Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

DynGraph2Seq: Dynamic-Graph-to-Sequence Interpretable Learning for Health Stage Prediction in Online Health Forums

T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Modeling a patient's subforum activity as a time-evolving graph predicts their health-stage sequence better than text-sequence or static-graph baselines.

desk verdict A plausible dynamic-graph-to-sequence architecture for a useful health-informatics task, but the empirical claims rest on a self-reported, self-selected label source and a thin evaluation. read the letter →

arxiv 1908.08497 v1 pith:7XRP6VAZ submitted 2019-08-22 cs.LG stat.ML

Yuyang Gao , Lingfei Wu , Houman Homayoun , Liang Zhao This is my paper · ORCID
classification cs.LGstat.ML
keywords dynamicgraphgraph-to-sequencehealthstagepredictiononlineforumshierarchicalattentionencoder-decoderinterpretabilitysubforumtransitionnetwork
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 tries to establish that a patient's movement between subforums of an online breast cancer forum, encoded as a time-evolving graph, can be translated into a sequence of health stages more accurately than either plain text activity or a static aggregate graph. The authors propose DynGraph2Seq, an encoder-decoder model in which each monthly snapshot is a graph whose nodes are subforums, whose edges are the user's transitions between subforums, and whose node features are content keywords. A hierarchical attention mechanism operates at two levels—which time window matters and which subforum within that window matters—making the prediction traceable. On eight years of Breast Cancer Community data, the model beats sequence-to-sequence and static graph-to-sequence baselines by 7–17% on BLEU and 6–13% on ROUGE scores, and the attention weights line up with clinically sensible subforum-to-stage correspondences. If the claim holds, health organizations could infer missing health-stage histories for the many users who never state one, directly from their forum activity.

What carries the argument

The central object is the dynamic activity-transition graph, in which each node is a subforum, each directed edge records a switch from one most-active subforum to another within a monthly time window, and each node carries features built from the 100 most frequent forum keywords. The load-bearing mechanism is the two-level dynamic graph hierarchical attention. After per-snapshot graph convolutions produce node embeddings, a node-to-graph attention layer weights which subforums best represent each month's graph; an LSTM then encodes the sequence of graph embeddings, and a graph-to-sequence attention layer weights which months matter when the decoder emits each health-stage token. This hierarchy is what makes the model both predictive and interpretable, because every decoded stage can be traced back to a specific subforum in a specific month.

What would settle it

Retrain the model and the baselines on a cohort whose health stages are confirmed by clinical records, matching the same monthly forum-activity graphs, and compare BLEU/ROUGE gaps. If the 7–17% BLEU advantage over sequence and static-graph baselines disappears or shrinks to noise, the dynamic-graph signal is not what the paper claims it is.

Watch

Extended reading notes

Core claim

The central claim is that the temporal structure of subforum transitions carries health-stage signal that sequence-only and static-graph representations discard, and that this signal can be captured by an end-to-end dynamic graph-to-sequence model. Formally, the task is a translation mapping from a dynamic graph $\mathcal{G} = \{G_1, \ldots, G_T\}$ to a target sequence of health-stage tokens, where each snapshot $G_t$ is a weighted directed graph of the user's subforum transitions in one month. The paper reports that DynGraph2Seq with dynamic graph hierarchical attention achieves the best BLEU-1 through BLEU-4 and ROUGE scores across 20 runs, outperforming baselines by 7–17% on BLEU and 6–13% on ROUGE. The paper also claims the learned attention is interpretable: at the node level, subforums such as the Radiation Therapy forum receive high attention exactly when the model decodes the Radiation stage.

Load-bearing premise

The load-bearing premise is that the health-stage sequences used for training and evaluation are accurate: they are extracted from what users write in their forum signatures, and those self-reports can lag behind real treatment changes, omit stages, or be missing entirely for many users.

Editorial extensions

If this is right

  • On the paper's dataset, DynGraph2Seq with hierarchical attention achieves the highest BLEU and ROUGE scores across all settings, beating baselines by 7–17% on BLEU and 6–13% on ROUGE.
  • The dynamic graph encoder plus LSTM captures temporal dependencies among monthly snapshots in a way that a static aggregate graph cannot, which explains the gap over the Graph2Seq baseline.
  • The two-level attention makes the inference traceable: each predicted health stage can be inspected for which month and which subforum mattered most.
  • The learned attention associations recover clinically sensible correspondences, such as the Radiation Therapy subforum for the Radiation stage and a 'Not Diagnosed but Worried' subforum for Dx.
  • The framework is a general dynamic-graph-to-sequence architecture that can be applied outside health forums to other sequence-labelled evolving networks.

Reading between the lines

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

  • Inference: Because the ground-truth health stages are self-reported and missing for many active users, the model may be learning forum conventions for describing stages rather than true clinical trajectories; validating on externally confirmed stage histories would separate these.
  • Inference: The same dynamic-graph-to-sequence formulation could be applied to other settings where a person's state is readable from transitions among activity categories, such as student topic drift in online courses or stage tracking in distributed software development, whenever a reference sequence of stages exists.
  • Inference: An operational extension the paper leaves implicit is using the two-level attention as an alert—when graph-to-sequence attention spikes on a month and node-to-graph attention shifts to a new subforum, a health organization could proactively deliver stage-appropriate resources to that user.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes DynGraph2Seq, a dynamic-graph-to-sequence neural architecture for predicting the sequence of health stages of users in an online breast cancer forum. User activity among subforums over monthly time windows is represented as a dynamic graph with subforum nodes and weighted edges for transitions of the most active forum. The model encodes each snapshot with graph convolutional networks, aggregates node embeddings into graph embeddings, applies an LSTM over the graph embedding sequence, and decodes the target health stage sequence with a hierarchical attention mechanism (node-to-graph attention and graph-to-sequence attention). Experiments on a single forum dataset compare DynGraph2Seq against sequence-to-sequence (NMT) and static Graph2Seq baselines, reporting higher BLEU and ROUGE scores. The paper also presents a qualitative case study of the learned attention weights for interpretability.

Significance. If the empirical claims are validated, the paper introduces a novel problem formulation (dynamic-graph-to-sequence learning) and a model that combines dynamic graph encoding with hierarchical attention, which is of interest to the graph representation learning and health informatics communities. The paper also provides a transparent two-level attention mechanism that, if properly evaluated, would be a genuine contribution to interpretable deep learning on temporal graphs. The strengths are the clear problem formalization, the extension of static Graph2Seq to dynamic settings, and the inclusion of an attention-based interpretability mechanism. However, the current evaluation has important gaps in label validity, baseline selection, statistical testing, and reproducibility that prevent the claims from being fully supported.

major comments (4)
  1. [Section V-A] The ground-truth health stage sequences are 'the real health stage history extracted from the users' signatures,' and the paper itself notes that over 36% of active users have not shared their disease history. Signatures are self-reported, may lag behind or omit clinical events, and their update times may be correlated with forum activity. If users edit signatures in the same periods in which they become active in treatment-specific subforums, the dynamic graph encoder could learn to predict signature-editing behavior rather than true health-stage progression. The paper does not report any analysis of label noise, such as the distribution of signature update times relative to forum activity, nor any independent validation against clinical records or manual annotation. Please provide such analyses, or at minimum discuss and empirically bound the effect of label noise on the reported gains.
  2. [Section V-B and Table I] The evaluation compares only against two neural baselines (NMT and static Graph2Seq). It lacks simple, strong baselines that are standard for categorical sequence prediction, such as predicting the most frequent health stage per user, a first-order Markov model over the stage sequence, and logistic regression on subforum visit counts. Without these, the reported 7%-17% BLEU and 6%-13% ROUGE improvements may be relative to weak baselines and do not establish that the model captures meaningful health-stage dynamics. Additionally, although scores from 20 runs are reported as mean ± SD, no statistical significance tests (e.g., paired bootstrap, Wilcoxon signed-rank, or McNemar) are provided. Please add such baselines and significance tests.
  3. [Section V-A.3] Hyperparameter settings are severely under-specified. The paper states that 'Hyper-parameters were searched based on the highest scores achieved on the validation set' but does not report the search space or the final values for the number of GCN hops, LSTM layers, hidden dimensions, attention dimensions, keyword vocabulary size, or any other model-specific hyperparameters. It is also unclear whether the baselines were given the same hyperparameter search budget. Without this information, the experiments are not reproducible and the fairness of the comparison cannot be assessed. Please provide the full hyperparameter configuration and, ideally, release code to enable reproduction.
  4. [Section V-C] The interpretability claim is supported only by a qualitative case study of a single user's attention weights. The statement that the hierarchical attention 'greatly enhances the interpretability of the model' requires quantitative evidence, such as measuring agreement between attention weights and domain-expert-annotated subforum-stage associations, or an ablation that removes each attention level and reports both predictive performance and interpretability metrics. As presented, the interpretability analysis is anecdotal and does not substantiate the paper's central interpretability contribution.
minor comments (7)
  1. [Section V-C] The section heading 'Interpretablity Analysis' contains a typo; it should be 'Interpretability Analysis'.
  2. [Abstract] The phrase 'dynamic graph-to-sequence neural networks architecture' is grammatically awkward; it should be 'dynamic graph-to-sequence neural network architecture'.
  3. [Section III-A] The definition of an activity transition edge is not fully operational. Please specify precisely how 'most active forum' is computed (e.g., number of posts, total time, or a combination) and how months with no activity or ties between subforums are handled.
  4. [Section V-A] The construction of node feature vectors from 'the 100 top frequency keywords' is vague. Please clarify whether the features are binary indicators, term frequencies, TF-IDF scores, or some other encoding, and whether these features are normalized.
  5. [Section V-A] The split into 70% training, 10% validation, and 20% testing is described only as 'randomly selected.' Please state whether the split is at the user level, whether any users were excluded for having too short or too long histories, and how many users remain in each partition.
  6. [Section V-A] The footnote defining 'Dx' as 'Short for Oncotype DX test' is confusing in the context of the health stage vocabulary, where 'Dx' likely denotes diagnosis. Please clarify whether the stage token refers to the initial diagnosis stage or specifically to the Oncotype DX test, and place the footnote at the first occurrence of the vocabulary.
  7. [References] Reference [32] is a TensorFlow tutorial; please also cite the original NMT paper by Luong et al. (2015) from which the architecture is adapted.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DynGraph2Seq's empirical claims rest on held-out user splits, and its architectural components are standard cited building blocks; label noise is an external-validity concern, not a circular derivation.

full rationale

The paper's derivation chain is an empirical supervised-learning pipeline, not a formal derivation. The input is a dynamic graph of subforum activity (Section III-A, Definition 1) and the target is a sequence of health stages extracted from user signatures (Section V-A). Training, validation, and test users are disjoint (70/10/20 split), so the reported BLEU/ROUGE numbers are out-of-sample predictions and are not forced by construction. The dynamic graph encoder applies the cited Graph2Seq convolution per snapshot (Section IV-A) and the decoder uses conventional LSTM attention (Section IV-B); the new ingredients are the temporal LSTM over graph embeddings and the two-level attention mechanism. Equations (1)-(3) and the attention definitions are standard forward computations with learnable parameters, with no parameter fitted to the test labels or to a quantity that is later renamed as a prediction. Self-citations, including [28] (Graph2Seq as architectural base and as a baseline) and [14]/[25], are present but not load-bearing: the central comparison is an empirical result against a static Graph2Seq baseline, and the claimed improvement is not a logical consequence of the cited prior work. The signature-derived labels are self-reported and may be noisy, and the subforum names overlap semantically with the stage vocabulary, but these are data-validity and task-construction concerns; they do not make the model's output equivalent to its input by definition. No uniqueness theorem, ansatz smuggled via citation, fitted-input-as-prediction step, or self-definitional reduction was found.

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

The central claim rests on the reliability of signature-derived labels, the informativeness of subforum transitions, a specific graph discretization, and translation-style evaluation metrics. No new physical or mathematical entities are introduced.

free parameters (4)
  • Adam learning rate = 0.001
    Set in Section V-A3; selected by validation search.
  • Batch size = 50
    Set in Section V-A3.
  • Unreported hyperparameters (hidden sizes, number of GCN hops, LSTM layers, attention dimensions)
    Section V-A3 says hyperparameters were searched on the validation set but the values are not reported.
  • Keyword vocabulary size = 100
    Top 100 frequency keywords used to build node features (Section V-A). A modeling choice.
assumptions (4)
  • domain assumption Subforum transitions correlate with health stage changes
    Introduction and Figure 2 present this as the motivating premise; if false, the input representation is uninformative.
  • domain assumption User signatures provide accurate health stage ground truth
    Section V-A states the predicted sequences are validated against 'real health stage history extracted from users' signatures'; no noise model is given.
  • ad hoc to paper Monthly time windows and most-active-forum edges adequately represent user activity
    Section III-A defines activity transition graphs with a one-month window and transitions based on the most active forum; this is a paper-specific discretization.
  • domain assumption BLEU and ROUGE are valid metrics for stage-sequence agreement
    Section V-A1 adopts MT evaluation metrics for a categorical health stage sequence without justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DynGraph2Seq: Dynamic-Graph-to-Sequence Interpretable Learning for Health Stage Prediction in Online Health Forums." pith.science (2026). https://pith.science/paper/7XRP6VAZ

@misc{pith2026190808497,
  author       = {Pith},
  title        = {Pith review of: DynGraph2Seq: Dynamic-Graph-to-Sequence Interpretable Learning for Health Stage Prediction in Online Health Forums},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7XRP6VAZ}},
  note         = {Machine review of arXiv:1908.08497}
}
read the original abstract

Online health communities such as the online breast cancer forum enable patients (i.e., users) to interact and help each other within various subforums, which are subsections of the main forum devoted to specific health topics. The changing nature of the users' activities in different subforums can be strong indicators of their health status changes. This additional information could allow health-care organizations to respond promptly and provide additional help for the patient. However, modeling complex transitions of an individual user's activities among different subforums over time and learning how these correspond to his/her health stage are extremely challenging. In this paper, we first formulate the transition of user activities as a dynamic graph with multi-attributed nodes, then formalize the health stage inference task as a dynamic graph-to-sequence learning problem, and hence propose a novel dynamic graph-to-sequence neural networks architecture (DynGraph2Seq) to address all the challenges. Our proposed DynGraph2Seq model consists of a novel dynamic graph encoder and an interpretable sequence decoder that learn the mapping between a sequence of time-evolving user activity graphs and a sequence of target health stages. We go on to propose dynamic graph hierarchical attention mechanisms to facilitate the necessary multi-level interpretability. A comprehensive experimental analysis of its use for a health stage prediction task demonstrates both the effectiveness and the interpretability of the proposed models.

Figures

Figures reproduced from arXiv: 1908.08497 by the authors.

Figure 1
Figure 1. An example of patient signature that contains cancer diagnosis and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An example of user forum activities and the corresponding health [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The proposed end-to-end dynamic graph-to-sequence learning [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The proposed dynamic graph hierarchical attention mechanism: node [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: An example of learned dynamic graph hierarchical attention by [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Reinforcement Learning Based Graph-to-Sequence Model for Natural Question Generation

    cs.CL 2019-08 conditional novelty 6.0 of 10

    A reinforcement-learning graph-to-sequence model with answer-aware alignment reports new state-of-the-art question generation scores on SQuAD, with the gain partly explained by BERT embeddings and direct BLEU-4 optimization.

Reference graph

Works this paper leans on

36 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

    Breast cancer community,

    “Breast cancer community,” https://community.breastcancer.org/

    work page

  2. [2]

    American cancer society,

    “American cancer society,” http://www.cancer.org

    work page

  3. [3]

    ehealth forum,

    “ehealth forum,” http://ehealthforum.com

    work page

  4. [4]

    Characterizing the sublanguage of online breast cancer forums for medications, symptoms, and emotions,

    N. Elhadad, S. Zhang, P. Driscoll, and S. Brody, “Characterizing the sublanguage of online breast cancer forums for medications, symptoms, and emotions,” in AMIA Annual Symposium Proceedings , vol. 2014. American Medical Informatics Association, 2014, p. 516

    work page 2014

  5. [5]

    Does sustained participation in an online health community affect sentiment?

    S. Zhang, E. Bantum, J. Owen, and N. Elhadad, “Does sustained participation in an online health community affect sentiment?” in AMIA Annual Symposium Proceedings , vol. 2014, 2014, p. 1970

    work page 2014

  6. [6]

    Longitudinal analysis of discussion topics in an online breast cancer community using convolu- tional neural networks,

    S. Zhang, E. Grave, E. Sklar, and N. Elhadad, “Longitudinal analysis of discussion topics in an online breast cancer community using convolu- tional neural networks,” Journal of biomedical informatics , vol. 69, pp. 1–9, 2017

    work page 2017

  7. [7]

    Cancer stage prediction based on patient online discourse,

    M. Jha and N. Elhadad, “Cancer stage prediction based on patient online discourse,” in Proceedings of the 2010 Workshop on Biomedical Natural Language Processing, 2010, pp. 64–71

    work page 2010

  8. [8]

    Hierarchi- cal incomplete multi-source feature learning for spatiotemporal event forecasting,

    L. Zhao, J. Ye, F. Chen, C.-T. Lu, and N. Ramakrishnan, “Hierarchi- cal incomplete multi-source feature learning for spatiotemporal event forecasting,” in KDD 2016. ACM, 2016, pp. 2085–2094

    work page 2016

Show all 36 references
  1. [9]

    Multi-instance domain adaptation for vaccine adverse event detection,

    J. Wang and L. Zhao, “Multi-instance domain adaptation for vaccine adverse event detection,” in WWW 2018, 2018, pp. 97–106

    2018

  2. [10]

    Representation learning over dynamic graphs,

    R. Trivedi, M. Farajtbar, P. Biswal, and H. Zha, “Representation learning over dynamic graphs,” arXiv preprint arXiv:1803.04051 , 2018

    2018 arXiv

  3. [11]

    Dynamic network embedding by modeling triadic closure process

    L.-k. Zhou, Y . Yang, X. Ren, F. Wu, and Y . Zhuang, “Dynamic network embedding by modeling triadic closure process.” in AAAI, 2018

    2018

  4. [12]

    dyngraph2vec: Capturing network dynamics using dynamic graph representation learning,

    P. Goyal, S. R. Chhetri, and A. Canedo, “dyngraph2vec: Capturing network dynamics using dynamic graph representation learning,” arXiv preprint arXiv:1809.02657, 2018

    2018 arXiv

  5. [13]

    Multiple structure- view learning for graph classification,

    J. Wu, S. Pan, X. Zhu, C. Zhang, and S. Y . Philip, “Multiple structure- view learning for graph classification,” IEEE Transactions on Neural Networks and Learning Systems , vol. 29, no. 7, pp. 3236–3251, 2018

    2018

  6. [14]

    Scalable global alignment graph kernel using random features: From node embedding to graph embedding,

    L. Wu, I. E.-H. Yen, Z. Zhang, K. Xu, L. Zhao, X. Peng, Y . Xia, and C. Aggarwal, “Scalable global alignment graph kernel using random features: From node embedding to graph embedding,” in KDD 2019 , 2019, pp. 1418–1428

    2019

  7. [15]

    Sql-to-text generation with graph-to-sequence model,

    K. Xu, L. Wu, Z. Wang, Y . Feng, and V . Sheinin, “Sql-to-text generation with graph-to-sequence model,” arXiv preprint arXiv:1809.05255, 2018

    2018 arXiv

  8. [16]

    Exploiting rich syntactic information for semantic parsing with graph-to-sequence model,

    K. Xu, L. Wu, Z. Wang, M. Yu, L. Chen, and V . Sheinin, “Exploiting rich syntactic information for semantic parsing with graph-to-sequence model,” arXiv preprint arXiv:1808.07624 , 2018

    2018 arXiv

  9. [17]

    Neural machine translation by jointly learning to align and translate,

    D. Bahdanau, K. Cho, and Y . Bengio, “Neural machine translation by jointly learning to align and translate,” arXiv preprint arXiv:1409.0473, 2014

    2014 arXiv

  10. [18]

    Effective ap- proaches to attention-based neural machine translation,

    M.-T. Luong, H. Pham, and C. D. Manning, “Effective ap- proaches to attention-based neural machine translation,” arXiv preprint arXiv:1508.04025, 2015

    2015 arXiv

  11. [19]

    Hierarchical attention networks for document classification,

    Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, and E. Hovy, “Hierarchical attention networks for document classification,” in NAACL, 2016, pp. 1480–1489

    2016

  12. [20]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903 , vol. 1, no. 2, 2017

    2017 arXiv

  13. [21]

    Graph r-cnn for scene graph generation,

    J. Yang, J. Lu, S. Lee, D. Batra, and D. Parikh, “Graph r-cnn for scene graph generation,” arXiv preprint arXiv:1808.00191 , vol. 2, 2018

    2018 arXiv

  14. [22]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation,

    K. Cho, B. Van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using rnn encoder-decoder for statistical machine translation,” arXiv preprint arXiv:1406.1078, 2014

    2014 arXiv

  15. [23]

    Tree-to-sequence at- tentional neural machine translation,

    A. Eriguchi, K. Hashimoto, and Y . Tsuruoka, “Tree-to-sequence at- tentional neural machine translation,” arXiv preprint arXiv:1603.06075, 2016

    2016 arXiv

  16. [24]

    Improved semantic represen- tations from tree-structured long short-term memory networks,

    K. S. Tai, R. Socher, and C. D. Manning, “Improved semantic represen- tations from tree-structured long short-term memory networks,” arXiv preprint arXiv:1503.00075, 2015

    2015 arXiv

  17. [25]

    Deep graph translation,

    X. Guo, L. Wu, and L. Zhao, “Deep graph translation,” arXiv preprint arXiv:1805.09980, 2018

    2018 arXiv

  18. [26]

    Graphvae: Towards genera- tion of small graphs using variational autoencoders,

    M. Simonovsky and N. Komodakis, “Graphvae: Towards genera- tion of small graphs using variational autoencoders,” arXiv preprint arXiv:1802.03480, 2018

    2018 arXiv

  19. [27]

    Learning deep generative models of graphs,

    Y . Li, O. Vinyals, C. Dyer, R. Pascanu, and P. Battaglia, “Learning deep generative models of graphs,” arXiv preprint arXiv:1803.03324 , 2018

    2018 arXiv

  20. [28]

    Graph2seq: Graph to sequence learning with attention-based neural networks,

    K. Xu, L. Wu, Z. Wang, and V . Sheinin, “Graph2seq: Graph to sequence learning with attention-based neural networks,” arXiv preprint arXiv:1804.00823, 2018

    2018 arXiv

  21. [29]

    The fundamental advantages of temporal networks,

    A. Li, S. P. Cornelius, Y .-Y . Liu, L. Wang, and A.-L. Barab ´asi, “The fundamental advantages of temporal networks,” Science, vol. 358, no. 6366, pp. 1042–1046, 2017

    2017

  22. [30]

    Sequence to sequence learning with neural networks,

    I. Sutskever, O. Vinyals, and Q. V . Le, “Sequence to sequence learning with neural networks,” in NIPS 2014, 2014, pp. 3104–3112

    2014

  23. [31]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  24. [32]

    Neural machine translation (seq2seq) tutorial,

    M. Luong, E. Brevdo, and R. Zhao, “Neural machine translation (seq2seq) tutorial,” https://github.com/tensorflow/nmt, 2017

    2017

  25. [33]

    Feed-forward networks with atten- tion can solve some long-term memory problems,

    C. Raffel and D. P. Ellis, “Feed-forward networks with atten- tion can solve some long-term memory problems,” arXiv preprint arXiv:1512.08756, 2015

    2015 arXiv

  26. [34]

    Bleu: a method for automatic evaluation of machine translation,

    K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, “Bleu: a method for automatic evaluation of machine translation,” in ACL 2002 , 2002, pp. 311–318

    2002

  27. [35]

    Rouge: A package for automatic evaluation of summaries,

    C.-Y . Lin, “Rouge: A package for automatic evaluation of summaries,” Text Summarization Branches Out , 2004

    2004

  28. [36]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.