Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Rhythm of Opinion: A Hawkes-Graph Framework for Dynamic Propagation Analysis

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a Hawkes process whose dimensions are (hierarchy level, sentiment) can be coupled with a graph neural network to capture temporal, structural, and sentiment dynamics of online opinion propagation.

desk verdict A potentially useful new opinion-propagation dataset is attached to a methods paper whose experimental evaluation is underspecified and probably circular. read the letter →

arxiv 2504.15072 v1 pith:3777B2HD submitted 2025-04-21 cs.SI cs.CL

Yulong Li , Zhixiang Lu , Feilong Tang , Simin Lai , Ming Hu , Yuxuan Zhang , Haochen Xue , Zhaodong Wu
show 3 more authors
Imran Razzak Qingxia Li Jionglong Su
This is my paper · ORCID
classification cs.SIcs.CL
keywords opiniondynamicsHawkesprocessgraphneuralnetworksentimentanalysissocialmediahierarchicalcommentsVISTAdatasettemporalpoint
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that social-media opinion propagation can be modeled as one coupled system of comment timing, thread hierarchy, and sentiment, rather than as separate forecasting problems. It builds a high-dimensional Hawkes process whose dimensions are the combined (hierarchy level, sentiment category) pairs, and feeds the resulting intensities into a graph neural network that classifies new comments and predicts propagation edges. To support this, it introduces VISTA, a dataset of 159 trending topics with 47,207 posts, 327,015 second-level comments, and 29,578 third-level comments annotated with 11 sentiment categories. The paper reports that on VISTA the joint model establishes a baseline that improves with training data, reaching 26.99 percent test sentiment accuracy and 35.76 percent structural consistency at the 25 percent data split. A reader should care because the framework ties temporal, hierarchical, and sentiment dynamics into a single interpretable model.

What carries the argument

The load-bearing mechanism is the high-dimensional self-exciting Hawkes process whose dimensions are the Cartesian product of hierarchy level and sentiment, ω=(l,c). Its intensity is λ_ω(t)=μ_ω+Σ_{ω'} α_{ω,ω'} $e^{{-β_{ω,ω'}}$(t-t')}, where α_{ω,ω'} is the cross-excitation from dimension ω' to ω; integrating the intensity over [T,T+Δ] gives predicted comment counts. The normalized intensity per sentiment, q_c(v), converts Hawkes intensities into a probability distribution over sentiment classes at each node, which is what makes the GNN's node classification trainable. This single object carries the argument by turning a count-forecasting model into a generator of both structure (edges through intensity couplings) and sentiment supervision (soft labels), so the whole claim depends on how well the fitted Hawkes intensities approximate real comment generation.

What would settle it

Recompute sentiment prediction accuracy on the VISTA test set using the human/LLM-annotated sentiment labels as ground truth instead of the Hawkes-derived q_c(v), and compare with the reported values. If accuracy falls toward chance or drops sharply, the sentiment component is measuring how well the GNN reproduces the Hawkes fit rather than how well it predicts real opinions.

Watch

Extended reading notes

Core claim

The central claim is that the coupling of a multidimensional Hawkes process with a graph neural network is enough to capture how opinions propagate through threaded social-media discussions. The Hawkes process is defined over dimensions ω=(l,c), where l is the comment hierarchy level and c is one of 11 sentiment categories; its intensity λ_ω(t) gives the expected arrival rate of comments of that type, and the integral of the intensity over a future window yields the predicted comment count. The normalized intensity q_c(v)=λ_(l,c)(t)/Σ_(c')λ_(l,c')(t) is then treated as the sentiment distribution of a node and used both as a node feature and as the 'true sentiment label' in the sentiment classification loss. The GNN updates node embeddings by message passing with edge features that include the time gap and the Hawkes excitation strength, and it is trained with a combined sentiment and structure loss. On the VISTA dataset, the authors report that this architecture predicts future comment volume, sentiment class, and parent-child structure, and they offer the results as a baseline.

Load-bearing premise

The framework assumes that the sentiment distribution q_c(v) computed from the fitted Hawkes intensity is a valid stand-in for the true sentiment of a comment, and that the same distribution can serve as the ground-truth label when measuring sentiment accuracy.

Editorial extensions

If this is right

  • On VISTA, the joint model's accuracy rises with the fraction of data used: test sentiment accuracy goes from 18.31 percent at a 15 percent split to 26.99 percent at a 25 percent split, with structural consistency rising from 21.22 to 35.76 percent.
  • The VISTA dataset itself provides a reusable resource: 159 topics with three-level comment trees and 11-category sentiment labels, which the paper positions as a benchmark for future opinion-dynamics models.
  • The framework yields interpretable quantities: per (hierarchy, sentiment) intensity functions, cross-excitation strengths α_{ω,ω'}, and predicted future comment counts, linking sentiment propagation to the comment tree and time.
  • Because the GNN predicts both node sentiment and parent-child edges, the method can generate a predicted future graph of a discussion, not just a count of future comments.
  • When training data is increased from 15 to 25 percent, structural consistency prediction accuracy improves more steeply than sentiment accuracy, suggesting that the model's structural component benefits more from additional comment-chain examples.

Reading between the lines

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

  • If the sentiment labels were replaced by independent human annotations, the GNN component could be tested as a genuine sentiment classifier, which would reveal whether the Hawkes-derived soft labels actually carry signal beyond the comment text itself.
  • The same (hierarchy, sentiment) dimensional pairing could be extended to other discrete attributes such as topic, stance, or user role, turning the architecture into a general template for multidimensional propagation modeling.
  • The VISTA annotations, produced by a large language model annotator with human spot checks, could themselves be evaluated for label noise; using the Hawkes soft labels in a teacher-student setup with human labels is a natural next test.
  • Because the Hawkes intensity integrates to predicted comment counts, the same model could be used to simulate counterfactual discussions under different excitation strengths, for instance to study how cross-topic influence changes opinion trajectories.
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 / 5 minor

Summary. The paper proposes a hybrid framework for modeling opinion propagation on social media, combining a high-dimensional Hawkes process with a graph neural network (GNN). The Hawkes process models the intensity of comments across (hierarchy level, sentiment) pairs, and its fitted intensities are used both as node features and as the training target for the GNN's sentiment classifier. The paper also introduces VISTA, a new Weibo dataset of 159 trending topics with more than 500,000 hierarchically nested comments annotated with 11 sentiment categories. The experimental section reports sentiment prediction accuracy (SA) and structural consistency accuracy (SCA) for the proposed model at three data proportions (15%, 20%, 25%). The authors claim that the method captures temporal evolution, structural changes, and sentiment diffusion, and that it provides a solid baseline for future studies.

Significance. If the empirical claims were sound, the paper would offer two contributions: a large, multi-level, sentiment-annotated dataset (VISTA) that is genuinely useful for opinion-dynamics research, and a plausible architectural idea of coupling Hawkes intensities with GNN message passing. However, the current evaluation is not sufficient to support the paper's central claims. The sentiment evaluation is potentially circular, the experimental design lacks baselines and statistical rigor, and the temporal prediction component is never evaluated. The dataset alone is valuable, but the paper's method-oriented claims require a substantially reworked experimental section before the work can be considered reliable.

major comments (4)
  1. [Sec. 5, Eq. (14); Sec. 6] The sentiment evaluation is circular. Eq. (14) defines L_sentiment using q_c(v) from Eq. (11), the normalized Hawkes intensity, and explicitly calls it 'the true sentiment label of node v.' However, Sec. 6 defines Acc_sentiment using 'the true sentiment label c(v)' without specifying whether c(v) is the GLM-4 annotation from Sec. 3 or the Hawkes-derived q_c(v). If c(v) = q_c(v), then the GNN is trained and evaluated against its own input features, so the SA numbers in Table 3 measure self-consistency, not predictive accuracy for human/GLM sentiment; if c(v) is the GLM annotation, then the training target in Eq. (14) differs from the evaluation target, and the model was never trained to optimize the reported metric. The manuscript does not resolve this ambiguity, so the headline claim of a robust sentiment-prediction baseline is unsupported.
  2. [Sec. 7.2, Table 3] The experimental comparison is insufficient. Table 3 reports only the proposed model at three data proportions, with no baselines such as a Hawkes-only model, a GNN-only model, standard time-series forecasting, or a majority-class predictor. There are also no error bars or significance tests. Consequently, the conclusion that 'a larger dataset effectively enhances the model's predictive capabilities' and the claim of establishing a 'solid baseline' are not supported by the evidence.
  3. [Sec. 5, Eq. (10); Sec. 7] The temporal prediction component is never evaluated. Eq. (10) defines the predicted comment count \hat{N}_\omega, but the experiments report only SA and SCA. No metric (e.g., MAE, NLL) is given for how well the Hawkes process predicts future comment volumes, despite the paper's stated goal of predicting 'when new comments will emerge and how the volume of comments will change over time.' The claim of capturing temporal evolution is therefore without empirical support.
  4. [Sec. 9.4] Appendix 9.4 is a verbatim copy of Appendix 9.3 and derives the gradient with respect to α_{ω,ω'}, not β_{ω,ω'}. The main text says 'For a detailed mathematical proof, refer to Appendix Section 9.4' for the β gradient, but that derivation is absent. The MLE training procedure is thus incompletely specified.
minor comments (5)
  1. [Sec. 2, para. 4] The phrase 'deep learning algorithm such as' should be 'deep learning algorithms such as'.
  2. [Fig. 2 caption] The caption contains a typo: 'VISTA Dateset' should be 'VISTA Dataset'.
  3. [Sec. 7.2, Table 3] The term 'data proportions' is not defined; the text says 15%, 20%, and 25% but does not state the base (e.g., fraction of training topics or comments) or how the split is constructed.
  4. [Sec. 5, Eq. (12)] The notation W_2 e_{uv} h_u is not dimensionally clear; if e_{uv} is a vector, the operation should be defined explicitly (e.g., concatenation or a tensor product followed by a linear layer).
  5. [References] The reference to 'De et al., 2016' is incomplete; the title is missing and the entry currently reads only 'c. Advances in neural information processing systems, 29.'

Circularity Check

1 steps flagged · score 6.0 of 10

Sentiment classification reduces to the fitted Hawkes output: Eq. 14 defines q_c(v), the normalized Hawkes intensity, as the 'true sentiment label,' and Section 6 never specifies whether the evaluated c(v) is that Hawkes label or the GLM annotation.

  1. self definitional [Section 5, Eq. (11) and Eq. (14); Section 6 (Evaluation Metrics).]
    "Each node’s features are composed of the intensity function values λω(t) generated by the Hawkes process and the sentiment distribution qc(v). ... qc(v) = λ(l,c)(t)/∑_{c′=1}^{C} λ(l,c′)(t). ... L_sentiment = −∑_{v∈Vpred} ∑_{c=1}^{C} qc(v)·lnP(c|v) ... where ... qc(v) is the normalized sentiment distribution output from the Hawkes process, representing the true sentiment label of node v."

    By Eq. (11), q_c(v) is just the normalized Hawkes intensity, i.e., an output of the model fitted to (level, sentiment) event counts. By Eq. (14), the GNN's sentiment training target is this same q_c(v), explicitly called 'the true sentiment label of node v.' The same q_c(v) is also listed among the node features, so the loss can be minimized by copying an input feature. Section 6 defines Acc_sentiment against 'the true sentiment label c(v)' without stating whether c(v) is the GLM-4 annotation or the Hawkes-derived q_c(v). If it is q_c(v), the reported SA measures how well the GNN reproduces the fitted Hawkes output, not human sentiment.

full rationale

The Hawkes count prediction itself is self-contained: MLE of Eq. (2) on observed event times is a standard, non-circular fitting procedure, and the VISTA dataset is an independent contribution. The circularity is confined to the sentiment-classification arm. Eq. (14) takes the normalized Hawkes intensity (Eq. 11) as the supervised target, while the same quantity is a node feature, so the sentiment 'prediction' reduces by construction to reproducing the fitted Hawkes process. Section 6's metric does not resolve whether the true label is the GLM annotation or this Hawkes output, so the headline claim of accurate sentiment classification is supported only in the circular reading or is untrained in the non-circular reading. No self-citation chain or imported uniqueness theorem is load-bearing. Because the central sentiment-prediction claim is partially forced by this definitional loop, the score is 6 rather than 0-2.

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

The model relies on a standard Hawkes MLE framework, with fitted intensity (mu), excitation (alpha), and decay (beta) parameters for 33 dimensions, plus loss hyperparameters. It assumes comment independence and noise-free collection (stated in Limitations), and depends on GLM-4 annotations as ground truth. The GNN's use of Hawkes-derived sentiment distributions as targets is an ad hoc choice that creates circularity.

free parameters (6)
  • µω baseline intensities = MLE estimates (33 values)
    Fitted via maximum likelihood using Eq. 2; one per (hierarchy, sentiment) dimension.
  • αω,ω′ excitation strengths = MLE estimates (33×33 values)
    Fitted via maximum likelihood using Eq. 2; measure cross-dimension triggering.
  • βω,ω′ decay coefficients = MLE estimates (33×33 values)
    Fitted via maximum likelihood using Eq. 2; gradient derivation promised in Appendix 9.4 is missing and the section duplicates 9.3.
  • λ1, λ2 loss weights = Not specified
    Hyperparameters in Eq. 16 balancing sentiment and structural losses; no tuning procedure given.
  • Hierarchy levels (L=3) and sentiment categories (C=11) = Fixed to 3 and 11
    Design choices based on the VISTA annotation scheme; they bound the dimension space and are not derived from data.
  • Exponential decay kernel = φ(τ)=e^{-βτ}
    Choice of kernel in Eq. 1; exponential decay is convenient but not compared to other kernels.
assumptions (5)
  • standard math Hawkes process log-likelihood is valid and MLE is consistent
    Used in Section 5 and Appendix; standard background for parameter estimation.
  • domain assumption Comment events are conditionally independent given the intensity
    Explicitly stated as an assumption in the Limitations section; needed for the factorization in Eq. 2.
  • domain assumption Data collection is essentially noise-free
    Stated in Limitations; however Appendix 9.1 admits residual sensitive or misleading content may remain, creating tension.
  • domain assumption GLM-4-plus automated labels are a reliable ground truth
    Cohen's Kappa 0.85 reported for ten spot checks, but sample sizes, disagreement handling, and human annotation subset are not detailed.
  • ad hoc to paper The Hawkes-derived q_c(v) can serve as the GNN sentiment target
    Eqs. 11 and 14 substitute the fitted intensity ratio for the true sentiment label, which is a circular modeling choice unique to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rhythm of Opinion: A Hawkes-Graph Framework for Dynamic Propagation Analysis." pith.science (2026). https://pith.science/paper/3777B2HD

@misc{pith2026250415072,
  author       = {Pith},
  title        = {Pith review of: Rhythm of Opinion: A Hawkes-Graph Framework for Dynamic Propagation Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3777B2HD}},
  note         = {Machine review of arXiv:2504.15072}
}
read the original abstract

The rapid development of social media has significantly reshaped the dynamics of public opinion, resulting in complex interactions that traditional models fail to effectively capture. To address this challenge, we propose an innovative approach that integrates multi-dimensional Hawkes processes with Graph Neural Network, modeling opinion propagation dynamics among nodes in a social network while considering the intricate hierarchical relationships between comments. The extended multi-dimensional Hawkes process captures the hierarchical structure, multi-dimensional interactions, and mutual influences across different topics, forming a complex propagation network. Moreover, recognizing the lack of high-quality datasets capable of comprehensively capturing the evolution of public opinion dynamics, we introduce a new dataset, VISTA. It includes 159 trending topics, corresponding to 47,207 posts, 327,015 second-level comments, and 29,578 third-level comments, covering diverse domains such as politics, entertainment, sports, health, and medicine. The dataset is annotated with detailed sentiment labels across 11 categories and clearly defined hierarchical relationships. When combined with our method, it offers strong interpretability by linking sentiment propagation to the comment hierarchy and temporal evolution. Our approach provides a robust baseline for future research.

Figures

Figures reproduced from arXiv: 2504.15072 by the authors.

Figure 1
Figure 1. Opinion Dynamics Modeling illustrating two [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustrative Diagram of VISTA Dateset.The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the Social Network Sentiment Analysis Model. Our framework analyzes sentiment in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 23 canonical work pages

  1. [1]

    Atif Aliak. 2025. https://www.kaggle.com/datasets/atifaliak/youtube-comments-dataset Youtube comments dataset . https://www.kaggle.com/datasets/atifaliak/youtube-comments-dataset. Accessed: 2025-02-12

    work page 2025

  2. [2]

    RM Anderson. 1991. Infectious diseases of humans: dynamics and control. Oxford University Press

    work page 1991

  3. [3]

    Alex Arenas, Albert D \' az-Guilera, Jurgen Kurths, Yamir Moreno, and Changsong Zhou. 2008. Synchronization in complex networks. Physics reports, 469(3):93--153

    work page 2008

  4. [4]

    Albert-L \'a szl \'o Barab \'a si and R \'e ka Albert. 1999. Emergence of scaling in random networks. science, 286(5439):509--512

    1999

  5. [5]

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung. 2015. Time series analysis: forecasting and control. John Wiley & Sons

    2015

  6. [6]

    Yan Cao, Yihong Dong, Shaoqing Wu, Yu Xin, and Jiangbo Qian. 2019. Dynamic network embedding for link prediction. In 2019 IEEE Intl Conf on Parallel & Distributed Processing with Applications, Big Data & Cloud Computing, Sustainable Computing & Communications, Social Computing & Networking (ISPA/BDCloud/SocialCom/SustainCom), pages 920--927. IEEE

    work page 2019

  7. [7]

    Zhihao Chen, Bingbing Xu, Tiecheng Cai, Zhou Yang, and Xiangwen Liao. 2023. A dynamic emotional propagation model over time for competitive environments. Electronics, 12(24):4937

    work page 2023

  8. [8]

    Matteo Cinelli, Walter Quattrociocchi, Alessandro Galeazzi, Carlo Michele Valensise, Emanuele Brugnoli, Ana Lucia Schmidt, Paola Zola, Fabiana Zollo, and Antonio Scala. 2020. The covid-19 social media infodemic. Scientific reports, 10(1):1--10

    work page 2020

Show all 34 references
  1. [9]

    Michael Conover, Jacob Ratkiewicz, Matthew Francisco, Bruno Gon c alves, Filippo Menczer, and Alessandro Flammini. 2011. Political polarization on twitter. In Proceedings of the international aaai conference on web and social media, volume 5, pages 89--96

    2011

  2. [10]

    Abir De, Isabel Valera, Niloy Ganguly, Sourangshu Bhattacharya, and Manuel Gomez Rodriguez. 2016. c. Advances in neural information processing systems, 29

    2016

  3. [11]

    Joshua M Epstein. 2012. Generative social science: Studies in agent-based computational modeling. Princeton University Press

    2012

  4. [12]

    O nder G \

    \"O nder G \"u rcan. 2024. Llm-augmented agent-based modelling for social simulations: Challenges and opportunities. HHAI 2024: Hybrid Human AI Systems for the Social Good, pages 134--144

    2024

  5. [13]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems, 30

    2017

  6. [14]

    Zhiwei Jin, Juan Cao, Yu-Gang Jiang, and Yongdong Zhang. 2014. News credibility evaluation on microblog with a hierarchical propagation model. In 2014 IEEE international conference on data mining, pages 230--239. IEEE

    2014

  7. [15]

    Elihu Katz, Paul F Lazarsfeld, and Elmo Roper. 2017. Personal influence: The part played by people in the flow of mass communications. Routledge

    2017

  8. [16]

    Seyed Mehran Kazemi, Rishab Goel, Kshitij Jain, Ivan Kobyzev, Akshay Sethi, Peter Forsyth, and Pascal Poupart. 2020. Representation learning for dynamic graphs: A survey. Journal of Machine Learning Research, 21(70):1--73

    2020

  9. [17]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

    2016 arXiv

  10. [18]

    Antonio Longa, Veronica Lachi, Gabriele Santin, Monica Bianchini, Bruno Lepri, Pietro Lio, Franco Scarselli, and Andrea Passerini. 2023. Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities. arXiv preprint arXiv:2302.01018

    2023 arXiv

  11. [19]

    Goran Muri \'c , Alexey Tregubov, Jim Blythe, Andr \'e s Abeliuk, Divya Choudhary, Kristina Lerman, and Emilio Ferrara. 2022. Large-scale agent-based simulations of online social networks. Autonomous Agents and Multi-Agent Systems, 36(2):38

    2022

  12. [20]

    Mark Newman, Albert-L \'a szl \'o Barab \'a si, and Duncan J Watts. 2011. The structure and dynamics of networks. Princeton university press

    2011

  13. [21]

    Elisabeth Noelle-Neumann. 1974. The spiral of silence a theory of public opinion. Journal of communication, 24(2):43--51

    1974

  14. [22]

    Sergey E Parsegov, Anton V Proskurnikov, Roberto Tempo, and Noah E Friedkin. 2016. Novel multidimensional models of opinion dynamics in social networks. IEEE Transactions on Automatic Control, 62(5):2270--2285

    2016

  15. [23]

    Bhavtosh Rath, Xavier Morales, and Jaideep Srivastava. 2021. Scarlet: explainable attention based graph neural network for fake news spreader prediction. In Pacific-Asia conference on knowledge discovery and data mining, pages 714--727. Springer

    2021

  16. [24]

    Kai Shu, Deepak Mahudeswaran, Suhang Wang, Dongwon Lee, and Huan Liu. 2018. Fakenewsnet: A data repository with news content, social context and dynamic information for studying fake news on social media. arXiv preprint arXiv:1809.01286

    2018 arXiv

  17. [25]

    Jos \'e F Torres, Dalil Hadjout, Abderrazak Sebaa, Francisco Mart \' nez- \'A lvarez, and Alicia Troncoso. 2021. Deep learning for time series forecasting: a survey. Big Data, 9(1):3--21

    2021

  18. [26]

    Twitter. 2021. https://www.kaggle.com/datasets/jp797498e/twitter-entity-sentiment-analysis Twitter dataset for sentiment analysis . https://www.kaggle.com/datasets/jp797498e/twitter-entity-sentiment-analysis. Accessed: 2025-02-12

    2021

  19. [27]

    Duncan J Watts and Steven H Strogatz. 1998. Collective dynamics of ‘small-world’networks. nature, 393(6684):440--442

    1998

  20. [28]

    Weibo. n.d. https://www.weibo.com Weibo . Accessed: 2024-12-30

    2024

  21. [29]

    Jiyoung Woo, Jaebong Son, and Hsinchun Chen. 2011. An sir model for violent topic diffusion in social media. In Proceedings of 2011 IEEE International Conference on Intelligence and Security Informatics, pages 15--19. IEEE

    2011

  22. [30]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4--24

    2020

  23. [31]

    Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchical attention networks for document classification. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language...

    2016

  24. [32]

    Fulian Yin, Xueying Shao, Meiqi Ji, and Jianhong Wu. 2020. Quantify influence of delay in opinion transmission of opinion leaders on covid-19 information propagation in the chinese sina-microblog. arXiv preprint arXiv:2011.06797

    2020 arXiv

  25. [33]

    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...

  26. [34]

    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 16, 2026 · model on record in the stance chip above.