Pith. sign in

REVIEW 5 major objections 5 minor 83 references

Limited Generalizability in Argument Mining: State-Of-The-Art Models Learn Datasets, Not Arguments

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

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

Pith's one-line read State-of-the-art argument-mining models learn dataset-specific lexical shortcuts, not a generalizable concept of arguments, and their performance collapses when applied to unseen datasets.

desk verdict Solid large-scale evidence that argument mining models transfer poorly; the shortcut-learning conclusion, however, outruns the manipulation design. read the letter →

arxiv 2505.22137 v1 pith:WJBBNKAG submitted 2025-05-28 cs.CL cs.AIcs.LG

Marc Feger , Katarina Boland , Stefan Dietze This is my paper
classification cs.CLcs.AIcs.LG
keywords argumentmininggeneralizationshortcutlearningtransformerslexicalshortcutscross-datasetevaluationbenchmarkre-evaluationcontrastivepre-training
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the strong benchmark results of transformer-based argument-mining models rest on lexical shortcuts tied to content words rather than on a generalizable concept of what makes a sentence an argument. In a large-scale re-evaluation of four transformers across 17 English sentence-level datasets, 97% of generalization experiments fell below the average benchmark score (0.79 macro F1), and 62% fell below 0.65, showing that familiar training data inflates apparent competence. Controlled manipulation that strips sentences of stop words, function words, discourse markers, and punctuation left standard models (BERT, RoBERTa, DistilBERT) nearly unchanged, while a contrastively pre-trained model (WRAP) showed the largest drop, which the paper reads as evidence that standard models were not using structural argument signals. The paper also shows that joint training on multiple benchmarks improves generalization and that WRAP's argument-oriented pre-training makes it more robust. The finding matters because argument mining should identify the logical scaffolding of discourse independent of topic; if models merely memorize topic cues, current benchmarks overstate real-world usefulness.

What carries the argument

The key mechanism is a controlled input-manipulation experiment labelled 'Disrupting Argument Signals': the authors strip each sentence of stop words, function words, discourse markers, and punctuation using a syntax parser, leaving a 'lexical skeleton' of content words and halving the sentence length. The experimental design combines this ablation with two transfer protocols—a 17×17 pairwise matrix in which a model is trained on one dataset and tested on every other, and a joint-training protocol in which a model is trained on all datasets except the target—so that in-distribution benchmarks and out-of-distribution generalization can be compared under the same ablation. The claim-bearing observation is that standard transformers' macro F1 barely moves after the ablation, whereas the contrastively pre-trained WRAP shows the largest drop; since the ablation is meant to suppress the structural-discourse cues that distinguish arguments from non-arguments, the non-response of the standard models is read as evidence that their benchmark success was carried by topic-related content words.

What would settle it

Run the complementary ablation: train a transformer on inputs from which all content words are removed, keeping only stop/function words, discourse markers, and punctuation, and measure its benchmark macro F1. If such a model retains most of its full-input accuracy on held-out datasets, then argument structure is sufficiently signalled by function words alone and the paper's claim that models rely on content-word shortcuts is falsified; if it collapses, the paper's interpretation is supported.

Watch

Extended reading notes

Core claim

The central claim is that state-of-the-art argument-mining models 'tend to rely on lexical shortcuts tied to content words,' so that apparent progress on benchmarks is often driven by dataset-specific cues rather than true task alignment. On three standard transformers (BERT, RoBERTa, DistilBERT) and one contrastively pre-trained model (WRAP), evaluated across 17 datasets under pairwise transfer, 97% of generalization outcomes fell below the mean benchmark result (M = 0.79), with 62% under 0.65 and 8% below 0.5 macro F1. When stop/function words, discourse markers, and punctuation were removed, BERT and DistilBERT showed nearly no change (Δ ≤ 0.02), RoBERTa none at all, while WRAP dropped by 0.05, which the paper treats as evidence that the standard models are not using argument-structure signals. Training jointly on all-but-one dataset improved average macro F1 to at least 0.63, and WRAP outperformed the others in most settings, suggesting that combining diverse benchmark data and task-specific pre-training mitigates, but does not eliminate, shortcut reliance. The paper also documents that datasets are largely disjoint in their semantic content (7–36% vocabulary Jaccard similarity) while their sentence-level structural features align closely (Spearman's ρ ≥ 0.68), which is what makes lexical shortcuts available and consequential.

Load-bearing premise

The paper's reading of the manipulation results depends on the assumption that the linguistic cues that let a sentence be recognised as an argument are carried mainly by stop words, function words, discourse markers, and punctuation; if argumentative structure can be conveyed solely by content words or other surface features, then the post-ablation performance drops cannot be cleanly attributed to shortcut reliance.

Editorial extensions

If this is right

  • Individual benchmark scores for argument mining overstate real capability; models that top one dataset typically fall far below that level on other datasets.
  • Training on a combined pool of argument-mining datasets raises average out-of-distribution macro F1 above what pairwise transfer achieves, suggesting that joint benchmarking is a practical route to more robust models.
  • Task-specific pre-training, such as WRAP's contrastive learning of inference/information structure, yields more consistent generalization and lower susceptibility to lexical shortcuts than standard language-model pre-training.
  • Divergent argument definitions and annotation practices across datasets are a real limiting factor independent of model choice, as evidenced by models consistently finding non-argument sentences easier to identify than argument sentences.

Reading between the lines

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

  • The ablation logic assumes that argument structure lives in function words, discourse markers, and punctuation; if argumentativeness can be signalled purely through content-word combinations (e.g., 'X should Y, because Z'), then the post-ablation drop understates shortcut reliance, and the paper's manipulation is a lower bound on lexical-shortcut effects.
  • The paper does not test topic-controlled transfer within a single dataset; a natural extension is to hold topics fixed across train and test while varying argumentative labelling, which would isolate topic-memorization from genuine argument-structure learning.
  • The finding that content-word vocabularies of argument and non-argument sentences overlap below 19% suggests a concrete remedy: adversarial training against topic discriminators, or fine-grained delexicalization, could force models toward structure and improve cross-dataset transfer; this is implied by the paper's results but not tested by it.
  • Because WRAP's advantage persists after joint training, the paper's results indirectly support the idea that embedding objectives that organise sentences by inference versus information structure may transfer to other discourse tasks such as stance detection or premise-conclusion extraction, though that application is outside the study.
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

5 major / 5 minor

Summary. The paper presents a large-scale evaluation of four transformer models (BERT, RoBERTa, DistilBERT, WRAP) on 17 sentence-level argument mining datasets, using a 17×17 pairwise transfer matrix, a leave-one-dataset-out joint-training experiment, and a controlled input manipulation that removes stop/function words, discourse markers, and punctuation. The authors report that models achieve strong within-benchmark results but degrade sharply in cross-dataset generalization, that post-manipulation performance stability indicates reliance on content-word lexical shortcuts, that WRAP (a contrastively pre-trained transformer) generalizes best, and that joint benchmark training improves robustness. They conclude that state-of-the-art models 'learn datasets, not arguments.'

Significance. The study is extensive and mostly well executed: it surveys 52 argument mining datasets, selects 17 under explicit criteria, runs repeated trials with ANOVA and corrected t-tests, and releases code and data. If the conclusions are valid, the work would be an important cautionary result for the argument mining community, providing a reusable benchmark suite and statistical methodology for generalization testing. The central contribution is the 17×17 transfer matrix, which is a valuable resource in itself. However, the shortcut-learning conclusion depends on an unvalidated assumption about the linguistic carriers of argument structure, and the 'state-of-the-art' comparison is not properly sourced. These issues are load-bearing and require substantive revision.

major comments (5)
  1. [Section 4, 'Disrupting Argument Signals'] The manipulation test assumes that removing stop/function words, discourse markers, and punctuation 'suppresses the linguistic cues that, in theory, enable the distinction between the elements that constitute an argument and those that do not.' This assumption is never validated. No control condition is provided—for instance, removing content words while retaining function words, or testing whether human annotators can still identify arguments from the content-word skeleton. The examples in Table 3 ('increase routes people transport easily', 'governments spend money improving roads highways') still appear to convey proposal/claim semantics, and the paper's own Limitations section acknowledges that implicit arguments 'are likely to depend on topical and content cues,' which undercuts the manipulation's premise. Because the post-manipulation stability of BERT, RoBERTa, and DistilBERT is the main direct evidence for shortcut reliance, the claim is not yet supported. Please add a validation study (e.g., human annotation of manipulated forms) or an alternative ablation that isolates content words versus function words.
  2. [Section 4, 'Benchmarking and Generalization'] The transfer evaluation protocol tunes hyperparameters on the target dataset's development set before reporting test performance. This leaks target-label information into model selection, which is inappropriate for an out-of-distribution generalization claim. A model intended to generalize to an unseen dataset would not have access to the target's development labels; tuning on target dev can inflate the reported generalization scores and makes the 'unseen datasets' framing in the abstract and Section 5 inaccurate. Please either select hyperparameters on the source dev set, use a fixed hyperparameter configuration, or explicitly justify why target dev tuning does not affect the generalization conclusions.
  3. [Tables 4 and 5] The 'state-of-the-art (SOTA)' baseline is not sourced. The captions say to 'compare diagonal of Figure 1,' which suggests the SOTA values are the authors' own within-dataset fine-tuning results, not published state-of-the-art numbers. If the diagonal is meant as the reference, this should be stated plainly (e.g., 'in-domain fine-tuning baseline') rather than labeled SOTA; the current label implies external state-of-the-art results that are never cited. This matters because the claim in Section 5 that all models 'underperformed relative to the state-of-the-art baselines' is a key interpretative point, and the reader cannot verify what is being compared. Please clarify the provenance of the SOTA column and ensure the comparison is meaningful and reproducible.
  4. [Section 5 and Figure 1] The conclusion that WRAP generalizes best is confounded by domain overlap. WRAP's pre-training uses Twitter text (Feger and Dietze, 2024a), and TACO, one of the 17 benchmark datasets, is a Twitter dataset introduced by the same research group (Feger and Dietze, 2024b). It is therefore expected that WRAP will perform strongly on TACO, and Figure 1 indeed shows exceptional WRAP performance on TACO. This overlap alone could explain much of WRAP's overall margin. Please report the pairwise results with TACO excluded, or include a domain-matched control to separate genuine argument-level generalization from domain familiarity.
  5. [Section 5, Q2 statistics] The headline statistic—'97% of generalization experiments fall below the mean benchmark result (M = 0.79), with 62% scoring under 0.65, and 8% dropping below 0.5 macro F1'—is presented without the underlying distribution or the exact denominator. The 'mean benchmark result' appears to be an average of the diagonal entries of Figure 1, but it is unclear whether this average is unweighted across the 17 datasets and four models, or computed in some other way. Also, because some off-diagonal entries exceed 0.75 (e.g., SCIARK-to-ABSTRCT, UKP-to-ABSTRCT, and WRAP-to-TACO), the 97% claim requires a precise definition of 'generalization experiments' and the benchmark mean. Please provide the full distribution, the number of experiments, and a per-model breakdown to make the statistic interpretable and falsifiable.
minor comments (5)
  1. [Throughout] There are numerous typographical and formatting issues, including 'ANOV A' (should be 'ANOVA'), 'M = 0 .79' and similar spacing artifacts, 'V ACC' with an embedded space, and inconsistent use of mathematical symbols (e.g., '¯∆max'). A careful copyedit is needed.
  2. [Section 4, 'Sampling'] The sentence 'selecting 850 instances per label, corresponding to 1,700 samples per dataset and 28,900 in total' is clear, but the rationale for capping all datasets to 850 per label (rather than using the full available data) should be stated in the text, not only implied by the 'Sufficiency' criterion in Section 2.2.
  3. [Section 2.1] The reference to the repository is given only as a footnote marker (1) without a visible URL; if the repository is publicly available, please include the URL in the camera-ready version.
  4. [References] The reference list includes multiple entries that are cited in the text but some formatting inconsistencies (e.g., 'Feger and Dietze, 2024a' and 'Feger and Dietze, 2024b' appear as separate entries, yet the text occasionally cites 'Feger and Dietze, 2024a/b' without disambiguation). Please standardize citations.
  5. [Limitations] The Limitations section appropriately acknowledges that implicit arguments were not separated and that this may affect interpretation; this is commendable, but the statement directly interacts with the manipulation assumption raised in my first major comment, and the paper should connect these two points explicitly.

Circularity Check

1 steps flagged · score 4.0 of 10

The central limited-generalization result is independently grounded in the 17x17 transfer matrix, but the highlighted sub-result that WRAP generalizes best is partly a same-author, same-domain artifact: WRAP's pretraining and TACO both come from the same authors' Twitter argument work, so WRAP's strong TACO scores are not a clean out-of-distribution prediction.

  1. self citation load bearing [Section 5 (Q2); Figure 1 caption; Section 6]
    "Similarly, TACO (inference-information, Twitter debate) consistently exceeds 0.8 across a vast range of definitions and genres with WRAP. ... WRAP excels in generalizing to TACO, as seen on the right."

    TACO is an author-created Twitter dataset (Feger and Dietze, 2024b), and WRAP is the authors' own model whose pre-training 'leverag[es] contrasts of inference and information signals' (Section 1) — the same inference/information contrast on Twitter that defines TACO. The paper singles out TACO as the clearest success of WRAP's generalization and later concludes WRAP 'shows more resilience, likely due to its pre-training for argument generalization.' Because WRAP's pretraining domain and TACO's construction are both from the same authors' Twitter argument-mining pipeline, the TACO scores are at least partially in-domain rather than independent cross-dataset evidence.

full rationale

The paper's central finding — that all four transformers drop markedly when transferred across the 17 benchmarks, with 97% of generalization experiments below the mean benchmark result — is a direct, self-contained measurement that does not reduce to any fitted parameter or definition. The controlled manipulation (removing stop/function words, discourse markers, and punctuation) is an interpretive assumption: the paper assumes those words carry the argument-distinguishing cues and then reads stable post-removal performance as evidence of content-word shortcut use. That is a validity threat, and the Limitations section honestly concedes that implicit arguments may depend on topical and content cues, but it is not circular in the paper's derivation chain because the assumption is supported by external linguistic work (Knott and Dale, 1994; Daxenberger et al., 2017; Opitz and Frank, 2019; Thorn Jakobsen et al., 2021). The main circularity-adjacent issue is the WRAP/TACO pairing: TACO is a same-author Twitter benchmark, WRAP is a same-author Twitter-pretrained model whose pretraining objective mirrors TACO's inference-information definition, and the paper repeatedly highlights WRAP's TACO scores as evidence of superior generalization. This raises the score to 4 because it weakens a prominent sub-claim, but the core conclusion does not depend on TACO alone and is independently supported by the broad transfer matrix and joint-training results.

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

The paper introduces no new free parameters or invented entities. Its main analytical assumption is that removing function words leaves only topic-related content, which is used to distinguish shortcut reliance from true argument understanding. The dataset selection and evaluation metric are additional domain assumptions rather than mathematically derived choices.

assumptions (3)
  • domain assumption Removing stop and function words, discourse markers, and punctuation suppresses the linguistic cues that signal argument structure.
    Invoked in Section 4 ('Disrupting Argument Signals') to interpret post-manipulation performance drops as evidence about whether models rely on argument signals vs. content words.
  • domain assumption Macro F1 is the appropriate evaluation metric for binary argument detection across imbalanced datasets.
    Used throughout Sections 4-5 to compare models; the choice favors balanced performance but is not the only possible metric.
  • domain assumption The 17 selected datasets are representative of the argument-mining benchmark landscape.
    Dataset selection is based on sentential, binary-label, and reproducibility criteria (Section 2.2), but the final set may not reflect the full diversity of argument types and genres.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Limited Generalizability in Argument Mining: State-Of-The-Art Models Learn Datasets, Not Arguments." pith.science (2026). https://pith.science/paper/WJBBNKAG

@misc{pith2026250522137,
  author       = {Pith},
  title        = {Pith review of: Limited Generalizability in Argument Mining: State-Of-The-Art Models Learn Datasets, Not Arguments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJBBNKAG}},
  note         = {Machine review of arXiv:2505.22137}
}
read the original abstract

Identifying arguments is a necessary prerequisite for various tasks in automated discourse analysis, particularly within contexts such as political debates, online discussions, and scientific reasoning. In addition to theoretical advances in understanding the constitution of arguments, a significant body of research has emerged around practical argument mining, supported by a growing number of publicly available datasets. On these benchmarks, BERT-like transformers have consistently performed best, reinforcing the belief that such models are broadly applicable across diverse contexts of debate. This study offers the first large-scale re-evaluation of such state-of-the-art models, with a specific focus on their ability to generalize in identifying arguments. We evaluate four transformers, three standard and one enhanced with contrastive pre-training for better generalization, on 17 English sentence-level datasets as most relevant to the task. Our findings show that, to varying degrees, these models tend to rely on lexical shortcuts tied to content words, suggesting that apparent progress may often be driven by dataset-specific cues rather than true task alignment. While the models achieve strong results on familiar benchmarks, their performance drops markedly when applied to unseen datasets. Nonetheless, incorporating both task-specific pre-training and joint benchmark training proves effective in enhancing both robustness and generalization.

Figures

Figures reproduced from arXiv: 2505.22137 by the authors.

Figure 1
Figure 1. The best macro F1 scores from the benchmark [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. The correlations of the individual datasets (as [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. The word overlaps, measured by the Jac￾card similarity between the vocabularies of two datasets, show that the datasets (as well as the labels) are gen￾erally distinct from each other. The overlaps range between 3–36%, with an average of 19%. B.1 Two-Way Repeated Measures ANOVA We employ a two-way repeated measures ANOVA to evaluate the effects of sampling (factor 1) and model choice (factor 2) on the macro F1 (depe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

83 extracted references · 45 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]

    Ehud Aharoni, Anatoly Polnarov, Tamar Lavee, Daniel Hershcovich, Ran Levy, Ruty Rinott, Dan Gutfreund, and Noam Slonim. 2014. https://doi.org/10.3115/v1/W14-2109 A benchmark dataset for automatic detection of claims and evidence in the context of controversial topics . In Proceedings of the First Workshop on Argumentation Mining, pages 64--68, Baltimore, ...

    work page doi:10.3115/v1/w14-2109 2014

  4. [4]

    Yamen Ajjour, Johannes Kiesel, Benno Stein, and Martin Potthast. 2023. https://doi.org/10.18653/v1/2023.findings-eacl.104 Topic ontologies for arguments . In Findings of the Association for Computational Linguistics: EACL 2023, pages 1411--1427, Dubrovnik, Croatia. Association for Computational Linguistics

    work page doi:10.18653/v1/2023.findings-eacl.104 2023

  5. [5]

    Yamen Ajjour, Henning Wachsmuth, Johannes Kiesel, Martin Potthast, Matthias Hagen, and Benno Stein. 2019. https://doi.org/10.1007/978-3-030-30179-8_4 Data acquisition for argument search: The args.me corpus . In KI 2019: Advances in Artificial Intelligence, pages 48--59, Cham. Springer International Publishing

    doi:10.1007/978-3-030-30179-8_4 2019

  6. [6]

    Khalid Al-Khatib, Henning Wachsmuth, Matthias Hagen, Jonas K \"o hler, and Benno Stein. 2016 a . https://doi.org/10.18653/v1/N16-1165 Cross-domain mining of argumentative text through distant supervision . In Proceedings of the 2016 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies , pa...

    work page doi:10.18653/v1/n16-1165 2016

  7. [7]

    Khalid Al-Khatib, Henning Wachsmuth, Johannes Kiesel, Matthias Hagen, and Benno Stein. 2016 b . https://aclanthology.org/C16-1324/ A news editorial corpus for mining argumentation strategies . In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers , pages 3433--3443, Osaka, Japan. The COLING 2016 Or...

    2016

  8. [8]

    Alaa Alhamzeh, Romain Fonck, Erwan Versm \'e e, El \"o d Egyed-Zsigmond, Harald Kosch, and Lionel Brunie. 2022. https://doi.org/10.18653/v1/2022.finnlp-1.22 It`s time to reason: Annotating argumentation structures in financial earnings calls: The F in A rg dataset . In Proceedings of the Fourth Workshop on Financial Technology and Natural Language Process...

    work page doi:10.18653/v1/2022.finnlp-1.22 2022

Show all 83 references
  1. [9]

    Roy Bar-Haim, Indrajit Bhattacharya, Francesco Dinuzzo, Amrita Saha, and Noam Slonim. 2017. https://aclanthology.org/E17-1024/ Stance classification of context-dependent claims . In Proceedings of the 15th Conference of the E uropean Chapter of the Association for Computationa...

    2017

  2. [10]

    Or Biran and Owen Rambow. 2011. https://doi.org/10.1142/S1793351X11001328 Identifying justifications in written dialogues by classifying text as argumentative . International Journal of Semantic Computing, 05(04):363--381

    2011 doi

  3. [11]

    Filip Boltu z i \'c and Jan S najder. 2014. https://doi.org/10.3115/v1/W14-2107 Back up your stance: Recognizing arguments in online discussions . In Proceedings of the First Workshop on Argumentation Mining, pages 49--58, Baltimore, Maryland. Association for Computational Linguistics

    2014 doi

  4. [12]

    J \'e r \'e mie Cabessa, Hugo Hernault, and Umer Mushtaq. 2025. https://aclanthology.org/2025.coling-main.442/ Argument mining with fine-tuned large language models . In Proceedings of the 31st International Conference on Computational Linguistics, pages 6624--6635, Abu Dhabi,...

    2025

  5. [13]

    Elena Cabrio and Serena Villata. 2018. Five years of argument mining: a data-driven analysis. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18, page 5427–5433. AAAI Press

    2018

  6. [14]

    Liying Cheng, Lidong Bing, Ruidan He, Qian Yu, Yan Zhang, and Luo Si. 2022. https://doi.org/10.18653/v1/2022.acl-long.162 IAM : A comprehensive and large-scale dataset for integrated argument mining tasks . In Proceedings of the 60th Annual Meeting of the Association for Compu...

    2022 doi

  7. [15]

    Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. https://doi.org/10.18653/v1/W19-4828 What does BERT look at? an analysis of BERT `s attention . In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP...

    2019 doi

  8. [16]

    Johannes Daxenberger, Steffen Eger, Ivan Habernal, Christian Stab, and Iryna Gurevych. 2017. https://doi.org/10.18653/v1/D17-1218 What is the essence of a claim? cross-domain claim identification . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language ...

    2017 doi

  9. [17]

    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

  10. [18]

    Marc Feger and Stefan Dietze. 2024 a . https://doi.org/10.18653/v1/2024.findings-naacl.146 BERT weet`s TACO fiesta: Contrasting flavors on the path of inference and information-driven argument mining on T witter . In Findings of the Association for Computational Linguistics: N...

    2024 doi

  11. [19]

    Marc Feger and Stefan Dietze. 2024 b . https://aclanthology.org/2024.lrec-main.1349/ TACO -- T witter arguments from CO nversations . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), p...

    2024

  12. [20]

    Aris Fergadis, Dimitris Pappas, Antonia Karamolegkou, and Haris Papageorgiou. 2021. https://doi.org/10.18653/v1/2021.argmining-1.10 Argumentation mining in scientific literature for sustainable development . In Proceedings of the 8th Workshop on Argument Mining, pages 100--111...

    2021 doi

  13. [21]

    Beatriz Fisas, Francesco Ronzano, and Horacio Saggion. 2016. https://aclanthology.org/L16-1492/ A multi-layered annotated corpus of scientific papers . In Proceedings of the Tenth International Conference on Language Resources and Evaluation ( LREC `16) , pages 3081--3088, Por...

    2016

  14. [23]

    Michael Fromm, Evgeniy Faerman, Max Berrendorf, Siddharth Bhargava, Ruoxia Qi, Yao Zhang, Lukas Dennert, Sophia Selle, Yang Mao, and Thomas Seidl. 2021 b . https://doi.org/10.1609/aaai.v35i6.16607 Argument mining driven analysis of peer-reviews . Proceedings of the AAAI Confer...

    2021 doi

  15. [24]

    Wichmann

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. 2020. https://doi.org/10.1038/s42256-020-00257-z Shortcut learning in deep neural networks . Nature Machine Intelligence, 2(11):665--673

    2020 doi

  16. [25]

    Nancy Green. 2018. https://doi.org/10.18653/v1/W18-5213 Proposed method for annotation of scientific arguments in terms of semantic relations and argument schemes . In Proceedings of the 5th Workshop on Argument Mining, pages 105--110, Brussels, Belgium. Association for Comput...

    2018 doi

  17. [26]

    Giulia Grundler, Piera Santin, Andrea Galassi, Federico Galli, Francesco Godano, Francesca Lagioia, Elena Palmieri, Federico Ruggeri, Giovanni Sartor, and Paolo Torroni. 2022. https://aclanthology.org/2022.argmining-1.14/ Detecting arguments in CJEU decisions on fiscal state a...

    2022

  18. [27]

    Ivan Habernal, Daniel Faber, Nicola Recchia, Sebastian Bretthauer, Iryna Gurevych, Indra Spiecker genannt D\" o hmann, and Christoph Burchard. 2023. https://doi.org/10.1007/s10506-023-09361-y Mining legal arguments in court decisions . Artif. Intell. Law, 32(3):1–38

    2023 doi

  19. [28]

    Ivan Habernal and Iryna Gurevych. 2015. https://doi.org/10.18653/v1/D15-1255 Exploiting debate portals for semi-supervised argumentation mining in user-generated web discourse . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 21...

    2015 doi

  20. [29]

    Ivan Habernal and Iryna Gurevych. 2017. https://doi.org/10.1162/COLI_a_00276 Argumentation mining in user-generated web discourse . Computational Linguistics, 43(1):125--179

    2017 doi

  21. [30]

    Shohreh Haddadan, Elena Cabrio, and Serena Villata. 2019. https://doi.org/10.18653/v1/P19-1463 Yes, we can! mining arguments in 50 years of US presidential campaign debates . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4684...

    2019 doi

  22. [31]

    Marcus Hansen and Daniel Hershcovich. 2022. https://doi.org/10.18653/v1/2022.nlp4pi-1.5 A dataset of sustainable diet arguments on T witter . In Proceedings of the Second Workshop on NLP for Positive Impact (NLP4PI), pages 40--58, Abu Dhabi, United Arab Emirates (Hybrid). Asso...

    2022 doi

  23. [32]

    Annette Hautli-Janisz, Zlata Kikteva, Wassiliki Siskou, Kamila Gorska, Ray Becker, and Chris Reed. 2022. https://aclanthology.org/2022.lrec-1.352/ QT 30: A corpus of argument and conflict in broadcast debate . In Proceedings of the Thirteenth Language Resources and Evaluation ...

    2022

  24. [33]

    Chris Hays, Zachary Schutzman, Manish Raghavan, Erin Walk, and Philipp Zimmer. 2023. https://doi.org/10.1145/3543507.3583214 Simplistic collection and labeling practices limit the utility of benchmark datasets for twitter bot detection . In Proceedings of the ACM Web Conferenc...

    2023

  25. [34]

    Christopher Hidey, Elena Musi, Alyssa Hwang, Smaranda Muresan, and Kathy McKeown. 2017. https://doi.org/10.18653/v1/W17-5102 Analyzing the semantic types of claims and premises in an online persuasive forum . In Proceedings of the 4th Workshop on Argument Mining, pages 11--21,...

    2017 doi

  26. [35]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. https://proceedings.mlr.press/v97/houlsby19a.html Parameter-efficient transfer learning for NLP . In Proceedings of the 36th In...

    2019

  27. [36]

    Hospice Houngbo and Robert Mercer. 2014. https://doi.org/10.3115/v1/W14-2103 An automated method to build a corpus of rhetorically-classified sentences in biomedical texts . In Proceedings of the First Workshop on Argumentation Mining, pages 19--23, Baltimore, Maryland. Associ...

    2014 doi

  28. [37]

    Xinyu Hua, Mitko Nikolov, Nikhil Badugu, and Lu Wang. 2019. https://doi.org/10.18653/v1/N19-1219 Argument mining for understanding peer reviews . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Langua...

    2019 doi

  29. [38]

    Alistair Knott and Robert Dale. 1994. https://doi.org/10.1080/01638539409544883 Using linguistic phenomena to motivate a set of coherence relations . Discourse Processes, 18(1):35--62

    1994 doi

  30. [39]

    Takahiro Kondo, Koki Washio, Katsuhiko Hayashi, and Yusuke Miyao. 2021. https://doi.org/10.18653/v1/2021.argmining-1.11 B ayesian argumentation-scheme networks: A probabilistic model of argument validity facilitated by argumentation schemes . In Proceedings of the 8th Workshop...

    2021 doi

  31. [40]

    Anne Lauscher, Goran Glava s , and Simone Paolo Ponzetto. 2018. https://doi.org/10.18653/v1/W18-5206 An argument-annotated corpus of scientific publications . In Proceedings of the 5th Workshop on Argument Mining, pages 40--46, Brussels, Belgium. Association for Computational ...

    2018 doi

  32. [41]

    John Lawrence, Floris Bex, Chris Reed, and Mark Snaith. 2012. https://doi.org/10.3233/978-1-61499-111-3-515 Aifdb: Infrastructure for the argument web . In Computational Models of Argument, Frontiers in Artificial Intelligence and Applications

    2012 doi

  33. [42]

    John Lawrence and Chris Reed. 2019. https://doi.org/10.1162/coli_a_00364 Argument mining: A survey . Computational Linguistics, 45(4):765--818

    2019 doi

  34. [43]

    Ran Levy, Ben Bogin, Shai Gretz, Ranit Aharonov, and Noam Slonim. 2018. https://aclanthology.org/C18-1176/ Towards an argumentative content search engine using weak supervision . In Proceedings of the 27th International Conference on Computational Linguistics, pages 2066--2081...

    2018

  35. [44]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized BERT pretraining approach . CoRR, abs/1907.11692

    2019 arXiv

  36. [45]

    Henrique Lopes Cardoso, Rui Sousa-Silva, Paula Carvalho, and Bruno Martins. 2023. https://doi.org/10.1017/S1351324923000062 Argumentation models and their use in corpus annotation: Practice, prospects, and challenges . Natural Language Engineering, 29(4):1150–1187

    2023 doi

  37. [46]

    Tobias Mayer, Elena Cabrio, Marco Lippi, Paolo Torroni, and Serena Villata. 2018. https://doi.org/10.3233/978-1-61499-906-5-137 Argument mining on clinical trials . In Computational Models of Argument, Frontiers in Artificial Intelligence and Applications, pages 137--148

    2018 doi

  38. [47]

    Tobias Mayer, Elena Cabrio, and Serena Villata. 2020 a . https://hal.science/hal-02879293 Transformer-based Argument Mining for Healthcare Applications . In ECAI 2020 - 24th European Conference on Artificial Intelligence , Santiago de Compostela / Online, Spain

    2020

  39. [48]

    Tobias Mayer, Elena Cabrio, and Serena Villata. 2020 b . https://ecai2020.eu/papers/1470_paper Transformer-based argument mining for healthcare applications . In European Conference on Artificial Intelligence

    2020

  40. [49]

    Middleton, Matt Ryan, Jiatong Zhu, and Timothy J

    Rafael Mestre, Razvan Milicin, Stuart E. Middleton, Matt Ryan, Jiatong Zhu, and Timothy J. Norman. 2021. https://doi.org/10.18653/v1/2021.argmining-1.8 M -arg: Multimodal argument mining dataset for political debates with audio and transcripts . In Proceedings of the 8th Works...

    2021 doi

  41. [50]

    Amita Misra, Brian Ecker, and Marilyn Walker. 2016. https://doi.org/10.18653/v1/W16-3636 Measuring the similarity of sentential arguments in dialogue . In Proceedings of the 17th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 276--287, Los Angele...

    2016 doi

  42. [51]

    Roser Morante, Chantal van Son, Isa Maks, and Piek Vossen. 2020. https://aclanthology.org/2020.lrec-1.611/ Annotating perspectives on vaccination . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4964--4973, Marseille, France. European Languag...

    2020

  43. [52]

    Vlad Niculae, Joonsuk Park, and Claire Cardie. 2017. https://doi.org/10.18653/v1/P17-1091 Argument mining with structured SVM s and RNN s . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 985--995, Vanco...

    2017 doi

  44. [53]

    Christopher Olshefski, Luca Lugini, Ravneet Singh, Diane Litman, and Amanda Godley. 2020. https://aclanthology.org/2020.lrec-1.130/ The discussion tracker corpus of collaborative argumentation . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages ...

    2020

  45. [54]

    Juri Opitz and Anette Frank. 2019. https://doi.org/10.18653/v1/W19-4503 Dissecting content and context in argumentative relation analysis . In Proceedings of the 6th Workshop on Argument Mining, pages 25--34, Florence, Italy. Association for Computational Linguistics

    2019 doi

  46. [55]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, a...

    2022

  47. [56]

    Sinno Jialin Pan and Qiang Yang. 2010. https://api.semanticscholar.org/CorpusID:740063 A survey on transfer learning . IEEE Transactions on Knowledge and Data Engineering, 22:1345--1359

    2010

  48. [57]

    Alexander Panchenko, Alexander Bondarenko, Mirco Franzek, Matthias Hagen, and Chris Biemann. 2019. https://doi.org/10.18653/v1/W19-4516 Categorizing comparative sentences . In Proceedings of the 6th Workshop on Argument Mining, pages 136--145, Florence, Italy. Association for ...

    2019 doi

  49. [58]

    Ponzetto, and Chris Biemann

    Alexander Panchenko, Eugen Ruppert, Stefano Faralli, Simone P. Ponzetto, and Chris Biemann. 2018. https://aclanthology.org/L18-1286/ Building a web-scale dependency-parsed corpus from C ommon C rawl . In Proceedings of the Eleventh International Conference on Language Resource...

    2018

  50. [59]

    Andreas Peldszus and Manfred Stede. 2015. https://doi.org/10.18653/v1/D15-1110 Joint prediction in MST -style discourse parsing for argumentation mining . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 938--948, Lisbon, Portuga...

    2015 doi

  51. [60]

    Prakash Poudyal, Jaromir Savelka, Aagje Ieven, Marie Francine Moens, Teresa Goncalves, and Paulo Quaresma. 2020. https://aclanthology.org/2020.argmining-1.8/ ECHR : Legal corpus for argument mining . In Proceedings of the 7th Workshop on Argument Mining, pages 67--75, Online. ...

    2020

  52. [61]

    Chris Reed, Raquel Mochales Palau, Glenn Rowe, and Marie-Francine Moens. 2008. https://aclanthology.org/L08-1553/ Language resources for studying argument . In Proceedings of the Sixth International Conference on Language Resources and Evaluation ( LREC `08) , Marrakech, Moroc...

    2008

  53. [62]

    Nils Reimers, Benjamin Schiller, Tilman Beck, Johannes Daxenberger, Christian Stab, and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/P19-1054 Classification and clustering of arguments with contextualized word embeddings . In Proceedings of the 57th Annual Meeting of the ...

    2019 doi

  54. [63]

    Steffen Rendle, Li Zhang, and Yehuda Koren. 2019. https://api.semanticscholar.org/CorpusID:146120960 On the difficulty of evaluating baselines: A study on recommender systems . ArXiv, abs/1905.01395

    2019 arXiv

  55. [64]

    Khapra, Ehud Aharoni, and Noam Slonim

    Ruty Rinott, Lena Dankin, Carlos Alzate Perez, Mitesh M. Khapra, Ehud Aharoni, and Noam Slonim. 2015. https://doi.org/10.18653/v1/D15-1050 Show me your evidence - an automatic method for context dependent evidence detection . In Proceedings of the 2015 Conference on Empirical ...

    2015 doi

  56. [65]

    Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. https://doi.org/10.1162/tacl_a_00349 A primer in BERT ology: What we know about how BERT works . Transactions of the Association for Computational Linguistics, 8:842--866

    2020 doi

  57. [66]

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. https://arxiv.org/abs/1910.01108 Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter . CoRR, abs/1910.01108

    2019 arXiv

  58. [67]

    Naomi Saphra, Eve Fleisig, Kyunghyun Cho, and Adam Lopez. 2024. https://doi.org/10.18653/v1/2024.naacl-long.128 First tragedy, then parse: History repeats itself in the new era of large language models . In Proceedings of the 2024 Conference of the North American Chapter of th...

    2024 doi

  59. [68]

    Robin Schaefer and Manfred Stede. 2021. https://doi.org/doi:10.1515/itit-2020-0053 Argument mining on twitter: A survey . it - Information Technology, 63(1):45--58

    2021 doi

  60. [69]

    Eyal Shnarch, Leshem Choshen, Guy Moshkowich, Ranit Aharonov, and Noam Slonim. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.243 Unsupervised expressive rules provide explainability and assist human experts grasping new domains . In Findings of the Association for Comp...

    2020 doi

  61. [70]

    Christian Stab and Iryna Gurevych. 2014. https://aclanthology.org/C14-1142/ Annotating argument components and relations in persuasive essays . In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers , pages 1501--1510, D...

    2014

  62. [71]

    Christian Stab and Iryna Gurevych. 2017. https://doi.org/10.1162/COLI_a_00295 Parsing argumentation structures in persuasive essays . Computational Linguistics, 43(3):619--659

    2017 doi

  63. [72]

    Christian Stab, Tristan Miller, Benjamin Schiller, Pranav Rai, and Iryna Gurevych. 2018. https://doi.org/10.18653/v1/D18-1402 Cross-topic argument mining from heterogeneous sources . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pag...

    2018 doi

  64. [73]

    Reid Swanson, Brian Ecker, and Marilyn Walker. 2015. https://doi.org/10.18653/v1/W15-4631 Argument mining: Extracting arguments from online dialogue . In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 217--226, Prague, Cze...

    2015 doi

  65. [74]

    Milagro Teruel, Cristian Cardellino, Fernando Cardellino, Laura Alonso Alemany, and Serena Villata. 2018. https://aclanthology.org/L18-1640/ Increasing argument annotation reproducibility by using inter-annotator agreement to improve guidelines . In Proceedings of the Eleventh...

    2018

  66. [75]

    Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. 2021. https://doi.org/10.18653/v1/2021.naacl-main.28 Augmented SBERT : Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks . In Proceedings of the 2021 Conference of the N...

    2021 doi

  67. [76]

    Terne Sasha Thorn Jakobsen, Maria Barrett, and Anders S gaard. 2021. https://doi.org/10.18653/v1/2021.starsem-1.25 Spurious correlations in cross-topic argument mining . In Proceedings of *SEM 2021: The Tenth Joint Conference on Lexical and Computational Semantics, pages 263--...

    2021 doi

  68. [77]

    Dietrich Trautmann. 2020. https://aclanthology.org/2020.argmining-1.5/ Aspect-based argument mining . In Proceedings of the 7th Workshop on Argument Mining, pages 41--52, Online. Association for Computational Linguistics

    2020

  69. [78]

    Dietrich Trautmann, Johannes Daxenberger, Christian Stab, Hinrich Schütze, and Iryna Gurevych. 2020. https://doi.org/10.1609/aaai.v34i05.6438 Fine-grained argument unit recognition and classification . Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):9048--9056

    2020 doi

  70. [79]

    Eva Maria Vecchi, Neele Falk, Iman Jundi, and Gabriella Lapesa. 2021. https://doi.org/10.18653/v1/2021.acl-long.107 Towards argument mining for social good: A survey . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Inter...

    2021 doi

  71. [80]

    Marilyn Walker, Jean Fox Tree, Pranav Anand, Rob Abbott, and Joseph King. 2012. https://aclanthology.org/L12-1643/ A corpus for research on deliberation and debate . In Proceedings of the Eighth International Conference on Language Resources and Evaluation ( LREC `12) , pages ...

    2012

  72. [81]

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/W18-5446 GLUE : A multi-task benchmark and analysis platform for natural language understanding . In Proceedings of the 2018 EMNLP Workshop B lackbox NLP : A...

    2018 doi

  73. [82]

    Michael Wojatzki and Torsten Zesch. 2016. https://api.semanticscholar.org/CorpusID:85555944 Stance-based argument mining - modeling implicit argumentation using stance . In Conference on Natural Language Processing

    2016

  74. [83]

    Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, and Guoyin Wang. 2024. https://arxiv.org/abs/2308.10792 Instruction tuning for large language models: A survey . Preprint, arXiv:2308.10792

    2024

  75. [84]

    Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. 2019. https://arxiv.org/abs/1911.02685 A comprehensive survey on transfer learning . CoRR, abs/1911.02685

    2019 arXiv

Pith tools

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