paynter, Ian Witten, Craig Nevill-Manning, and Eibe Frank. 1999. Improving browsing in digital libraries with keyphrase indexes. Decision Support Systems, 27(1-2):81--104

    1999

  25. [33]

    Geoffrey E Hinton. 2002. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8):1771--1800

    2002

  26. [34]

    Anette Hulth. 2003. https://aclanthology.org/W03-1028/ Improved automatic keyword extraction given more linguistic knowledge . In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, pages 216--223

    2003

  27. [35]

    Steve Jones and Mark S Staveley. 1999. Phrasier: a system for interactive document retrieval using keyphrases. In Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval, pages 160--167

    1999

  28. [36]

    Byungha Kang and Youhyun Shin. 2024. https://aclanthology.org/2024.lrec-main.775/ Improving low-resource keyphrase generation through unsupervised title phrase generation . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resourc...

    2024

  29. [37]

    Su Nam Kim, Olena Medelyan, Min-Yen Kan, and Timothy Baldwin. 2010. https://aclanthology.org/S10-1004/ S em E val-2010 task 5 : Automatic keyphrase extraction from scientific articles . In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 21--26, Upps...

    2010

  30. [38]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  31. [39]

    Aobo Kong, Shiwan Zhao, Hao Chen, Qicheng Li, Yong Qin, Ruiqi Sun, and Xiaoyan Bai. 2023. https://doi.org/10.18653/v1/2023.acl-long.545 P rompt R ank: Unsupervised keyphrase extraction using prompt . In Proceedings of the 61st Annual Meeting of the Association for Computationa...

    2023 doi

  32. [40]

    Mikalai Krapivin, Aliaksandr Autaeu, Maurizio Marchese, et al. 2009. Large dataset for keyphrases extraction

    2009

  33. [41]

    Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep, and Rodrigo Nogueira. 2021. Pyserini: A python toolkit for reproducible information retrieval research with sparse and dense representations. In Proceedings of the 44th International ACM SIGIR Conference ...

    2021

  34. [42]

    Sean MacAvaney, Franco Maria Nardini, Raffaele Perego, Nicola Tonellotto, Nazli Goharian, and Ophir Frieder. 2020. Expansion via prediction of importance with contextualization. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Inform...

    2020

  35. [43]

    Rui Meng, Sanqiang Zhao, Shuguang Han, Daqing He, Peter Brusilovsky, and Yu Chi. 2017. https://doi.org/10.18653/v1/P17-1054 Deep keyphrase generation . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 582...

    2017 doi

  36. [44]

    Rada Mihalcea and Paul Tarau. 2004. https://aclanthology.org/W04-3252/ T ext R ank: Bringing order into text . In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 404--411, Barcelona, Spain. Association for Computational Linguistics

    2004

  37. [45]

    Thuy Dung Nguyen and Min-Yen Kan. 2007. Keyphrase extraction in scientific publications. In International conference on Asian digital libraries, pages 317--326. Springer

    2007

  38. [46]

    Rodrigo Nogueira, Jimmy Lin, and AI Epistemic. 2019 a . From doc2query to doctttttquery. Online preprint, 6(2)

    2019

  39. [47]

    Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019 b . Document expansion by query prediction. arXiv preprint arXiv:1904.08375

    2019 arXiv

  40. [48]

    Matteo Pagliardini, Prakhar Gupta, and Martin Jaggi. 2018. https://doi.org/10.18653/v1/N18-1049 Unsupervised learning of sentence embeddings using compositional n-gram features . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Comput...

    2018 doi

  41. [49]

    Biswajit Paria, Chih-Kuan Yeh, Ian EH Yen, Ning Xu, Pradeep Ravikumar, and Barnab \'a s P \'o czos. 2020. Minimizing flops to learn efficient sparse representations. arXiv preprint arXiv:2004.05665

    2020 arXiv

  42. [50]

    Martin F Porter. 1980. An algorithm for suffix stripping. Program, 14(3):130--137

    1980

  43. [51]

    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

  44. [52]

    Tarek Saier, Johan Krause, and Michael F\" a rber. 2023. https://doi.org/10.1109/JCDL57899.2023.00020 unarXive 2022: All arXiv Publications Pre-Processed for NLP, Including Structured Full-Text and Citation Network . In 2023 ACM/IEEE Joint Conference on Digital Libraries (JCDL...

    2023

  45. [53]

    V Sanh. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108

    2019 arXiv

  46. [54]

    Xianjie Shen, Yinghan Wang, Rui Meng, and Jingbo Shang. 2022. Unsupervised deep keyphrase generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11303--11311

    2022

  47. [55]

    Amanpreet Singh, Mike D ' Arcy, Arman Cohan, Doug Downey, and Sergey Feldman. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.338 S ci R ep E val: A multi-format benchmark for scientific document representations . In Proceedings of the 2023 Conference on Empirical Methods in...

    2023 doi

  48. [56]

    Mingyang Song, Huafeng Liu, Yi Feng, and Liping Jing. 2023. https://doi.org/10.18653/v1/2023.findings-acl.66 Improving embedding-based unsupervised keyphrase extraction by incorporating structural information . In Findings of the Association for Computational Linguistics: ACL ...

    2023 doi

  49. [57]

    Karen Sparck Jones. 1972. A statistical interpretation of term specificity and its application in retrieval. Journal of documentation, 28(1):11--21

    1972

  50. [58]

    Yi Sun, Hangping Qiu, Yu Zheng, Zhongwei Wang, and Chaoran Zhang. 2020. Sifrank: a new baseline for unsupervised keyphrase extraction based on pre-trained language model. IEEE Access, 8:10896--10906

    2020

  51. [59]

    Nandan Thakur, Nils Reimers, Andreas R \"u ckl \'e , Abhishek Srivastava, and Iryna Gurevych. 2021. https://openreview.net/forum?id=wCu6T5xFjeJ BEIR : A heterogeneous benchmark for zero-shot evaluation of information retrieval models . In Thirty-fifth Conference on Neural Info...

    2021

  52. [60]

    Ellen Voorhees, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, William R Hersh, Kyle Lo, Kirk Roberts, Ian Soboroff, and Lucy Lu Wang. 2021. Trec-covid: constructing a pandemic information retrieval test collection. In ACM SIGIR Forum, volume 54, pages 1--12. ACM New York, NY, USA

    2021

  53. [61]

    David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.609 Fact or fiction: Verifying scientific claims . In Proceedings of the 2020 Conference on Empirical Methods in Na...

    2020 doi

  54. [62]

    Xiaojun Wan and Jianguo Xiao. 2008. Single document keyphrase extraction using neighborhood knowledge. In AAAI, volume 8, pages 855--860

    2008

  55. [63]

    Jianyou Andre Wang, Kaicheng Wang, Xiaoyue Wang, Prudhviraj Naidu, Leon Bergen, and Ramamohan Paturi. 2024. Scientific document retrieval using multi-level aspect-based queries. Advances in Neural Information Processing Systems, 36

    2024

  56. [64]

    Ian H Witten, David Bainbridge, and David M Nichols. 2009. How to build a digital library. Morgan Kaufmann

    2009

  57. [65]

    Di Wu, Wasi Uddin Ahmad, and Kai-Wei Chang. 2022 a . Pre-trained language models for keyphrase generation: A thorough empirical study. arXiv preprint arXiv:2212.10233

    2022 arXiv

  58. [66]

    Huanqin Wu, Baijiaxin Ma, Wei Liu, Tao Chen, and Dan Nie. 2022 b . Fast and constrained absent keyphrase generation by prompt-based learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11495--11503

    2022

  59. [67]

    Jiacheng Ye, Ruijian Cai, Tao Gui, and Qi Zhang. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.213 Heterogeneous graph neural networks for keyphrase generation . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2705--2715, On...

    2021 doi

  60. [68]

    Chengxiang Zhai. 1997. Fast statistical parsing of noun phrases for document indexing. arXiv preprint cmp-lg/9702009

    1997 arXiv

  61. [69]

    Linhan Zhang, Qian Chen, Wen Wang, Chong Deng, ShiLiang Zhang, Bing Li, Wei Wang, and Xin Cao. 2022. https://doi.org/10.18653/v1/2022.findings-acl.34 MDER ank: A masked document embedding rank approach for unsupervised keyphrase extraction . In Findings of the Association for ...

    2022 doi

Pith tools

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