Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Multi-Hop Question Generation via Dual-Perspective Keyword Guidance

T0 review · 4 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 Splitting keywords into question-intent and document-content roles improves multi-hop question generation, and the DPKG framework built on that split ranks first or second on most of six HotpotQA metrics.

desk verdict Dual-perspective keyword guidance for multi-hop question generation is a real, well-ablated contribution; the SOTA claim is plausible but rests on metrics the design can inflate, and the paper still deserves a serious referee. read the letter →

arxiv 2505.15299 v2 pith:AOSU52BL submitted 2025-05-21 cs.CL

Maodong Li , Longyin Zhang , Fang Kong This is my paper · ORCID
classification cs.CL
keywords multi-hopquestiongenerationkeyword-guidedkeywordsdocumentanswer-awareattentionHotpotQAcontrollabletexttransformer
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

Multi-hop question generation — writing a question that forces the reader to stitch together several facts in a document to reach a given answer — is currently guided by keywords that most systems treat as one undifferentiated blob. This paper claims that the keywords actually come in two roles: question keywords, which encode the questioner's intent and must surface in the generated question, and document keywords, which flag the content in the document that the answer depends on. The authors build the Dual-Perspective Keyword-Guided (DPKG) framework, an expanded transformer encoder feeding two answer-aware decoders, one that produces the keyword pairs and one that turns them, plus the document and answer, into the final question. On HotpotQA, DPKG in its hard mode (explicit role tags) and soft mode (roles learned) ranks first or second on most of six evaluation metrics, and the same recipe also beats strong baselines on MusiQue-2hop. If the claim holds, keyword guidance of this split form is a cheap and controllable lever for generating the multi-hop questions that question-answering systems need for training and evaluation.

What carries the argument

The load-bearing mechanism is the Dual-Perspective Keyword-Guided (DPKG) architecture: an expanded transformer encoder that encodes the document, the answer, and a concatenated document-answer view in parallel with shared self-attention, followed by two answer-aware transformer decoders. The keyword decoder emits question and document keywords (in hard mode with <qes>/<doc> prefixes, in soft mode plain); the question decoder then re-encodes those keywords with a vanilla BART encoder and fuses them into the generation. Two components carry the argument: the answer-aware attention mechanism, which weights document tokens by their learned relation to the answer so the decoder reads the crucial snippets, and the gated fusion module that mixes answer-aware states with ordinary cross-attention states. A joint training loss $L = \beta_1 L_1 + \beta_2 L_2 + \beta_3 L_3$ combines keyword cross-entropy, question cross-entropy, and the representation-matching term $L_3$ that closes the train/inference gap between ground-truth and self-generated keywords.

What would settle it

Run the released DPKG checkpoints on the full-document test set with document keywords replaced by random entities sampled from the same paragraphs (question keywords left intact) and compare BLEU-4, ROUGE-L, and QA answerability (EM/F1 via the same Qwen-plus setup the paper uses); if scores barely move, the claimed pinpointing function of document keywords is not doing the work the paper attributes to it.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that separating keywords by perspective — question intent versus document content — and enforcing that question keywords appear in the output makes a difference that no single-keyword scheme captures. The authors demonstrate this with the DPKG framework, whose hard mode tags each generated keyword as <qes> or <doc> and whose soft mode lets the question decoder infer the roles, and they show the two modes trade off: hard mode gives the stronger guidance once keywords are known (best results with ground-truth keywords), while soft mode generates keywords slightly better and wins in the noisy full-document setting. The ablation results claim that both the answer-aware attention mechanism and the auxiliary loss $L_3$, which pulls the generated keyword representations toward the ground-truth keyword representations, contribute real gains. Question keywords alone outperform document keywords alone, yet both together beat either, which the paper reads as evidence that the two roles carry complementary information rather than being redundant.

Load-bearing premise

The whole comparison leans on the SpaCy-based annotation rule that throws away extracted keywords not found in the document; if that filter, rather than the question/document role split, is what produces the better questions, the framework's advantage could evaporate once the annotation changes.

Editorial extensions

If this is right

  • With generated keywords on HotpotQA, DPKGhard and DPKGsoft rank first or second on most of the six metrics, with hard mode best in the supporting-fact setting and soft mode best in the full-document setting.
  • When ground-truth keywords are supplied, hard mode's explicit role tags give the strongest guidance, especially on the long, noisy full-document inputs, implying the framework's value is partly in pinpointing snippets amid irrelevant text.
  • Question keywords are the richer signal: DPKG_Q beats DPKG_D on question quality even with weaker keyword generation, but both together beat either alone, so the two roles are complementary.
  • The recipe transfers: on MusiQue-2hop, both DPKG modes beat CQG, MulQG, E2EQR, and DP-Graph by wide margins.
  • Prompting an LLM (Qwen-Plus) with dual-perspective keywords greatly lifts its generated-question quality over no-keyword prompting, yet the fine-tuned DPKG still beats both prompted and LoRA-fine-tuned LLMs, suggesting the task rewards dedicated training.

Reading between the lines

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

  • The reported gains could be partly an artifact of the annotation filter, since SpaCy keywords that do not appear in the document are discarded; swapping in a different extractor or keeping the discarded keywords to see whether the hard/soft gap survives would separate the dual-perspective idea from the filtering effect.
  • The $L_3$ loss's large ablation impact hints that the pipelined keyword bottleneck, not the decoder design, is the main performance cost; a single-pass decoder that attends directly to keyword prototypes might capture the same gain more cheaply.
  • Because the same surface word can be tagged a question keyword by hard mode and a document keyword by soft mode, role assignment is not a fixed property of the word but of its function in the question; that opens the door to learning role labels end-to-end from the soft mode's attention patterns.
  • The dual-perspective split is a natural template for other controllable generation tasks where intent tokens and content tokens play distinct roles — for instance, rationale generation or contrastive summarization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes DPKG, a dual-perspective keyword-guided framework for multi-hop question generation. It distinguishes question keywords (which capture questioner intent and are expected to appear in the generated question) from document keywords (which mark QA-relevant content). The model consists of an expanded transformer encoder and two answer-aware decoders, one for keyword generation and one for question generation, with hard and soft modes for signaling keyword type. A joint loss includes an L3 alignment term intended to bridge the train/inference gap between ground-truth and generated keywords. Experiments on HotpotQA (SF and Full) and MuSiQue-2hop report improved BLEU-4 and ROUGE-L over prior systems, with ablations showing benefits from the L3 loss, the answer-aware attention, and the dual-keyword design.

Significance. The paper has several strengths: it is a clearly specified plan-then-generate system; the code is released; the ablations are consistent with the design choices; and the authors include a second-dataset generalization study, QA-based evaluation, human evaluation, and an explicit limitations section. I do not see a circularity problem, since keywords are an intermediate representation, the final evaluation uses generated keywords, and the L3 loss is a legitimate attempt to close the train/inference gap. However, the central SOTA claim is currently supported mainly by small margins on lexical-overlap metrics, and the semantic evaluations do not include the closest n-gram competitors. If these gaps are closed, the contribution would be a solid, useful advance for MQG.

major comments (4)
  1. [§4.3, Table 2] The claim of state-of-the-art performance rests on small margins that are not tested for significance. For example, SF BLEU-4 is 26.80 for DPKGhard vs. 26.16 for SGCM, SF ROUGE-L is 46.50 vs. 46.48 for QA4QG, Full BLEU-4 is 23.33 for DPKGsoft vs. 22.61 for SGCM, and Full ROUGE-L is 43.29 vs. 42.44. No variance, confidence intervals, or significance tests are reported for any comparison. Moreover, on METEOR the method is not state of the art: SF DPKGhard has 27.87 vs. SGCM's 28.51, and Full DPKGsoft has 25.21 vs. SGCM's 26.04. Please add significance tests or confidence intervals for the main comparisons and temper the 'SOTA in most cases' statement accordingly.
  2. [§4.3, Appendices D.2 and D.3] The semantic evaluations do not include the strongest n-gram competitors. Table 10 compares DPKG only with BART and ground-truth questions on a QA system, and Table 11 compares DPKG only with BART and TS-BART in human evaluation. Because question keywords are extracted from the gold question and the model is trained to include them in the output, the small BLEU-4/ROUGE-L gains over CQG, SGCM, and QA4QG could be explained by increased lexical overlap with the reference without improving answerability or multi-hop complexity. To rule out this alternative explanation, please evaluate DPKG and at least CQG, SGCM, and QA4QG under the same QA-system and human-evaluation protocols.
  3. [§3.4, Figure 2] The paper repeatedly states that question keywords 'must appear' in the generated question, but the architecture and losses in Eqs. (2)-(7) and (11) do not enforce this as a hard constraint, and no experiment reports the rate at which gold or generated question keywords actually appear in the output. If this property is intended as a soft inductive bias, the wording should be revised; if it is intended as a hard requirement, please describe the mechanism (e.g., constrained decoding, coverage loss) and report the coverage metric.
  4. [§4.1, Appendix B] The annotation procedure excludes the 'Not utilized' portion of SpaCy-extracted keywords, requiring both question and document keywords to appear in the document. This is presented as an intuitive principle and then justified empirically with DPKGv2 on HotpotQA. The concern is that this is a post hoc data-selection rule fitted to the same benchmark, and the DPKGv2 comparison only shows a trade-off on one dataset, not that the filtering rule generalizes. Please provide a robustness analysis on a second dataset (e.g., MuSiQue-2hop with an unfiltered annotation variant) or otherwise justify that the exclusion rule is not a source of overfitting to HotpotQA.
minor comments (6)
  1. [Throughout] The token 'DPKGsof t' appears in text and tables; it should be 'DPKGsoft' or 'DPKG_soft' consistently.
  2. [§4.2, Table 2] BERTScore is listed as an evaluation metric in §4.2 but is not reported in the main comparison in Table 2; please add it to the main table or remove it from the metric list.
  3. [Appendix E] The dataset name 'MusiQue' should be typeset as 'MuSiQue' to match the reference by Trivedi et al. (2022).
  4. [§3.3] The notation H^{t-1}_k and H^t_k in Eqs. (4)-(6) suggests a step index for keyword representations, but the decoding steps are never formally defined; please clarify the index and the update process.
  5. [Appendix A] The loss weights β1, β2, β3 are all set to 1.0, but no sensitivity analysis is reported; a brief note on whether the result is stable around this setting would be useful.
  6. [§4.1] Treating all interrogative words as question keywords is a reasonable choice, but it means the question-keyword set is partially determined by the question type; the paper could discuss how this interacts with the 'must appear' property.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DPKG is a standard supervised plan-then-generate pipeline with external baselines; the BLEU/ROUGE-overlap concern is an evaluation-validity issue, not a circularity.

full rationale

The paper's derivation chain is self-contained. The framework is a conventional plan-then-generate pipeline: Section 3.1 defines keyword generation as K_i = f_theta1(D_i, A_i) and question generation as Q_i = f_theta2(D_i, A_i, K_i), with keywords annotated from the gold question and supporting sentences via SpaCy during dataset construction. At inference, keywords are first generated by the model and then fed to the question decoder, so the final prediction is not read back from a fitted label. The L3 loss (Section 3.5, Eq. 8) is a representation-matching regularizer between generated and ground-truth keyword encodings; it is a training objective, not a definitional identity that forces the output to equal an input. The requirement that question keywords appear in the generated question is a design constraint on decoding, not a circular equivalence: the question tokens are not given as model input during inference, and BLEU-4/ROUGE-L are evaluated against the gold question rather than against the input keywords. No fitted parameter is renamed as a prediction, and no central result is imported from the authors' own prior work: the reference list contains no self-citations by Li, Zhang, or Kong, and the data split, answer-aware attention, and soft-mode idea are attributed to external works (Fei et al. 2022a; Wang et al. 2024; Liu et al. 2024a). The skeptical concern that question keywords derived from gold questions may inflate lexical-overlap metrics is a legitimate empirical-validity threat, and the paper's human and QA evaluations compare only against BART and TS-BART rather than the closest n-gram competitors; however, that is a completeness and robustness limitation, not a reduction of the claimed result to its inputs. Accordingly, no circular step can be exhibited, and the honest finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a manual keyword annotation scheme and a hand-set loss weighting. The annotation scheme is the most consequential assumption, since the framework's advantage over baselines depends on the specific split of keywords into question and document types and on discarding a subset of extracted keywords.

free parameters (1)
  • Loss weights β1, β2, β3 = 1.0 (all)
    Set to equal weights by hand; no sensitivity analysis is reported, so the central result may depend on this choice.
assumptions (3)
  • domain assumption SpaCy en_core_web_sm provides adequate keyword extraction for question and document keywords.
    The entire keyword annotation pipeline relies on SpaCy's entity and phrase extraction; no independent validation of extraction quality is provided.
  • domain assumption Question keywords must appear in the generated question.
    This is imposed as a hard constraint in the task formulation and is not derived from user studies; it is an assumption about what makes a good multi-hop question.
  • ad hoc to paper The 'Not utilized' portion of extracted keywords should be discarded.
    Section 4.1 excludes keywords that are not in the document, and Appendix B shows including them (DPKGv2) hurts performance under generated keywords. The exclusion is motivated post hoc and may be dataset-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Hop Question Generation via Dual-Perspective Keyword Guidance." pith.science (2026). https://pith.science/paper/AOSU52BL

@misc{pith2026250515299,
  author       = {Pith},
  title        = {Pith review of: Multi-Hop Question Generation via Dual-Perspective Keyword Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AOSU52BL}},
  note         = {Machine review of arXiv:2505.15299}
}
read the original abstract

Multi-hop question generation (MQG) aims to generate questions that require synthesizing multiple information snippets from documents to derive target answers. The primary challenge lies in effectively pinpointing crucial information snippets related to question-answer (QA) pairs, typically relying on keywords. However, existing works fail to fully utilize the guiding potential of keywords and neglect to differentiate the distinct roles of question-specific and document-specific keywords. To address this, we define dual-perspective keywords (i.e., question and document keywords) and propose a Dual-Perspective Keyword-Guided (DPKG) framework, which seamlessly integrates keywords into the multi-hop question generation process. We argue that question keywords capture the questioner's intent, whereas document keywords reflect the content related to the QA pair. Functionally, question and document keywords work together to pinpoint essential information snippets in the document, with question keywords required to appear in the generated question. The DPKG framework consists of an expanded transformer encoder and two answer-aware transformer decoders for keyword and question generation, respectively. Extensive experiments demonstrate the effectiveness of our work, showcasing its promising performance and underscoring its significant value in the MQG task.

Figures

Figures reproduced from arXiv: 2505.15299 by the authors.

Figure 1
Figure 1. (a) shows the definition of dual-perspective [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of DPKG framework. In (c), both modules receive the same inputs as in (b), [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) shows our data annotation process; (b) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: An example of multi-hop question generation is provided, where the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Prompt template. "w/ keyword guidance" refers to dual-perspective keyword guidance. According to the results in [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 22 canonical work pages

  1. [1]

    Satanjeev Banerjee and Alon Lavie. 2005. https://aclanthology.org/W05-0909 METEOR : An automatic metric for MT evaluation with improved correlation with human judgments . In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization , pages 65--72, Ann Arbor, Michigan. Association for Comput...

    2005

  2. [2]

    Shuyang Cao and Lu Wang. 2021. https://doi.org/10.18653/v1/2021.acl-long.502 Controllable open-ended question generation with a new question type ontology . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 642...

    doi:10.18653/v1/2021.acl-long.502 2021

  3. [3]

    Yllias Chali and Sadid A. Hasan. 2012. https://aclanthology.org/C12-1030 Towards automatic topical question generation . In Proceedings of COLING 2012 , pages 475--492, Mumbai, India. The COLING 2012 Organizing Committee

    work page 2012

  4. [4]

    Yu Chen, Lingfei Wu, and Mohammed J. Zaki. 2020. https://openreview.net/forum?id=HygnDhEtvr Reinforcement learning based graph-to-sequence model for natural question generation . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net

    work page 2020

  5. [5]

    Yu Chen, Lingfei Wu, and Mohammed J Zaki. 2023. Toward subgraph-guided knowledge graph question generation with graph neural networks. IEEE Transactions on Neural Networks and Learning Systems

    work page 2023

  6. [6]

    Chuyao Ding, Yu Hong, and Jianmin Yao. 2024. https://aclanthology.org/2024.lrec-main.1285 SGCM : Salience-guided context modeling for question generation . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 14755--14762, Torino, Italia. ELRA and ICCL

    work page 2024

  7. [7]

    Xinya Du, Junru Shao, and Claire Cardie. 2017. https://doi.org/10.18653/v1/P17-1123 Learning to ask: Neural question generation for reading comprehension . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1342--1352, Vancouver, Canada. Association for Computational Linguistics

    doi:10.18653/v1/p17-1123 2017

  8. [8]

    Barah Fazili, Koustava Goswami, Natwar Modani, and Inderjeet Nair. 2024. Gensco: Can question decomposition based passage alignment improve question answering? arXiv preprint arXiv:2407.10245

    arXiv 2024

Show all 41 references
  1. [9]

    Zichu Fei, Qi Zhang, Tao Gui, Di Liang, Sirui Wang, Wei Wu, and Xuanjing Huang. 2022 a . https://doi.org/10.18653/v1/2022.acl-long.475 CQG : A simple and effective controlled generation framework for multi-hop question generation . In Proceedings of the 60th Annual Meeting of ...

    2022 doi

  2. [10]

    Zichu Fei, Qi Zhang, and Yaqian Zhou. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.201 Iterative GNN -based decoder for question generation . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2573--2582, Online and Punta Cana...

    2021 doi

  3. [11]

    Zichu Fei, Xin Zhou, Tao Gui, Qi Zhang, and Xuanjing Huang. 2022 b . https://aclanthology.org/2022.coling-1.572 LFKQG : A controlled generation framework with local fine-tuning for question generation over knowledge bases . In Proceedings of the 29th International Conference o...

    2022

  4. [12]

    Shasha Guo, Lizi Liao, Cuiping Li, and Tat-Seng Chua. 2024. https://doi.org/10.24963/ijcai.2024/889 A survey on neural question generation: Methods, applications, and prospects . In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCA...

    2024 doi

  5. [13]

    Michael Heilman and Noah A. Smith. 2010. https://aclanthology.org/N10-1086 Good question! statistical ranking for question generation . In Human Language Technologies: The 2010 Annual Conference of the North A merican Chapter of the Association for Computational Linguistics , ...

    2010

  6. [14]

    Seonjeong Hwang, Yunsu Kim, and Gary Geunbae Lee. 2024. https://aclanthology.org/2024.lrec-main.599 Explainable multi-hop question generation: An end-to-end approach without intermediate question labeling . In Proceedings of the 2024 Joint International Conference on Computati...

    2024

  7. [15]

    Kyungho Kim, Seongmin Park, Junseo Lee, and Jihwa Lee. 2024. https://aclanthology.org/2024.lrec-main.1075 Non-essential is NE cessary: Order-agnostic multi-hop question generation . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Languag...

    2024

  8. [16]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. https://doi.org/10.18653/v1/2020.acl-main.703 BART : Denoising sequence-to-sequence pre-training for natural language generation, translatio...

    2020 doi

  9. [17]

    Yuanyuan Liang, Jianing Wang, Hanlun Zhu, Lei Wang, Weining Qian, and Yunshi Lan. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.263 Prompting large language models with chain-of-thought for few-shot knowledge base question generation . In Proceedings of the 2023 Conference...

    2023 doi

  10. [18]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013 ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  11. [19]

    Jintao Liu, Ruixue Ding, Linhao Zhang, Pengjun Xie, and Fie Huang. 2024 a . https://arxiv.org/abs/2410.12248 Cofe-rag: A comprehensive full-chain evaluation framework for retrieval-augmented generation with enhanced data diversity . Preprint, arXiv:2410.12248

    2024 arXiv

  12. [20]

    Naiming Liu, Zichao Wang, and Richard Baraniuk. 2024 b . https://arxiv.org/abs/2406.13188 Synthetic context generation for question generation . Preprint, arXiv:2406.13188

    2024 arXiv

  13. [21]

    Karen Mazidi and Rodney D. Nielsen. 2014. https://doi.org/10.3115/v1/P14-2053 Linguistic considerations in automatic question generation . In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 321--326, Balti...

    2014 doi

  14. [22]

    Lidiya Murakhovs ' ka, Chien-Sheng Wu, Philippe Laban, Tong Niu, Wenhao Liu, and Caiming Xiong. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.111 M ix QG : Neural question generation with mixed answer types . In Findings of the Association for Computational Linguistics...

    2022 doi

  15. [23]

    Liangming Pan, Yuxi Xie, Yansong Feng, Tat-Seng Chua, and Min-Yen Kan. 2020. https://doi.org/10.18653/v1/2020.acl-main.135 Semantic graphs for generating deep questions . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1463--14...

    2020 doi

  16. [24]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei - Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 Bleu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, ...

    2002

  17. [25]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. https://doi.org/10.18653/v1/D16-1264 SQ u AD : 100,000+ questions for machine comprehension of text . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-...

    2016 doi

  18. [26]

    Linfeng Song, Zhiguo Wang, Wael Hamza, Yue Zhang, and Daniel Gildea. 2018. https://doi.org/10.18653/v1/N18-2090 Leveraging context information for natural question generation . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computat...

    2018 doi

  19. [27]

    Dan Su, Peng Xu, and Pascale Fung. 2022. Qa4qg: using question answering to constrain multi-hop question generation. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8232--8236. IEEE

    2022

  20. [28]

    Dan Su, Yan Xu, Wenliang Dai, Ziwei Ji, Tiezheng Yu, and Pascale Fung. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.416 Multi-hop question generation with graph convolutional network . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 463...

    2020 doi

  21. [29]

    Xingwu Sun, Jing Liu, Yajuan Lyu, Wei He, Yanjun Ma, and Shi Wang. 2018. https://doi.org/10.18653/v1/D18-1427 Answer-focused and position-aware neural question generation . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3930--3...

    2018 doi

  22. [30]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. https://doi.org/10.1162/tacl_a_00475 M u S i Q ue: Multihop questions via single-hop question composition . Transactions of the Association for Computational Linguistics, 10:539--554

    2022 doi

  23. [31]

    Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2023. https://doi.org/10.18653/v1/2023.acl-demo.8 A practical toolkit for multilingual question and answer generation . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V...

    2023 doi

  24. [32]

    A Vaswani. 2017. Attention is all you need. Advances in Neural Information Processing Systems

    2017

  25. [33]

    Jian Wang, Dongding Lin, and Wenjie Li. 2024. https://doi.org/10.1109/TNNLS.2023.3242071 A target-driven planning approach for goal-directed dialog systems . IEEE Transactions on Neural Networks and Learning Systems, 35(8):10475--10487

    2024

  26. [34]

    Zehua Xia, Qi Gou, Bowen Yu, Haiyang Yu, Fei Huang, Yongbin Li, and Cam - Tu Nguyen. 2023. https://doi.org/10.18653/V1/2023.FINDINGS-EMNLP.57 Improving question generation with multi-level content planning . In Findings of the Association for Computational Linguistics: EMNLP 2...

    2023 doi

  27. [35]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/v1/D18-1259 H otpot QA : A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference...

    2018 doi

  28. [36]

    Ruqing Zhang, Jiafeng Guo, Lu Chen, Yixing Fan, and Xueqi Cheng. 2021. https://doi.org/10.1145/3468889 A review on question generation from natural language text . ACM Trans. Inf. Syst., 40(1)

    2021 doi

  29. [37]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://openreview.net/forum?id=SkeHuCVFDr Bertscore: Evaluating text generation with BERT . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April...

    2020

  30. [38]

    Yao Zhao, Xiaochuan Ni, Yuanyuan Ding, and Qifa Ke. 2018. https://doi.org/10.18653/v1/D18-1424 Paragraph-level neural question generation with maxout pointer and gated self-attention networks . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Proc...

    2018 doi

  31. [39]

    Qingyu Zhou, Nan Yang, Furu Wei, Chuanqi Tan, Hangbo Bao, and Ming Zhou. 2018. Neural question generation from text: A preliminary study. In Natural Language Processing and Chinese Computing, pages 662--671, Cham. Springer International Publishing

    2018

  32. [40]

    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...

  33. [41]

    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 gl...

Pith tools

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