Pith. sign in

REVIEW 2 major objections 6 minor 50 references

Multi-Lingual Implicit Discourse Relation Recognition with Multi-Label Hierarchical Learning

T0 review · 2 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read HArch, a hierarchical multi-task model, predicts multi-label implicit discourse relations better than few-shot GPT-4o and Llama-4-Maverick across English, German, French, and Czech.

desk verdict First multilingual multi-label IDRR on DiscoGeM 2.0, with a modest but sensible hierarchical extension; numbers are plausible, but the split-caption mix-up and missing cross-lingual comparability discussion need attention. read the letter →

arxiv 2508.20712 v1 pith:PQIDGCSC submitted 2025-08-28 cs.CL

Nelson Filipe Costa , Leila Kosseim This is my paper · ORCID
classification cs.CL
keywords implicitdiscourserelationrecognitionmulti-labelclassificationmultilingualNLPconnectivesPDTB3.0hierarchicalmulti-tasklearningDiscoGeM2.0few-shotprompting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that recognizing implicit discourse relations — the unspoken links, such as cause, contrast, or expansion, between two sentences — improves when a model is trained to predict the full three-level PDTB 3.0 sense hierarchy at once, feeding coarse sense predictions into finer-grained heads. The task is multi-label: each relation is represented as a probability distribution over several possible senses, and the model must reproduce that distribution. To test this, the paper introduces HArch, a multi-task encoder architecture trained on DiscoGeM 2.0, the first parallel four-language corpus of crowd-sourced multi-label discourse relations (English, German, French, Czech). HArch is presented as the first multilingual and multi-label system for implicit discourse relation recognition, and it reports new state-of-the-art results on the related English DiscoGeM 1.0 benchmark. The headline comparison is that fine-tuned RoBERTa-HArch and XLM-RoBERTa-HArch (125M parameters) produce sense distributions closer to the crowd annotators' distributions than few-shot GPT-4o and Llama-4-Maverick do, in every language and at every sense level. If true, this means task-specific hierarchical fine-tuning of small encoders still beats very large prompted generalist models for this discourse task, and that multilingual training can compensate for scarce German, French, and Czech data.

What carries the argument

The cascade of augmentation blocks and learnable weighted sums. Each lower-level classification head's probability vector is projected up to the encoder embedding dimension through a two-step linear-plus-GELU augmentation block, then combined with the shared representation by weights α (for level 2) and β1,β2 (for level 3) before the next head. This is what makes the finer-grained sense predictions depend on the coarse senses predicted below, and it is the mechanism by which the model exploits the PDTB 3.0 hierarchy.

What would settle it

Take a random sample of 200 parallel relations from DiscoGeM 2.0 and have expert annotators in each of the four languages re-annotate the senses directly (without the connective proxy), or back-translate one language's arguments into another and compare predicted distributions. If expert distributions for the same parallel relation diverge by language more than HArch's reported error margins, or if re-annotation changes the ranking between HArch and the LLMs, the claimed cross-lingual advantage rests on incomparable annotation targets rather than on the model.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that explicitly modeling the three-level PDTB 3.0 sense hierarchy through cascading classification heads yields probability distributions over discourse senses that match human annotations more closely than predicting the levels independently. In HArch, a pretrained transformer encodes the concatenated argument pair; a shared representation feeds three softmax heads, one per sense level, and each lower-level head's output is projected back into embedding space and merged with that shared representation through learnable weighted sums. The level-2 and level-3 heads thus receive the coarse sense predictions made below them. On DiscoGeM 2.0, RoBERTa-HA

Load-bearing premise

The load-bearing premise is that the crowd-sourced connective choices in DiscoGeM 2.0 encode the same sense boundaries in English, German, French, and Czech — that the language-specific connective lists are faithful translations of the same 28 senses — so cross-lingual training signals and comparisons are based on comparable targets.

Editorial extensions

If this is right

  • Multilingual multi-label IDRR becomes feasible as a single task: a model trained on all four languages of DiscoGeM 2.0 applies to English, German, French, and Czech without per-language fine-tuning.
  • For German, French, and Czech, training on all languages together beats training on the target language alone, suggesting multilingual data compensates for scarce in-language annotations.
  • Task-specific fine-tuning of small encoders (RoBERTa-base and XLM-RoBERTa-base, 125M parameters) outperforms few-shot prompting of much larger LLMs (GPT-4o and Llama-4-Maverick) on multi-label IDRR at every sense level.
  • Adding hierarchical dependency between sense levels improves level-2 and level-3 prediction over a non-hierarchical multi-task model, while leaving level-1 performance unchanged.
  • The hierarchical approach also sets a new state of the art on the English DiscoGeM 1.0 corpus, extending the result beyond the four-language corpus.

Reading between the lines

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

  • The reported multilingual-vs-single-language gap grows as target-language training data shrinks (Czech benefits most; English is slightly better single-language), suggesting a data-efficiency law that could be tested by subsampling English to Czech size and rerunning the comparison.
  • Because the prompt mirrors DiscoGeM 2.0's connective-selection annotation method, the LLM comparison measures protocol-following as much as discourse reasoning; prompting with sense definitions instead of connectives could produce a different ranking.
  • The trained values of α, β1, and β2 would reveal how much each level actually leans on the lower-level heads; ablating the augmentation blocks or inspecting these weights could isolate how much of the gain is the hierarchy itself rather than shared representations.
  • If the four languages really share the same sense distributions for parallel relations, HArch could serve as a weak-supervision bridge: labels from one language could initialize models for new languages without additional annotation.
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 HArch, a hierarchical multi-task encoder model for multi-label implicit discourse relation recognition (IDRR) that predicts probability distributions over all three levels of the PDTB 3.0 sense hierarchy. The model concatenates ARG1 and ARG2, encodes them with a pre-trained language model, and passes the shared representation to three classification heads; the outputs of the lower-level heads are projected back into the embedding space and combined with the shared representation via learnable weighted sums. The approach is evaluated on DiscoGeM 2.0 in English, German, French, Czech, and a multilingual setting, comparing several encoder backbones, and against GPT-4o and Llama-4-Maverick under few-shot prompting. The paper also reports results on DiscoGeM 1.0. The reported findings are that RoBERTa-HArch is best for English, XLM-RoBERTa-HArch is best for the multilingual setting, the hierarchical architecture improves over the prior non-hierarchical multi-label model of Costa and Kosseim (2025), fine-tuned models consistently outperform the prompted LLMs, and RoBERTa-HArch achieves SOTA on DiscoGeM 1.0.

Significance. If the results hold, this is the first multilingual multi-label IDRR model and the first published evaluation on DiscoGeM 2.0 across all four languages. The architecture is simple and the empirical program is broad: it compares multiple encoder backbones, includes a multi-task ablation, and releases code and prompts. The result that a 125M-parameter fine-tuned encoder outperforms much larger few-shot LLMs is practically relevant and consistent with prior work. The main weaknesses are the absence of statistical significance testing for several close comparisons and an unexamined assumption that the DiscoGeM 2.0 annotations are cross-lingually comparable; neither is fatal, but both need to be addressed for the stronger claims to be fully supported.

major comments (2)
  1. [Section 5, Tables 2 and 4] The paper reports means and standard deviations over three runs but provides no significance tests or paired comparisons. Several load-bearing comparisons are numerically small: English level-1 RoBERTa-HArch vs. Costa and Kosseim (2025) is 0.327±0.004 vs. 0.329±0.005; in Table 4, English XLM-RoBERTa-HArch All vs. Eng is 0.349±0.004 vs. 0.347±0.005 at level-1; German All vs. Ger at level-2 is 0.566±0.006 vs. 0.586±0.006. While the level-2 and level-3 improvements are generally larger, without per-instance paired tests (e.g., bootstrap over test items) the claim that hierarchical structure and multilingual training are consistently beneficial is not statistically grounded. I request paired significance tests or explicit statements of which differences are not significant.
  2. [Section 4, Table 6, Section 6] The reference distributions are derived by averaging crowdworker choices over language-specific lists of discourse connectives (§4). The paper does not establish that the adapted connective lists induce the same PDTB 3.0 sense boundaries in all four languages. Table 6 shows large cross-lingual differences in sense counts even after accounting for corpus size, e.g., CAUSE sums to 1,690.7 in English but only 457.5 in German, while CONDITION sums to 1.2 in English and 126.1 in German. If these differences reflect annotation or translation artifacts rather than genuine discourse differences, the conclusion in Section 6 that multilingual training helps German, French, and Czech (Table 4) could be an artifact of label distributions rather than evidence of transferable discourse knowledge. The authors should either provide annotation-reliability evidence per language (e.g., agreement, consisten
minor comments (6)
  1. [Table 2 caption vs. Section 5] Table 2's caption says results are on the test split, while Section 5 states that the validation split was used for encoder selection, and Section 6 references validation performance in Table 2. Please correct the caption to 'validation split' or clarify the actual evaluation split.
  2. [Table 4] The 'All All' row for XLM-RoBERTa-HArch reports level-3 as 0.606±0.04; the standard deviation is presumably 0.004 to match the other rows.
  3. [Equations (1) and (2)] The parameters α, β1, and β2 are called 'learnable parameters' in a 'weighted sum,' but no constraints are stated. If they are meant as convex weights, their ranges should be constrained or specified; otherwise, calling them 'learnable coefficients' would avoid confusion.
  4. [Section 4, footnote 1] The footnote lists a different total and English count than the DiscoGeM 2.0 paper (13,063 vs. 12,834; 5,847 vs. 5,618). Please explain the discrepancy explicitly, as the reader cannot tell whether this is a counting convention or an error.
  5. [Table 4, Table 5] For a new corpus like DiscoGeM 2.0, a simple majority-class or marginal-distribution baseline would calibrate the reported JS distances and help the reader assess absolute performance. Currently only comparisons to prior work and LLMs are given.
  6. [Conclusion] The 'SOTA' claim on DiscoGeM 1.0 is based on a comparison with a single prior model. Please qualify it as 'SOTA among current distribution-prediction methods' or list any other methods that predict full sense distributions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical, held-out evaluations against external baselines and ablations, not derivations from their inputs.

full rationale

The central claims are empirical rather than derived. HArch's parameters, including the learnable interpolation weights alpha, beta1, and beta2 (Equations 1 and 2), are trained on the training split and evaluated on held-out test splits of DiscoGeM 2.0 and DiscoGeM 1.0; nothing is fitted to the test predictions and then renamed as a prediction. The comparison to Costa and Kosseim (2025) is a baseline from prior work using the same encoder and training conditions, not an input to the model, and the claimed improvement is reported as measured JS distance on test data. The multi-task ablation in Section 7 isolates the architectural contribution by training single-level models with the same encoder and comparing them on the same test split. The few-shot LLM comparisons use examples drawn from the training split with temperature 0 and report scores on the test split; the prompt mirrors the annotation methodology but this is a design choice for the comparison, not a way of importing the reference labels into the model. The stated limitations about the label set and language-specific connective lists concern data validity and generalizability, not circularity. Self-citations appear as prior baseline identification and as a cited design choice for the MAE loss, but they are not load-bearing in the sense of forcing the reported results; the results are independently computable from the released corpus and code. No step reduces by construction to its own inputs, so the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper makes no mathematical derivation; its claims are empirical. The ledger records the data assumptions, metric assumptions, and hand-selected hyperparameters that the benchmark numbers rest on. The learned interpolation weights are normal trained parameters, not post-hoc fitted constants, but they are still the mechanism by which the hierarchical claim is implemented.

free parameters (6)
  • alpha
    Learnable scalar in Equation 1 that weights the augmented level-1 output against the shared hidden representation. Trained on data, value not reported.
  • beta1, beta2
    Learnable scalars in Equation 2 that weight augmented level-1 and level-2 outputs for the level-3 head. Trained on data, values not reported.
  • Augmentation hidden dimension
    Set to half the embedding size and chosen by hand based on early experiments (Section 3.1), not reported as an absolute number.
  • Learning rate = 1e-5
    Chosen for training in Section 5; used identically for compared models.
  • Training epochs = 10
    Fixed number of epochs used for all runs in Section 5.
  • Batch size = 16
    Chosen in Section 5; used for all experiments.
assumptions (5)
  • domain assumption DiscoGeM 2.0 crowd annotations are reliable, comparable multi-label ground truth across all four languages.
    The entire evaluation in Section 4 and all tables depend on the corpus annotations being valid and cross-lingually comparable.
  • domain assumption PDTB 3.0 sense hierarchy applies consistently across English, German, French, and Czech, including replacing missing level-3 senses with their level-2 parent.
    Section 4 and the Limitations state this projection/replacement strategy; cross-lingual claims inherit it.
  • domain assumption Jensen-Shannon distance is the appropriate metric for comparing predicted and reference sense distributions.
    Used throughout Sections 3, 5, and 6; all conclusions about model ranking follow this metric choice.
  • domain assumption Pre-trained encoder representations transfer across the four languages well enough for multilingual training to help.
    Section 5 assumes multilingual encoders like XLM-RoBERTa can share representations across languages; no probing or alignment analysis is provided.
  • standard math Softmax at each classification head yields a valid probability distribution over the sense labels at that level.
    Standard softmax assumption used by the three heads in Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Lingual Implicit Discourse Relation Recognition with Multi-Label Hierarchical Learning." pith.science (2026). https://pith.science/paper/PQIDGCSC

@misc{pith2026250820712,
  author       = {Pith},
  title        = {Pith review of: Multi-Lingual Implicit Discourse Relation Recognition with Multi-Label Hierarchical Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQIDGCSC}},
  note         = {Machine review of arXiv:2508.20712}
}
read the original abstract

This paper introduces the first multi-lingual and multi-label classification model for implicit discourse relation recognition (IDRR). Our model, HArch, is evaluated on the recently released DiscoGeM 2.0 corpus and leverages hierarchical dependencies between discourse senses to predict probability distributions across all three sense levels in the PDTB 3.0 framework. We compare several pre-trained encoder backbones and find that RoBERTa-HArch achieves the best performance in English, while XLM-RoBERTa-HArch performs best in the multi-lingual setting. In addition, we compare our fine-tuned models against GPT-4o and Llama-4-Maverick using few-shot prompting across all language configurations. Our results show that our fine-tuned models consistently outperform these LLMs, highlighting the advantages of task-specific fine-tuning over prompting in IDRR. Finally, we report SOTA results on the DiscoGeM 1.0 corpus, further validating the effectiveness of our hierarchical approach.

Figures

Figures reproduced from arXiv: 2508.20712 by the authors.

Figure 1
Figure 1. Architecture of our hierarchical multi-task model, HArch, for multi-label IDRR. The model takes as input [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Prompt-response template for predicting multi-label IDRR in English using DiscoGeM 2.0. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 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]

    Nicolas Boizard, Hippolyte Gisserot-Boukhlef, Duarte M. Alves, André Martins, Ayoub Hammal, Caio Corro, Céline Hudelot, Emmanuel Malherbe, Etienne Malaboeuf, Fanny Jourdan, Gabriel Hautreux, João Alves, Kevin El-Haddad, Manuel Faysse, Maxime Peyrard, Nuno M. Guerreiro, Patrick Fernandes, Ricardo Rei, and Pierre Colombo. 2025. https://arxiv.org/pdf/2503.05...

    arXiv 2025

  4. [4]

    Peter Bourgonje and Manfred Stede. 2020. https://aclanthology.org/2020.lrec-1.133/ The Potsdam Commentary Corpus 2.2: Extending Annotations for Shallow Discourse Parsing . In Proceedings of the 12th Language Resources and Evaluation Conference (LREC'20) , pages 1061--1066, Marseille, France. European Language Resources Association (ELRA)

    work page 2020

  5. [5]

    Martin Juan Jos \'e Bucher and Marco Martini. 2024. https://arxiv.org/pdf/2406.08660 Fine-Tuned 'Small' LLMs (Still) Significantly Outperform Zero-Shot Generative AI Models in Text Classification . arXiv preprint arXiv:2406.08660

    arXiv 2024

  6. [6]

    Chunkit Chan, Cheng Jiayang, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2024. https://aclanthology.org/2024.findings-eacl.47/ Exploring the Potential of ChatGPT on Sentence Level Relations: A Focus on Temporal, Causal, and Discourse Relations . In Findings of the 18th Conference of the European Chapter of the Association for Comput...

    work page 2024

  7. [7]

    Chunkit Chan, Xin Liu, Jiayang Cheng, Zihan Li, Yangqiu Song, Ginny Y Wong, and Simon See. 2023. https://aclanthology.org/2023.findings-acl.4.pdf DiscoPrompt: Path Prediction Prompt Tuning for Implicit Discourse Relation Recognition . In Findings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL'23) , pages 35--57, Toronto, ...

    work page 2023

  8. [8]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, and 1 others. 2024. https://jmlr.org/papers/volume25/23-0870/23-0870.pdf Scaling Instruction-Finetuned Language Models . Journal of Machine Learning Research (JMLR) , 25(70):1--53

    work page 2024

Show all 50 references
  1. [9]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://aclanthology.org/2020.acl-main.747 Unsupervised Cross-lingual Representation Learning at ...

    2020

  2. [10]

    Nelson Filipe Costa and Leila Kosseim. 2024. https://aclanthology.org/2024.codi-1.11 Exploring Soft-Label Training for Implicit Discourse Relation Recognition . In Proceedings of the 5th Workshop on Computational Approaches to Discourse (CODI'24) , pages 120--126, St. Julians,...

    2024

  3. [11]

    Nelson Filipe Costa and Leila Kosseim. 2025. A Multi-Task and Multi-Label Classification Model for Implicit Discourse Relation Recognition . In Proceedings of the 26th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL'25) , Avignon, France. Associ...

    2025

  4. [12]

    Nelson Filipe Costa, Nadia Sheikh, and Leila Kosseim. 2023. https://aclanthology.org/2023.ranlp-1.39 Mapping Explicit and Implicit Discourse Relations between the RST-DT and the PDTB 3.0 . In Proceedings of the 14th International Conference on Recent Advances in Natural Langua...

    2023

  5. [13]

    Vera Demberg, Merel CJ Scholman, and Fatemeh Torabi Asr. 2019. https://doi.org/10.5087/dad.2019.104 How compatible are our discourse annotation frameworks? Insights from mapping RST-DT and PDTB annotations . Dialogue & Discourse , 10(1):87--135

    2019 doi

  6. [14]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. https://arxiv.org/pdf/2407.21783 The Llama 3 Herd of Models . arXiv preprint arXiv:2407.21783

    2024 arXiv

  7. [15]

    Dan Hendrycks and Kevin Gimpel. 2016. https://arxiv.org/pdf/1606.08415 Gaussian Error Linear Units (GELUs) . arXiv preprint arXiv:1606.08415

    2016 arXiv

  8. [16]

    Scholman, and Ted J.M

    Jet Hoek, Merel C.J. Scholman, and Ted J.M. Sanders. 2021. https://aclanthology.org/2021.discann-1.1 Is there less annotator agreement when the discourse relation is underspecified? In Proceedings of the 1st Workshop on Integrating Perspectives on Discourse Annotation , pages ...

    2021

  9. [17]

    Najoung Kim, Song Feng, Chulaka Gunasekara, and Luis Lastras. 2020. https://aclanthology.org/2020.acl-main.480 Implicit Discourse Relation Classification: We Need to Talk about Evaluation . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguist...

    2020

  10. [18]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. http://arxiv.org/abs/1412.6980 Adam: A Method for Stochastic Optimization . In Proceedings of the 3rd International Conference on Learning Representations (ICLR'15) , pages 1--15, San Diego, California, USA

    2015 arXiv

  11. [19]

    Wei Liu and Michael Strube. 2023. https://aclanthology.org/2023.acl-long.874 Annotation-Inspired Implicit Discourse Relation Classification with Auxiliary Discourse Connective Generation . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguisti...

    2023

  12. [20]

    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/pdf/1907.11692.pdf RoBERTa: A Robustly Optimized BERT Pretraining Approach . arXiv preprint arXiv:1907.11692

    2019 arXiv

  13. [21]

    Wanqiu Long, N Siddharth, and Bonnie Webber. 2024. https://aclanthology.org/2024.findings-acl.500 Multi-Label Classification for Implicit Discourse Relation Recognition . In Findings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL'24), pages 84...

    2024

  14. [22]

    Wanqiu Long and Bonnie Webber. 2022. https://aclanthology.org/2022.emnlp-main.734 Facilitating Contrastive Learning of Discourse Relational Senses by Exploiting the Hierarchy of Sense Relations . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Pr...

    2022

  15. [23]

    Wanqiu Long and Bonnie Webber. 2024. https://arxiv.org/pdf/2411.14880 Leveraging Hierarchical Prototypes as the Verbalizer for Implicit Discourse Relation Recognition . arXiv preprint arXiv:2411.14880

    2024 arXiv

  16. [24]

    Eleni Miltsakaki, Rashmi Prasad, Aravind Joshi, and Bonnie Webber. 2004. http://www.lrec-conf.org/proceedings/lrec2004/pdf/618.pdf The Penn Discourse Treebank . In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04) , pages 2237--2...

    2004

  17. [25]

    OpenAI. 2024. https://arxiv.org/pdf/2410.21276 GPT-4o System Card . arXiv preprint arXiv:2410.21276

    2024 arXiv

  18. [26]

    Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008 a . http://www.lrec-conf.org/proceedings/lrec2008/pdf/754_paper.pdf The Penn Discourse TreeBank 2.0. In Proceedings of the Sixth International Conference on Language...

    2008

  19. [27]

    Rashmi Prasad, Alan Lee, Nikhil Dinesh, Eleni Miltsakaki, Geraud Campion, Aravind Joshi, and Bonnie Webber. 2008 b . https://doi.org/10.35111/nbvh-1n26 Penn Discourse Treebank Version 2.0 . LDC2008T05. Web Download. Philadelphia: Linguistic Data Consortium

    2008 doi

  20. [28]

    Rashmi Prasad, Bonnie Webber, Alan Lee, and Aravind Joshi. 2019. https://doi.org/10.35111/qebf-gk47 Penn Discourse Treebank Version 3.0 . LDC2019T05. Web Download. Philadelphia: Linguistic Data Consortium

    2019 doi

  21. [29]

    Valentina Pyatkin, Ayal Klein, Reut Tsarfaty, and Ido Dagan. 2020. https://aclanthology.org/2020.emnlp-main.224 QADiscourse - Discourse Relations as QA Pairs: Representation, Crowdsourcing and Baselines . In Proceedings of the 2020 Conference on Empirical Methods in Natural La...

    2020

  22. [30]

    Valentina Pyatkin, Frances Yung, Merel C. J. Scholman, Reut Tsarfaty, Ido Dagan, and Vera Demberg. 2023. https://doi.org/10.1162/tacl\_a\_00586 Design Choices for Crowdsourcing Implicit Discourse Relations: Revealing the Biases Introduced by Task Design . Transactions of the A...

    2023 doi

  23. [31]

    Ahmed Ruby, Christian Hardmeier, and Sara Stymne. 2025. https://aclanthology.org/2025.coling-main.363/ Multimodal Extraction and Recognition of A rabic Implicit Discourse Relations . In Proceedings of Context and Meaning: Navigating Disagreements in NLP Annotation , pages 5415...

    2025

  24. [32]

    Muhammed Saeed, Peter Bourgonje, and Vera Demberg. 2025. https://aclanthology.org/2025.coling-main.174/ Implicit Discourse Relation Classification For Nigerian Pidgin . In Proceedings of Context and Meaning: Navigating Disagreements in NLP Annotation , pages 2561--2574, Abu Dh...

    2025

  25. [33]

    Merel Scholman and Vera Demberg. 2017. https://doi.org/10.5087/dad.2017.203 Examples and Specifications that Prove a Point: Identifying Elaborative and Argumentative Discourse Relations . Dialogue & Discourse , 8(2):56--83

    2017 doi

  26. [34]

    Merel Scholman, Tianai Dong, Frances Yung, and Vera Demberg. 2022 a . https://aclanthology.org/2022.lrec-1.351 DiscoGeM: A Crowdsourced Corpus of Genre-Mixed Implicit Discourse Relations . In Proceedings of the 13th Language Resources and Evaluation Conference (LREC'22) , page...

    2022

  27. [35]

    Merel Scholman, Valentina Pyatkin, Frances Yung, Ido Dagan, Reut Tsarfaty, and Vera Demberg. 2022 b . https://aclanthology.org/2022.lrec-1.231 Design Choices in Crowdsourcing Discourse Relation Annotations: The Effect of Worker Selection and Training . In Proceedings of the 13...

    2022

  28. [36]

    Manfred Stede. 2008. https://doi.org/10.1007/s11168-008-9053-7 Disambiguating Rhetorical Structure . Research on Language and Computation , 6(3):311--332

    2008 doi

  29. [37]

    Pavl \'i na Synkov \'a , Ji r \'i M \'i rovsk \'y , Lucie Pol \'a kov \'a , and Magdal \'e na Rysov \'a . 2024. https://aclanthology.org/2024.lrec-main.114/ Announcing the Prague Discourse Treebank 3.0 . In Proceedings of the 2024 Joint International Conference on Computationa...

    2024

  30. [38]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. https://arxiv.org/pdf/2302.13971.pdf LLaMA: Open and Efficient Foundation Language Models...

    2023 arXiv

  31. [39]

    Michiel van der Meer, Neele Falk, Pradeep K Murukannaiah, and Enrico Liscio. 2024. https://aclanthology.org/2024.emnlp-main.1031/ Annotator-Centric Active Learning for Subjective NLP Tasks . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process...

    2024

  32. [40]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention Is All You Need . In Proceedings of...

    2017

  33. [41]

    Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. 2024. https://arxiv.org/pdf/2412.13663 Smarter, Better, F...

    2024 arXiv

  34. [42]

    Bonnie Webber, Rashmi Prasad, Alan Lee, and Aravind Joshi. 2019. https://catalog.ldc.upenn.edu/docs/LDC2019T05/PDTB3-Annotation-Manual.pdf The Penn Discourse Treebank 3.0 Annotation Manual . Technical report, University of Pennsylvania

    2019

  35. [43]

    Frances Yung, Mansoor Ahmad, Merel Scholman, and Vera Demberg. 2024 a . https://aclanthology.org/2024.law-1.15/ Prompting Implicit Discourse Relation Annotation . In Proceedings of the 18th Linguistic Annotation Workshop (LAW-XVIII) , pages 150--165, St. Julian's, Malta. Assoc...

    2024

  36. [44]

    Frances Yung, Kaveri Anuranjana, Merel Scholman, and Vera Demberg. 2022. https://aclanthology.org/2022.codi-1.7 Label distributions help implicit discourse relation classification . In Proceedings of the 3rd Workshop on Computational Approaches to Discourse (CODI'22) , pages 4...

    2022

  37. [45]

    Frances Yung and Vera Demberg. 2025. https://aclanthology.org/2025.comedi-1.2/ On Crowdsourcing Task Design for Discourse Relation Annotation . In Proceedings of Context and Meaning: Navigating Disagreements in NLP Annotation , pages 12--19, Abu Dhabi, UAE. International Commi...

    2025

  38. [46]

    Frances Yung, Vera Demberg, and Merel Scholman. 2019. https://aclanthology.org/W19-4003 Crowdsourcing Discourse Relation Annotations by a Two-Step Connective Insertion Task . In Proceedings of the 13th Linguistic Annotation Workshop (LAW'19) , pages 16--25, Florence, Italy. As...

    2019

  39. [47]

    Frances Yung, Merel Scholman, Sarka Zikanova, and Vera Demberg. 2024 b . https://aclanthology.org/2024.lrec-main.443/ DiscoGeM 2.0: A Parallel Corpus of English, German, French and Czech Implicit Discourse Relations . In Proceedings of the 2024 Joint International Conference o...

    2024

  40. [48]

    Lei Zeng, Ruifang He, Haowen Sun, Jing Xu, Chang Liu, and Bo Wang. 2024. https://aclanthology.org/2024.lrec-main.686/ Global and Local Hierarchical Prompt Tuning Framework for Multi-level Implicit Discourse Relation Recognition . In Proceedings of the 2024 Joint International ...

    2024

  41. [49]

    e Val \= u nait \

    Deniz Zeyrek, Giedr \. e Val \= u nait \. e Ole s kevi c ien \. e , and Amália Mendes. 2024. https://aclanthology.org/2024.bucc-1.14/ Multiple Discourse Relations in English TED Talks and Their Translation into Lithuanian, Portuguese and Turkish . In Proceedings of the 2024 Jo...

    2024

  42. [50]

    Haodong Zhao, Ruifang He, Mengnan Xiao, and Jing Xu. 2023. https://aclanthology.org/2023.acl-long.357 Infusing Hierarchical Guidance into Prompt Tuning: A Parameter-Efficient Framework for Multi-level Implicit Discourse Relation Recognition . In Proceedings of the 61st Annual ...

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.