Pith. sign in

REVIEW 5 major objections 5 minor 54 references

Conversation Kernels: A Flexible Mechanism to Learn Relevant Context for Online Conversation Understanding

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

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

Pith's one-line read Context-augmented conversation kernels — small, learned windows of nearby comments — outperform transformer baselines and zero-shot GPT-4 on four Slashdot comment-labeling tasks, with absolute accuracy gains up to 20 percentage points.

desk verdict Useful proof-of-concept for learned context-window selection, but the headline gains are inflated by weak baselines and a heavily constrained retrieval universe. read the letter →

arxiv 2505.20482 v1 pith:J4QI42ZI submitted 2025-05-26 cs.CL cs.AI

Vibhor Agarwal , Arjoo Gupta , Suparna De , Nishanth Sastry This is my paper
classification cs.CLcs.AI
keywords conversationkernelstreescontextretrievalcommentlabelingSlashdotRoBERTaGPT-4comparisononlineunderstanding
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 tries to establish that the right way to understand a comment in an online thread is to first retrieve a small set of nearby comments from the conversation tree, then classify the comment together with that retrieved context. It proposes 'conversation kernels,' two families of hand-designed window shapes — ancestors, siblings, and children; or one- and two-hop neighbors — that select up to three comments around a target post, and a two-module model that marginalizes label predictions over retrieved windows. On four Slashdot comment-labeling tasks ('funny', 'informative', 'insightful', 'interesting'), the paper reports that context-augmented kernels beat LSTM, BERT, RoBERTa (with and without parent context), and zero-shot GPT-3.5 and GPT-4, with absolute accuracy gains up to 20% and macro-F1 gains up to 19%. The claim matters because it suggests a cheap, reusable mechanism — learn which part of a conversation is relevant for a given task — can substitute for much larger models and task-specific architectures.

What carries the argument

The central object is the conversation kernel, a mechanism that maps a target post to a small context window in the conversation tree. Two families are defined: ancestor/sibling/children windows, which take up to $L=3$ ancestors (starting from the parent), siblings, or children of the target, and one-hop/two-hop neighborhood windows, which take the first $L=3$ direct or two-step neighbors ordered by timestamp (the window size is fixed after a sweep over 2–10). The architecture couples a context retriever $p(w|x)=\mathrm{softmax}_w(\mathrm{Embed}_{\mathrm{comment}}(x)^\top \mathrm{Embed}_{\mathrm{window}}(w))$ with a context-augmented encoder $p(y|w,x)$ built on RoBERTa; the full prediction is the marginalization $p(y|x)=\sum_{w\in W} p(y|w,x)\,p(w|x)$. The retriever's job is to weight each window by its relevance to the task, and the encoder classifies the comment given that window; both are trained end-to-end, so the model learns which context matters per task.

What would settle it

An oracle test would settle it: for a random sample of labeled comments, ask annotators which other comments (or the article root) are needed to justify the label, then measure how often all necessary context falls within the model's $L=3$ windows. If a substantial share of comments need context from beyond those windows, and those comments are misclassified while a full-context model gets them right, the fixed-window assumption is the binding constraint.

Watch

Extended reading notes

Core claim

The central discovery is that a task-flexible context-retrieval step, rather than a bigger or more expressive encoder, is what drives performance in conversation understanding. Formally, the paper defines the prediction for a target comment $x$ as $p(y|x)=\sum_{w\in W} p(y|w,x)\,p(w|x)$, where $W$ is a small set of windows, each of up to $L=3$ comments drawn from one of two kernel shapes: the ancestor/sibling/children windows or the one-hop/two-hop neighborhood windows. The retriever scores each window by the inner product of RoBERTa embeddings of the target and the window (softmax over windows), and the encoder concatenates $x$ with the window separated by [SEP] tokens and classifies with RoBERTa plus an MLP. The paper reports that this architecture outperforms strong transformer baselines and zero-shot GPT-3.5 and GPT-4 on all four Slashdot label tasks, and that different kernel shapes win for different tasks — ancestor/sibling/children windows win for insightful, informative, and interesting; one-hop/two-hop windows win for funny — which the authors take as evidence that the mechanism is learning task-relevant context rather than relying on a single fixed neighborhood definition.

Load-bearing premise

The approach assumes that the context needed to judge a comment always sits within three immediately surrounding comments in the reply tree, so labels whose deciding context lies farther up the thread or spread across many posts cannot be captured.

Editorial extensions

If this is right

  • The same architecture, with no task-specific engineering, beats the best baseline for all four labels; the winning kernel family differs by label, which the authors read as evidence the retriever learns task-relevant context.
  • The model trained on 2014–2022 data holds its accuracy on an unseen January–November 2023 Slashdot snapshot, indicating the learned context-retrieval transfers across time.
  • Conversation kernels outperform zero-shot GPT-3.5 and GPT-4 even when the LLMs receive the entire conversation (truncated only by the token limit), by 10–15% on the insightful, informative, and interesting tasks.
  • Because Reddit and X conversations also form reply trees, the paper expects the same mechanism to transfer to those platforms.
  • The kernel-shape families are explicitly proof-of-concept; designing new shapes or mixing existing windows is the paper's stated next step.

Reading between the lines

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

  • Beyond the paper: the $L=3$ window cap is a hard ceiling; a retriever that can attend over the whole thread or learn variable-size windows would test whether long-range context (e.g., the root article) carries label signal the current design cannot reach.
  • Beyond the paper: the LLM comparison is zero-shot and run on a 10% sample with long conversations truncated, so the claim that LLMs cannot find the right context is protocol-specific; fine-tuned LLMs or retrieval-augmented prompts using the same windows could close the gap.
  • Beyond the paper: community subjectivity — flagged in the paper's ethical statement for 'funny' labels on Slashdot — means the 'general-purpose' claim may require retraining the retriever per community or per platform rather than transferring one kernel unchanged.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes Conversation Kernels, a two-module architecture for classifying comments in tree-structured online conversations. A context retriever scores a small set of fixed-size windows taken from two families of tree neighborhoods (ancestors/siblings/children and one-/two-hop neighbors) via an inner-product softmax, and a RoBERTa-based encoder classifies the target comment conditioned on the chosen window, marginalizing over windows as in Eq. (1). The authors crawl Slashdot (2014-2022, 509,669 comments; 70,316 labeled), train binary classifiers for 'funny', 'informative', 'insightful', and 'interesting', and report accuracy and macro-F1 improvements over LSTM/BERT/RoBERTa baselines, as well as over zero-shot GPT-3.5/4. A separate 2023 snapshot is used to claim temporal generalizability.

Significance. The idea of making context selection task-dependent is sensible, and the dataset release and end-to-end formulation are useful contributions. The paper is also transparent about several limitations (non-tree graphs, early conversation cold start, multilingual settings) and ships code and data. However, the central quantitative claims currently rest on comparisons that are not apples-to-apples: different test sets for the LLM comparison, truncated transformer baselines, and test-set-based selection of the kernel family. If the authors re-run the evaluation in a matched, held-out protocol and add an analysis of the retrieval search space, the result could be a solid empirical contribution; as it stands, the magnitude of the claimed gains is not credible.

major comments (5)
  1. [Table 1 / Section 5.5] The GPT-3.5/4 comparison is not made on the CK test split: the footnote to Table 1 states that the GPT results are based on a random 10% stratified sample of the entire dataset, whereas the CK results are from the 10% test split described in Section 5.1. Because the two evaluations use different sets of comments, the abstract's claim that conversation kernels outperform GPT-4 is not established. In addition, the LLMs are prompted zero-shot while CK is trained on labeled examples, so the comparison conflates model quality with training supervision. Please evaluate GPT-3.5/4 on the same test split used for CK (or report paired results on the overlap) and state clearly that this is a zero-shot comparison.
  2. [Section 5.2 vs Section 3.2] The BERT and RoBERTa baselines are run with maximum sequence length 75, but the average comment length in the corpus is 99.38 tokens (Section 3.2), and the CK encoder receives the target comment plus up to L context comments. The baselines are therefore systematically truncated, and the reported gains of up to 20% may be largely a truncation artifact. Please re-run the transformer baselines with a maximum length that covers the full comments (for example 128 or 256) using the same tokenizer and preprocessing as the CK encoder, and report the updated numbers.
  3. [Section 5.3] The kernel family is selected per task from the test-set results: ancestor-sibling-child windows are reported for insightful, informative, and interesting, while one/two-hop windows are reported for funny. No validation-based selection or multiple-comparison correction is described, so the reported 'best family' performance is optimistically biased. Please choose the kernel family on the validation split before computing test numbers, or report both families for every task together with confidence intervals and significance tests.
  4. [Sections 4.1 / 5.1 / 6.1] The retriever can only choose among fixed windows of at most L=3 comments (ancestor/sibling/children or one/two-hop neighborhoods, selected by timestamp), and Section 6.1 states that non-local windows are not tested. The paper provides no measurement of how often the label-relevant context falls outside this search space. Because the central claim is that the model learns 'the right context,' the current evidence only shows that local windows help on Slashdot. Please add an oracle or upper-bound experiment with an unconstrained retriever, or an analysis of the distance distribution of relevant context, to bound what the fixed kernel shapes can capture.
  5. [Section 5.4 / Table 2] The temporal generalizability claim is not supported by Table 2, which reports only CK's absolute accuracy and macro-F1 on the Jan-Nov 2023 snapshot. There is no baseline on the same 2023 data and no statistical comparison with the original test-set numbers, so the statement that the model 'performs just as well' is not demonstrated. Please include matched baselines on the 2023 snapshot and a significance test comparing temporal performance.
minor comments (5)
  1. [Abstract / Section 5.3] In the abstract and Section 5.3, the reported gains of 'up to 20%' and 'up to 19%' should be labeled as absolute percentage points and tied to the specific baseline and task; as written, the numbers are ambiguous and appear inconsistent with several entries in Table 1.
  2. [Section 5.2] The LSTM baseline is trained with mean squared error loss for a binary classification task; standard cross-entropy would be a fairer and more conventional choice.
  3. [Section 6] Section 6 states that 'an attention mechanism is used to give additional weight to those that are more relevant,' but no attention mechanism over comments is described in Section 4; Eq. (2) is a softmax over windows. Please align the conclusion with the model description.
  4. [Figure 6 / Section 5.5] The prompts in Figure 6 do not specify the output format or parsing procedure for GPT-3.5/4 responses; please describe how free-text responses were mapped to labels and whether any invalid responses were discarded.
  5. [Section 5.1 / Table 1] No confidence intervals or multiple-seed results are reported for any method; given that the test set is a single 10% split, the reader cannot assess the stability of the ranking between CK and the baselines.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Equation (1) defines a genuine latent-variable model whose retriever and encoder are jointly trained on an 80% split and evaluated on held-out and future-time-period data; the reported gains are held-out measurements, and self-citations are background only.

full rationale

The paper's central prediction equation p(y|x) = sum_w p(y|w,x)p(w|x) (Eq. 1) is a genuine latent-variable marginalization in which both the retrieval distribution p(w|x) and the context-augmented encoder p(y|w,x) are trained jointly on an 80% training split and evaluated on held-out conversations (Section 5.1) plus a previously unseen Jan-Nov 2023 snapshot (Section 5.4). No quantity reported as a 'prediction' is a fitted constant from the test set: the retriever weights and encoder parameters are learned on the training data, and the reported accuracy/F1 gains over BERT, RoBERTa, LSTM and GPT-3.5/4 are measured on examples the model has not seen. The kernel shapes (ancestor/sibling/children, 1-hop/2-hop) and the window size L=3 are design choices justified empirically by an internal sweep, not imported from a self-citation, and no cited uniqueness theorem does load-bearing work. Self-citations (e.g., Agarwal et al. for graph-based conversation modeling and conversation-tree notation) are background references; the conversation-tree construction here follows directly from Slashdot's parent IDs. The concerns that do arise — the retriever can only choose among timestamp-truncated windows of at most L=3 comments, the LLM comparison truncates long conversations to the 8192-token limit, and the per-task selection of the best kernel family — are evaluation-fairness and generalization-bound issues, not circularity: the model's outputs are not equivalent to its inputs by construction. The limitations section openly concedes that non-local context and non-tree (graph) conversation structures are untested, which further confirms that the paper claims less than a derived universal mechanism and does not smuggle its conclusion into its assumptions.

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

No physical entities are postulated, so invented_entities is empty. The model itself relies on the assumptions listed above.

free parameters (4)
  • Context window size L = 3
    Chosen by sweeping L from 2 to 10 and taking the best value, as reported in Section 5.1. This caps how much context the model can ever see.
  • Kernel family per task = ancestor-sibling-child for insightful, informative, interesting; 1-hop/2-hop for funny
    The best family is selected separately for each task after seeing test-set results in Table 1, which is a post hoc model selection that inflates reported performance.
  • Balanced class sampling ratio = 1:1
    Negative examples are sampled from other Slashdot categories to balance each binary task, defining the evaluation distribution in Section 5.1.
  • Training epochs = 3
    The paper fixes three epochs without reporting early stopping or validation-based epoch selection; results may depend on this choice.
assumptions (4)
  • domain assumption Slashdot user-applied tags are treated as ground truth labels for funny, informative, insightful, and interesting.
    The model is trained and evaluated against community tags with no annotation-quality or inter-rater reliability analysis (Section 3).
  • domain assumption The reconstructed conversation tree faithfully represents reply structure.
    The crawler extracts comment IDs and parent IDs from HTML, so any error in the reply-to structure would misplace nodes in kernel windows (Section 3.1).
  • domain assumption The first L comments by timestamp in a window are an adequate representative sample of that window's context.
    Section 4.1 fixes the window size L and truncates by timestamp, with no analysis of whether later or random nodes would be better context.
  • ad hoc to paper The two hand-designed kernel shape families cover the relevant context for all four tasks.
    The paper describes these shapes as proof-of-concept and provides no principle for why they suffice for all tasks (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conversation Kernels: A Flexible Mechanism to Learn Relevant Context for Online Conversation Understanding." pith.science (2026). https://pith.science/paper/J4QI42ZI

@misc{pith2026250520482,
  author       = {Pith},
  title        = {Pith review of: Conversation Kernels: A Flexible Mechanism to Learn Relevant Context for Online Conversation Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4QI42ZI}},
  note         = {Machine review of arXiv:2505.20482}
}
read the original abstract

Understanding online conversations has attracted research attention with the growth of social networks and online discussion forums. Content analysis of posts and replies in online conversations is difficult because each individual utterance is usually short and may implicitly refer to other posts within the same conversation. Thus, understanding individual posts requires capturing the conversational context and dependencies between different parts of a conversation tree and then encoding the context dependencies between posts and comments/replies into the language model. To this end, we propose a general-purpose mechanism to discover appropriate conversational context for various aspects about an online post in a conversation, such as whether it is informative, insightful, interesting or funny. Specifically, we design two families of Conversation Kernels, which explore different parts of the neighborhood of a post in the tree representing the conversation and through this, build relevant conversational context that is appropriate for each task being considered. We apply our developed method to conversations crawled from slashdot.org, which allows users to apply highly different labels to posts, such as 'insightful', 'funny', etc., and therefore provides an ideal experimental platform to study whether a framework such as Conversation Kernels is general-purpose and flexible enough to be adapted to disparately different conversation understanding tasks.

Figures

Figures reproduced from arXiv: 2505.20482 by the authors.

Figure 1
Figure 1. Distribution of comments with respect to different [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Conversation Kernels dian of the ‘insightful’ category suggests that users are more likely to engage with and reply to insightful comments. Sim￾ilarly, users are more engaged with comments having scores of 3 and 5. Following these findings, we concentrate our analysis of slashdot.org conversations to the four major comment categories which attract the highest distribution of com￾ments and replies: ‘informative’, ‘in… view at source ↗
Figure 2
Figure 2. Box plots showing the number of replies for each [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of different kernel shapes. Different [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: An example ‘funny’ Slashdot conversation. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Prompts for comment nature prediction tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 42 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  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]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Agarwal, P.; Hawkins, O.; Amaxopoulou, M.; Dempsey, N.; Sastry, N.; and Wood, E. 2021. Hate Speech in Political Discourse: A Case Study of UK MPs on Twitter . In Proceedings of the 32nd ACM Conference on Hypertext and Social Media, HT '21, 5–16. New York, NY, USA: Association for Computing Machinery. ISBN 9781450385510

    work page 2021

  5. [5]

    Agarwal, P.; Sastry, N.; and Wood, E. 2019. Tweeting MPs : Digital engagement between citizens and members of parliament in the uk. In Proc.\ the International AAAI Conference on Web and Social Media, volume 13, 26--37

    work page 2019

  6. [6]

    Agarwal, V.; Chen, Y.; and Sastry, N. 2023. Haterephrase: Zero-and few-shot reduction of hate intensity in online posts using large language models. arXiv preprint arXiv:2310.13985

    arXiv 2023

  7. [7]

    Agarwal, V.; Chen, Y.; and Sastry, N. 2024. GASCOM: Graph-based Attentive Semantic Context Modeling for Online Conversation Understanding. Online Social Networks and Media, 43: 100290

    work page 2024

  8. [8]

    Agarwal, V.; Jin, Y.; Chandra, M.; De Choudhury, M.; Kumar, S.; and Sastry, N. 2024 a . MedHalu: Hallucinations in Responses to Healthcare Queries by Large Language Models. arXiv preprint arXiv:2409.19492

    arXiv 2024

Show all 54 references
  1. [9]

    P.; and Sastry, N

    Agarwal, V.; Joglekar, S.; Young, A. P.; and Sastry, N. 2022. GraphNLI : A Graph-based Natural Language Inference Model for Polarity Prediction in Online Debates. In Proc.\ the ACM Web Conference 2022, 2729--2737

    2022

  2. [10]

    Agarwal, V.; Pei, Y.; Alamir, S.; and Liu, X. 2024 b . Codemirage: Hallucinations in code generated by large language models. arXiv preprint arXiv:2408.08333

    2024 arXiv

  3. [11]

    M.; Tyson, G.; and Castro, I

    Agarwal, V.; Raman, A.; Sastry, N.; Abdelmoniem, A. M.; Tyson, G.; and Castro, I. 2024 c . Decentralised Moderation for Interoperable Social Networks: A Conversation-based Approach for Pleroma and the Fediverse. In Proceedings of the International AAAI Conference on Web and So...

    2024

  4. [12]

    P.; Joglekar, S.; and Sastry, N

    Agarwal, V.; Young, A. P.; Joglekar, S.; and Sastry, N. 2023. A graph-based context-aware model to understand online conversations. ACM Transactions on the Web, 18(1): 1--27

    2023

  5. [13]

    Allen, K.; Carenini, G.; and Ng, R. 2014. Detecting Disagreement in Conversations using Pseudo-Monologic Rhetorical Structure. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing ( EMNLP ) , 1169--1180. Doha, Qatar: Association for Computa...

    2014

  6. [14]

    Awadallah, R.; Ramanath, M.; and Weikum, G. 2012. Harmony and Dissonance: Organizing the People's Voices on Political Controversies. In Proceedings of the Fifth ACM International Conference on Web Search and Data Mining, WSDM '12, 523–532. New York, NY, USA: Association for Co...

    2012

  7. [15]

    Basu Roy Chowdhury, S.; and Chaturvedi, S. 2021. Does Commonsense help in detecting Sarcasm? In Proceedings of the Second Workshop on Insights from Negative Results in NLP, 9--15. Online and Punta Cana, Dominican Republic: Association for Computational Linguistics

    2021

  8. [16]

    P.; Joglekar, S.; Cammarota, C.; and Sastry, N

    Boschi, G.; Young, A. P.; Joglekar, S.; Cammarota, C.; and Sastry, N. 2021. Who has the last word? Understanding how to sample online discussions. ACM Transactions on the Web (TWEB), 15(3): 1--25

    2021

  9. [17]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  10. [18]

    Caselli, T.; Basile, V.; Mitrovi \'c , J.; Kartoziya, I.; and Granitzer, M. 2020. I Feel Offended, Don ' t Be Abusive! Implicit/Explicit Messages in Offensive and Abusive Language. In Proceedings of the Twelfth Language Resources and Evaluation Conference, 6193--6202. Marseill...

    2020

  11. [19]

    Choi, Y.; Jung, Y.; and Myaeng, S.-H. 2010. Identifying Controversial Issues and Their Sub-topics in News Articles. In Chen, H.; Chau, M.; Li, S.-h.; Urs, S.; Srinivasa, S.; and Wang, G. A., eds., Intelligence and Security Informatics. Springer Berlin Heidelberg

    2010

  12. [20]

    Dave, K.; Lawrence, S.; and Pennock, D. M. 2003. Mining the Peanut Gallery: Opinion Extraction and Semantic Classification of Product Reviews. In Proceedings of the 12th International Conference on World Wide Web, WWW '03, 519–528. New York, NY, USA: Association for Computing ...

    2003

  13. [21]

    Davidson, T.; Warmsley, D.; Macy, M.; and Weber, I. 2017. Automated Hate Speech Detection and the Problem of Offensive Language. Proceedings of the International AAAI Conference on Web and Social Media, 11(1): 512--515

    2017

  14. [22]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Burstein, J.; Doran, C.; and Solorio, T., eds., NAACL-HLT (1), 4171--4186. Association for Computational Linguistics. ISBN 978-1-950737-13-0

    2019

  15. [23]

    Fan, X.; Lin, H.; Yang, L.; Diao, Y.; Shen, C.; Chu, Y.; and Zou, Y. 2020. Humor detection via an internal and external neural network. Neurocomputing, 394: 105--111

    2020

  16. [24]

    Ghosh, S.; Suri, M.; Chiniya, P.; Tyagi, U.; Kumar, S.; and Manocha, D. 2023. CoSyn: Detecting Implicit Hate Speech in Online Conversations Using a Context Synergized Hyperbolic Network. arXiv:2303.03387

    2023 arXiv

  17. [25]

    G\' o mez, V.; Kaltenbrunner, A.; and L\' o pez, V. 2008. Statistical Analysis of the Social Network and Discussion Threads in Slashdot. In Proceedings of the 17th International Conference on World Wide Web, WWW '08, 645–654. New York, NY, USA: Association for Computing Machin...

    2008

  18. [26]

    Gu, J.-C.; Ling, Z.-H.; Liu, Q.; Liu, C.; and Hu, G. 2023. GIFT: Graph-Induced Fine-Tuning for Multi-Party Conversation Understanding. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 11645–11658. Association f...

    2023

  19. [27]

    Hochreiter, S.; and Schmidhuber, J. 1997. Long Short-Term Memory . Neural Computation, 9(8): 1735--1780

    1997

  20. [28]

    Hoque, E.; Carenini, G.; and Joty, S. 2014. Interactive Exploration of Asynchronous Conversations: Applying a User-centered Approach to Design a Visual Text Analytic System. In Proceedings of the Workshop on Interactive Language Learning, Visualization, and Interfaces, 45--52....

    2014

  21. [29]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  22. [30]

    R.; Liu, S.; Wang, X.; and Xu, G

    Islam, M. R.; Liu, S.; Wang, X.; and Xu, G. 2020. Deep learning for misinformation detection on online social networks: a survey and new perspectives. Social Network Analysis and Mining, 10: 1--20

    2020

  23. [31]

    Karamshuk, D.; Lokot, T.; Pryymak, O.; and Sastry, N. 2016. Identifying partisan slant in news articles and twitter during political crises. In Social Informatics: 8th International Conference, SocInfo 2016, Bellevue, WA, USA, November 11-14, 2016, Proceedings, Part I 8, 257--...

    2016

  24. [32]

    Lin, J. 2022. Leveraging World Knowledge in Implicit Hate Speech Detection. In Proceedings of the Second Workshop on NLP for Positive Impact (NLP4PI), 31--39. Abu Dhabi, United Arab Emirates (Hybrid): Association for Computational Linguistics

    2022

  25. [33]

    Liu, Y.; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy, O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  26. [34]

    C.; and Thompson, S

    Mann, W. C.; and Thompson, S. A. 1988. Rhetorical structure theory: Toward a functional theory of text organization. Text, 8(3): 243--281

    1988

  27. [35]

    T.; and Joty, S

    Mehdad, Y.; Carenini, G.; Ng, R. T.; and Joty, S. 2013. Towards Topic Labeling with Phrase Entailment and Aggregation. In Proceedings of the 2013 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies , 179--189....

    2013

  28. [36]

    Mihaylov, T.; Georgiev, G.; and Nakov, P. 2015. Finding Opinion Manipulation Trolls in News Community Forums. In Proceedings of the Nineteenth Conference on Computational Natural Language Learning, 310--314. Beijing, China: Association for Computational Linguistics

    2015

  29. [37]

    Mozafari, M.; Farahbakhsh, R.; and Crespi, N. 2020. Hate speech detection and racial bias mitigation in social media based on BERT model. PLOS ONE, 15(8): 1--26

    2020

  30. [38]

    Mukherjee, A.; and Liu, B. 2012. Modeling Review Comments. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 320--329. Jeju Island, Korea: Association for Computational Linguistics

    2012

  31. [39]

    U.; Qiu, S.; Saqib, M.; Anwar, S.; Usman, M.; Akhtar, N.; Barnes, N.; and Mian, A

    Naveed, H.; Khan, A. U.; Qiu, S.; Saqib, M.; Anwar, S.; Usman, M.; Akhtar, N.; Barnes, N.; and Mian, A. 2023. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435

    2023 arXiv

  32. [40]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730--27744

    2022

  33. [41]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E.; DeVito, Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B.; Fang, L.; Bai, J.; and Chintala, S. 2019. PyTorch: An ...

    2019

  34. [42]

    A.; Montero-D \' az, J.; and Moreno-Delgado, A

    Paz, M. A.; Montero-D \' az, J.; and Moreno-Delgado, A. 2020. Hate speech: A systematized review. Sage Open, 10(4): 2158244020973022

    2020

  35. [43]

    M.; Luque, F

    P \'e rez, J. M.; Luque, F. M.; Zayat, D.; Kondratzky, M.; Moro, A.; Serrati, P. S.; Zajac, J.; Miguel, P.; Debandi, N.; Gravano, A.; et al. 2023. Assessing the impact of contextual information in hate speech detection. IEEE Access, 11: 30575--30590

    2023

  36. [44]

    Popescu, A.-M.; and Etzioni, O. 2005. Extracting Product Features and Opinions from Reviews. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, 339--346. Vancouver, British Columbia, Canada: Association fo...

    2005

  37. [45]

    Qin, K.; Li, C.; Pavlu, V.; and Aslam, J. 2019. Adapting RNN Sequence Prediction Model to Multi-label Set Prediction. In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (L...

    2019

  38. [46]

    Reyes, A.; Potthast, M.; Rosso, P.; and Stein, B. 2010. Evaluating Humour Features on Web Comments. In Proceedings of the Seventh International Conference on Language Resources and Evaluation ( LREC '10) . Valletta, Malta: European Language Resources Association (ELRA)

    2010

  39. [47]

    Su, Q.; Wan, M.; Liu, X.; Huang, C.-R.; et al. 2020. Motivations, methods and metrics of misinformation detection: an NLP perspective. Natural Language Processing Research, 1(1-2): 1--13

    2020

  40. [48]

    Vanroy, B.; Labat, S.; Kaminska, O.; Lefever, E.; and Hoste, V. 2020. LT 3 at S em E val-2020 Task 7: Comparing Feature-Based and Transformer-Based Approaches to Detect Funny Headlines. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, 1033--1040. Barcelona (on...

    2020

  41. [49]

    Wang, L.; and Ling, W. 2016. Neural Network-Based Abstract Generation for Opinions and Arguments. In Proceedings of the 2016 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies , 47--57. San Diego, California:...

    2016

  42. [50]

    Waseem, Z.; and Hovy, D. 2016. Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on T witter. In Proceedings of the NAACL Student Research Workshop , 88--93. San Diego, California: Association for Computational Linguistics

    2016

  43. [51]

    L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A

    Wolf, T.; Debut, L.; Sanh, V.; Chaumond, J.; Delangue, C.; Moi, A.; Cistac, P.; Rault, T.; Louf, R.; Funtowicz, M.; Davison, J.; Shleifer, S.; von Platen, P.; Ma, C.; Jernite, Y.; Plu, J.; Xu, C.; Scao, T. L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A. M. 2020. Transforme...

    2020

  44. [52]

    Xu, Z.; and Zhu, S. 2010. Filtering Offensive Language in Online Communities using Grammatical Relations. In Proceedings of the Seventh Annual Collaboration, Electronic Messaging, Anti-Abuse and Spam Conference, CEAS. Redmond, Washington, US

    2010

  45. [53]

    Yin, W.; Agarwal, V.; Jiang, A.; Zubiaga, A.; and Sastry, N. 2023. Annobert: Effectively representing multiple annotators’ label choices to improve hate speech detection. In Proceedings of the International AAAI Conference on Web and Social Media, volume 17, 902--913

    2023

  46. [54]

    Zhu, Y.; Zhang, P.; Haq, E.-U.; Hui, P.; and Tyson, G. 2023. Can chatgpt reproduce human-generated labels? a study of social computing tasks. arXiv preprint arXiv:2304.10145

    2023 arXiv

Pith tools

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