Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Self-Anchored Attention Model for Sample-Efficient Classification of Prosocial Text Chat

T0 review · 4 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 The paper claims a self-anchored attention model detects prosocial game chat with an AUC of 0.836, beating the previous best by 7.9%.

desk verdict Useful new dataset and honest baseline comparison for a genuinely underexplored task, but the headline generalization claim is not supported by the keyword-filtered, unclear-merging evaluation, and the attention mechanism is less novel than claimed. read the letter →

arxiv 2506.09259 v1 pith:CAQIHGET submitted 2025-06-10 cs.CL cs.AIcs.CY

Zhuofang Li , Rafal Kocielnik , Fereshteh Soltani , Penphob (Andrea) Boonyarungsrit , Animashree Anandkumar , R. Michael Alvarez This is my paper
classification cs.CLcs.AIcs.CY
keywords prosocialbehaviorgamechatlow-resourcetextclassificationself-anchoredattentioncontrastivelearningCallofDutytopicmodelingsample-efficient
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 is trying to establish that prosocial communication in competitive game chat—messages that build community, invite teammates, or share resources—can be automatically detected even when fewer than 200 labeled examples are available. It proposes a Self-Anchored Attention Model (SAAM) that reuses the entire training anchor set as attention keys and values, first to fine-tune sentence embeddings with contrastive learning and then to classify new messages. On a hand-labeled set of 960 Call of Duty chat histories, the best SAAM variant reaches a mean AUC of 0.836 with a standard deviation of 0.012, a 7.9% improvement over the strongest existing technique. If the result holds, moderation systems could shift from only punishing toxicity to actively rewarding and encouraging positive interactions.

What carries the argument

The central mechanism is Self-Anchored Attention (SAAM), in which the entire labeled anchor set serves as the attention memory: each input embedding becomes a query, while all anchor embeddings serve as both keys and values, and the resulting context vector is passed to a classification head. A simplified K-attention variant that learns only the key weighting matrix reduces overfitting while keeping similar accuracy. This self-anchoring is what lets every new message be compared against every labeled example, and the paper attributes the sample-efficiency gain to that exhaustive reuse.

What would settle it

Take a fresh sample of public game chat that was not pre-filtered by community-builder keywords, label it with the original three categories kept separate, and run the trained SAAM model on it; if the AUC on that unfiltered mix falls to the level of generic sentiment or toxicity baselines, the reported advantage would be an artifact of the filter rather than of prosocial detection.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a small labeled anchor set, instead of being merely training data, can be converted into the model's reference memory. SAAM uses the whole anchor set twice: once for contrastive fine-tuning of embeddings to game-domain language, and once as the key/value store in a lightweight attention layer. The best configuration, SAAM with K attention, reports a mean AUC of 0.836 over three seeded 20%/80% splits, outperforming fine-tuned transformers, toxicity detectors, few-shot prompting, and representation-based baselines, with the remaining errors concentrated on messages that human annotators themselves had labeled as 'unclear'.

Load-bearing premise

The evaluation rests on the assumption that the keyword-filtered, consensus-labeled 960 instances fairly represent 'prosocial behavior' in real, unfiltered game chat, and that collapsing 'unclear' into 'not-prosocial' does not distort the task.

Editorial extensions

If this is right

  • A prosocial detection system can flag community-building messages with roughly 80% precision at 60–80% recall, which is practical for reward and recognition loops.
  • The anchor-set size matters directly: reducing the anchor set from 20% to 10% of the data drops AUC by 16.9%, so the benefit depends on keeping a reasonably sized reference pool.
  • The attention layer is the main source of the gain: removing it cuts AUC by 9.1%, while removing the embedding fine-tuning step costs 2.4%.
  • The approach is positioned to work in other low-resource settings where a small curated set defines a behavior worth encouraging, such as supportive or collaborative language on other platforms.

Reading between the lines

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

  • If the K-attention variant generalizes, the same self-anchoring trick could be applied to other low-resource classification targets where a small curated set defines a style or behavior, such as detecting constructive comments or collaborative speech.
  • Because 57–71% of the model's false positives were judged 'unclear' by human raters, a three-way output (prosocial / unclear / not-prosocial) may be more faithful to the underlying label structure than the binary collapse the paper uses.
  • The keyword pre-screen means the recall on prosocial messages that do not contain the selected vocabulary remains untested; that boundary could be probed directly with an unfiltered held-out set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper addresses the under-explored task of detecting prosocial communication in online game chat. Using Call of Duty: Modern Warfare II chat data, the authors combine unsupervised topic modeling with domain-expert input to define five prosocial categories, then construct a 960-instance labeled dataset for the "community builder" category. They propose a Self-Anchored Attention Model (SAAM) that reuses the small labeled anchor set twice: first to fine-tune sentence embeddings with contrastive learning, and then as keys/values in an attention layer for classification. On three seeded 20%/80% splits, the best variant (SAAM K attention) is reported to achieve a mean AUC of 0.836±0.012, a 7.9% relative improvement over the strongest baseline (Deberta-v3+adapter at 0.775±0.014). Ablations show that the attention component contributes substantially and that performance degrades as the anchor set shrinks. The paper also provides confusion matrices, precision-recall curves, and a manual error analysis.

Significance. If the reported result is valid for the intended deployment setting, this is a meaningful contribution: it introduces a novel task, a new low-resource classification architecture that reuses the entire training set, and a carefully constructed expert-in-the-loop labeling pipeline with substantial inter-annotator agreement (Cohen's kappa 0.755). The paper deserves credit for comparing against a wide range of baselines, for isolating the effect of the attention mechanism through ablations, for reporting precision-recall trade-offs, and for being transparent about the ambiguity of the "unclear" label. However, the central claim that SAAM detects prosocial behavior in real in-game chat is only as strong as the evaluation distribution, and the current evaluation is restricted to a keyword-filtered, one-week sample with a debatable label-merging choice. The significance would be high if the method were shown to generalize beyond that filtered distribution.

major comments (4)
  1. [§3.3, Table 3] The evaluation distribution is constructed by design to contain only messages that match the Community Builder keyword list. Section 3.3 filters the preprocessed chat to messages containing the keywords in Table 3, and Section 3.4 labels only those candidates. Consequently, the 0.836 AUC in Table 6 is a measure of discrimination among keyword-containing, non-toxic, non-spam messages of at least 11 words; it is not evidence about detecting prosocial messages that lack these keywords, which is exactly the setting the abstract's claim ('classifying prosocial behaviors in in-game chats') implies. This is load-bearing for the headline result. The authors should either evaluate on an unfiltered random sample of chat (reporting how many prosocial instances fall outside the keyword filter), or explicitly rescope the claims to 'prosocial community-builder chat that matches the keyword filter.'
  2. [§3.5.1 and §4.2] Merging the 'unclear' label into 'not-prosocial' is not a neutral preprocessing choice. The paper's own error analysis reports that 57–71% of false positives were originally labeled 'unclear' by human annotators. Since 'unclear' cases are treated as negatives, the reported AUC partly measures how well the model agrees with a labeling convention that classifies genuinely ambiguous messages as negative. The authors should report sensitivity analyses: for example, exclude 'unclear' instances from test evaluation, or treat them as a third class, or report precision/recall separately for the subset originally labeled 'unclear.' Without such analysis, the practical meaning of the 0.836 AUC for deployment is unclear.
  3. [Table 6 and Table 7] The headline comparison rests on only three seeded splits with no significance testing. The difference between SAAM (K attention) at 0.836±0.012 and Deberta-v3+adapter at 0.775±0.014 is large relative to the reported standard deviations, but with three folds the reader cannot assess whether the ordering is stable across splits. In addition, Table 7 shows that multiple SAAM variants (Cos-Sim, K attention, QKV attention) were compared, and the best variant was selected after seeing the results; this selection is not accounted for in the reported gains. The authors should provide per-fold paired comparisons (e.g., the three per-fold AUC differences) and either use a fixed model-selection rule or discuss the selection bias.
  4. [§3.4, Table 4] The labeled dataset consists of 960 instances from a single week (2023-12-11 to 2023-12-17) of one game title, after several exclusion steps. While the paper frames this as a low-resource benchmark, the generalizability of the absolute AUC number to other time periods, game modes, or games is not established. At minimum, the limitations section should state explicitly that the evaluation is a convenience sample and that the reported AUC is a within-filter estimate, not a population-level performance measure for all in-game chat.
minor comments (5)
  1. [Table 4] In the Character Length rows, the Min and Median values are identical (49, 53, 49, 67), which appears to be a data-entry or formatting error and should be corrected.
  2. [Table 5] The table heading contains a typo: 'Exmaples' should be 'Examples.'
  3. [§4.2] The sentence beginning 'Given that only 11%-16% of the labeled test set was categorized as false positives ... and only 18%-20% of the test set data is labeled as unclear' is confusing because the false-positive rate and the 'unclear' rate are percentages with different denominators and are not directly comparable. The argument that the misclassified instances are genuinely ambiguous would be clearer if the authors computed, among the false positives, the proportion that were originally 'unclear' (which they report as 57–71%) and then discussed the base rates separately.
  4. [§3.5.4, Step 1] The contrastive sampling description is ambiguous: it says 10 anchors are sampled 'for each pro-social keyword and for each label,' but it is not clear how the anchor set is partitioned by keyword, whether sampling is with replacement, or how the pairwise positive and negative pairs are constructed if a keyword has fewer than 10 anchors in a label. Please clarify the exact sampling procedure and any exclusion rules.
  5. [§3.5.3] For the ConvLSTM model, 'learning epochs' should read 'number of training epochs,' and the sentence 'we experimentally determined the LR of 1e-3 and learning epochs of 30' should be rephrased for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SAAM reuses the anchor set as a training mechanism, and the reported AUC gain is measured on held-out test splits against external baselines.

full rationale

The derivation chain is self-contained. SAAM's 'self-anchoring' uses the 20% anchor set as attention keys and values and for contrastive SBERT fine-tuning; the three-seeded-split protocol evaluates on a disjoint 80% test set, so the reported AUC (0.836 ± 0.012 for K attention) is a held-out performance estimate and not a quantity that reduces to the training labels by construction. The 7.9% improvement is computed relative to external baselines (SetFit, Deberta-v3+adapter, ToxicBert+adapter, ConvLSTM, KNN, frozen and prompted models), so the central claim is not imported from a self-citation. Self-citations such as [41] and [42] support data provenance and few-shot selection details but are not load-bearing for the claimed result. The acknowledged choice to collapse 'unclear' into 'not-prosocial' (§3.5.1) and the keyword-filtered construction of the labeled set (§3.3) are construct-validity limitations about the evaluation distribution, not circularity: the model is trained on labels and tested on held-out instances from that same filtered distribution. The only mild concern is that the best SAAM variant (K attention) was selected after comparing SAAM variants on the evaluation folds; this is a model-selection and overfitting concern rather than a definitional or fitted-parameter circularity, and it does not rise to the level of a circular step under the stated criteria.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on several domain and modeling assumptions that are not proven by the paper: keyword pre-filtering identifies candidate prosocial chat; unclear labels can be merged into not-prosocial; the chosen prosocial category (Community Builder) is representative; SBERT embeddings plus contrastive fine-tuning transfer to game chat; and three 80/20 splits are a reliable evaluation. No new physical or mathematical entities are introduced. Free parameters are mostly data-splitting and sampling choices that materially affect the reported gain.

free parameters (4)
  • Anchor set split ratio = 20% (192 of 960 instances)
    Defines the low-resource setup; ablation shows AUC drops 16.9% at 10% and 4.8% at 15% (Table 7), so this choice directly shapes the headline gain.
  • Contrastive positive samples per keyword per label = 10
    In Step 1 (Section 3.5.4), 10 prosocial and 10 non-prosocial anchors are sampled per keyword; pairwise positive and negative pairs are generated from this count.
  • SAAM variant selection = K attention (also tested QKV and cosine-similarity variants)
    The authors tried several self-anchoring variants and report the best one (K attention, AUC 0.836) in Table 6; this is model selection on the same evaluation splits.
  • Baseline tuning choices = KNN k=24; few-shot k=4
    Selected by grid search on the evaluation data (Section 3.5.3), so the comparison includes baselines tuned to the test distribution.
assumptions (5)
  • domain assumption Keyword-based pre-filtering identifies prosocial candidate chat instances
    Sections 3.2 and 3.3: spam, toxic, and short chats are removed, and only messages containing expert-selected prosocial keywords are labeled; the evaluation therefore measures classification on this filtered distribution, not on all game chat.
  • ad hoc to paper The 'unclear' label can be merged into 'not-prosocial' for binary classification
    Section 3.5.1: the binary task collapses unclear with not-prosocial; later analysis shows 57-71% of false positives were originally unclear, so this coding choice affects the ground truth the model optimizes.
  • domain assumption Sentence-BERT embeddings plus contrastive fine-tuning transfer to game chat language
    Section 3.5.4 assumes general-domain sentence embeddings are a useful feature space after contrastive adaptation; ablation shows a 2.4% AUC drop when fine-tuning is removed.
  • domain assumption Three seeded randomized splits with 80% test provide a reliable performance estimate
    Section 3.5.1 chooses three splits without significance testing; standard deviations across splits are reported but no statistical test separates the 7.9% gain from noise.
  • domain assumption Community Builder category is a valid target for prosocial detection
    Section 3.3 selects this category based on expert interest and topic-cluster strength; results are not shown for the other four proposed prosocial categories.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Anchored Attention Model for Sample-Efficient Classification of Prosocial Text Chat." pith.science (2026). https://pith.science/paper/CAQIHGET

@misc{pith2026250609259,
  author       = {Pith},
  title        = {Pith review of: Self-Anchored Attention Model for Sample-Efficient Classification of Prosocial Text Chat},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CAQIHGET}},
  note         = {Machine review of arXiv:2506.09259}
}
read the original abstract

Millions of players engage daily in competitive online games, communicating through in-game chat. Prior research has focused on detecting relatively small volumes of toxic content using various Natural Language Processing (NLP) techniques for the purpose of moderation. However, recent studies emphasize the importance of detecting prosocial communication, which can be as crucial as identifying toxic interactions. Recognizing prosocial behavior allows for its analysis, rewarding, and promotion. Unlike toxicity, there are limited datasets, models, and resources for identifying prosocial behaviors in game-chat text. In this work, we employed unsupervised discovery combined with game domain expert collaboration to identify and categorize prosocial player behaviors from game chat. We further propose a novel Self-Anchored Attention Model (SAAM) which gives 7.9% improvement compared to the best existing technique. The approach utilizes the entire training set as "anchors" to help improve model performance under the scarcity of training data. This approach led to the development of the first automated system for classifying prosocial behaviors in in-game chats, particularly given the low-resource settings where large-scale labeled data is not available. Our methodology was applied to one of the most popular online gaming titles - Call of Duty(R): Modern Warfare(R)II, showcasing its effectiveness. This research is novel in applying NLP techniques to discover and classify prosocial behaviors in player in-game chat communication. It can help shift the focus of moderation from solely penalizing toxicity to actively encouraging positive interactions on online platforms.

Figures

Figures reproduced from arXiv: 2506.09259 by the authors.

Figure 1
Figure 1. Text Chat Data Processing Pipeline. The pipeline preprocesses data, uses topic modeling to identify prosocial candidates, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Self-anchored Attention Model. Note that for our self-anchoring models, the anchor set is [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Confusion Matrices for the test set across the 3 data [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Precision-recall curves for the test set across the 3 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Topic Modeling Results and Cluster-keywords [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Sign in 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. Prosocial Behavior Detection in Player Game Chat: From Aligning Human-AI Definitions to Efficient Annotation at Scale

    cs.CL 2025-08 conditional novelty 6.0 of 10

    A hybrid system using a calibrated SVM plus selective GPT-4o fallback detects prosocial game chat at roughly 0.90 precision while cutting LLM inference cost by about 70%.

Reference graph

Works this paper leans on

75 extracted references · 47 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Robert Bamler, Farnood Salehi, and Stephan Mandt. 2020. Augmenting and tuning knowledge graph embeddings. InUncertainty in Artificial Intelligence. PMLR, 508–518

    work page 2020

  3. [3]

    Rachel Baumsteiger and Jason Siegel. 2018. Measuring Prosociality: The Develop- ment of a Prosocial Behavioral Intentions Scale.Journal of Personality Assessment 101 (02 2018). https://doi.org/10.1080/00223891.2017.1411918

    arXiv 2018

  4. [4]

    Nicole A Beres, Julian Frommel, Elizabeth Reid, Regan L Mandryk, and Madison Klarkowski. 2021. Don’t you know that you’re toxic: Normalization of toxicity in online gaming. InProceedings of the 2021 CHI conference on human factors in computing systems. 1–15

    work page 2021

  5. [5]

    Palwasha Bibi. 2024. Legal and Ethical Challenges of Content Moderation: Bal- ancing Privacy and Free Speech in the AI Era. (2024)

    work page 2024

  6. [6]

    Blei, Andrew Y

    David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent Dirichlet Allocation.J. Mach. Learn. Res.3, null (mar 2003), 993–1022

    work page 2003

  7. [7]

    Sebastian Bodenstedt, Dominik Rivoir, Alexander Jenke, Martin Wagner, Michael Breucha, Beat Müller-Stich, Sören Torge Mees, Jürgen Weitz, and Stefanie Speidel

    work page

  8. [8]

    Nan-Chen Chen, Margaret Drouhard, Rafal Kocielnik, Jina Suh, and Cecilia R Aragon. 2018. Using machine learning to support qualitative coding in social sci- ence: Shifting the focus to ambiguity.ACM Transactions on Interactive Intelligent Systems (TiiS)8, 2 (2018), 1–20

    work page 2018

Show all 75 references
  1. [9]

    Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. 2020. Recall and Learn: Fine-tuning Deep Pretrained Language Models with Less Forgetting. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 7870–7881

    2020

  2. [10]

    Jean Decety, Inbal Bartal, Florina Uzefovsky, and Ariel Knafo-Noam. 2016. Em- pathy as a driver of prosocial behaviour: Highly conserved neurobehavioural mechanisms across species.Philosophical Transactions of the Royal Society B: Biological Sciences371 (01 2016), 20150077. ht...

    2016

  3. [11]

    Nicole M Deterding and Mary C Waters. 2021. Flexible coding of in-depth interviews: A twenty-first-century approach.Sociological methods & research50, 2 (2021), 708–739

    2021

  4. [12]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805(2018)

    2018 arXiv

  5. [13]

    Tobias Eder, Viktor Hangya, and Alexander Fraser. 2020. Anchor-based bilingual word embeddings for low-resource languages.arXiv preprint arXiv:2010.12627 (2020)

    2020 arXiv

  6. [14]

    Fabes, and Tracy Spinrad

    Nancy Eisenberg, R.A. Fabes, and Tracy Spinrad. 2006. Prosocial development. In N. Eisenberg, W. Damon, & R. M. Lerner (Eds.),: Vol 3, Social, emotional, and personality development.Handbook of child psychology(01 2006), 646–718

    2006

  7. [15]

    Shihab Elbagir and Jing Yang. 2019. Twitter sentiment analysis using natural language toolkit and VADER sentiment.Proceedings of the international multi- conference of engineers and computer scientists122, 16 (2019)

    2019

  8. [16]

    Hongchao Fang and Pengtao Xie. 2023. CERT: Contrastive Self-supervised Learning for Language Understanding.Authorea Preprints(2023)

    2023

  9. [17]

    David Farr, Iain Cruickshank, Nico Manzonelli, Nicholas Clark, Kate Starbird, and Jevin West. 2024. LLM Confidence Evaluation Measures in Zero-Shot CSS Classification.arXiv preprint arXiv:2410.13047(2024)

    2024 arXiv

  10. [18]

    Findlay, Alberta Girardi, and Robert J

    Leanne C. Findlay, Alberta Girardi, and Robert J. Coplan. 2006. Links between empathy, social behavior, and social understanding in early childhood.Early Childhood Research Quarterly21, 3 (2006), 347–359. https://doi.org/10.1016/j. ecresq.2006.07.009

    2006 doi

  11. [19]

    Neda Firoz, Olga Grigorievna Beresteneva, Aksyonov Sergey Vladimirovich, Mohammad Sadman Tahsin, and Faiza Tafannum. 2023. Automated text-based depression detection using hybrid ConvLSTM and Bi-LSTM model. In2023 Third International Conference on Artificial Intelligence and Sm...

    2023

  12. [20]

    Julian Frommel and Regan L Mandryk. 2022. Daily quests or daily pests? The benefits and Pitfalls of engagement rewards in games.Proceedings of the ACM on Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato and Tobin, et al. Human-Computer Interaction6, CHI PLAY (20...

    2022

  13. [21]

    Furianto Furianto and Risa Rumentha Simanjuntak. 2023. Gaming Language as a Language Variations in Digital Humanities. InE3S Web of Conferences, Vol. 388. EDP Sciences, 04010

    2023

  14. [22]

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. 2024. Bias and fairness in large language models: A survey.Computational Linguistics(2024), 1–79

    2024

  15. [23]

    T Gao, X Yao, and Danqi Chen. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. InEMNLP 2021-2021 Conference on Empirical Methods in Natural Language Processing, Proceedings

    2021

  16. [24]

    Douglas A Gentile, Craig A Anderson, Shintaro Yukawa, Nobuko Ihori, Muniba Saleem, Lim Kam Ming, Akiko Shibuya, Albert K Liau, Angeline Khoo, Brad J Bushman, et al. 2009. The effects of prosocial video games on prosocial behaviors: International evidence from correlational, lo...

    2009

  17. [25]

    Gentile, Craig A

    Douglas A. Gentile, Craig A. Anderson, Shintaro Yukawa, Nobuko Ihori, Mu- niba Saleem, Lim Kam Ming, Akiko Shibuya, Albert K. Liau, Angeline Khoo, Brad J. Bushman, L. Rowell Huesmann, and Akira Sakamoto. 2009. The Effects of Prosocial Video Games on Prosocial Behaviors: Intern...

    2009 doi

  18. [26]

    Maarten Grootendorst. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv:2203.05794 [cs.CL]

    2022 arXiv

  19. [27]

    Hamilton

    W.D. Hamilton. 1964. The genetical evolution of social behaviour. I.Journal of Theoretical Biology7, 1 (1964), 1–16. https://doi.org/10.1016/0022-5193(64)90038- 4

    1964 doi

  20. [28]

    Pengrui Han, Rafal Kocielnik, Adhithya Saravanan, Roy Jiang, Or Sharir, and Anima Anandkumar. 2024. ChatGPT Based Data Augmentation for Improved Parameter-Efficient Debiasing of LLMs.arXiv preprint arXiv:2402.11764(2024)

    2024 arXiv

  21. [29]

    Laura Hanu and Unitary team. 2020. Detoxify. Github. https://github.com/unitaryai/detoxify

    2020

  22. [30]

    Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. DeBERTaV3: Improv- ing DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing.CoRRabs/2111.09543 (2021). arXiv:2111.09543 https: //arxiv.org/abs/2111.09543

    2021 arXiv

  23. [31]

    Zorah Hilvert-Bruce and James T. Neill. 2020. I’m just trolling: The role of normative beliefs in aggressive behaviour in online gaming.Computers in Human Behavior102 (2020), 303–311. https://doi.org/10.1016/j.chb.2019.09.003

    2020 doi

  24. [32]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-Efficient Transfer Learning for NLP. arXiv:1902.00751 [cs.LG]

    2019 arXiv

  25. [33]

    Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Lee. 2023. LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models. InProceedings of the 2023 Conference on Empirical Methods in Natural...

    2023

  26. [34]

    Yongxin Huang, Kexin Wang, Sourav Dutta, Raj Nath Patel, Goran Glavaš, and Iryna Gurevych. 2023. AdaSent: Efficient Domain-Adapted Sentence Embeddings for Few-Shot Classification. arXiv:2311.00408 [cs.CL]

    2023 arXiv

  27. [35]

    HuggingFace. 2024. Trainer. https://huggingface.co/docs/transformers/en/main_ classes/trainer. (Accessed on 11/05/2024)

    2024

  28. [36]

    Hutto and Eric Gilbert

    C. Hutto and Eric Gilbert. 2014. VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text.Proceedings of the International AAAI Conference on Web and Social Media8, 1 (May 2014), 216–225. https: //doi.org/10.1609/icwsm.v8i1.14550

    2014 doi

  29. [37]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al . 2023. Mistral 7B.arXiv preprint arXiv:2310.06825(2023)

    2023 arXiv

  30. [38]

    Low-resource

    Zhiying Jiang, Matthew Yang, Mikhail Tsirlin, Raphael Tang, Yiqin Dai, and Jimmy Lin. 2023. “Low-resource” text classification: A parameter-free classifica- tion method with compressors. InFindings of the Association for Computational Linguistics: ACL 2023. 6810–6828

    2023

  31. [39]

    Xisen Jin, Francesco Barbieri, Brendan Kennedy, Aida Mostafazadeh Davani, Leonardo Neves, and Xiang Ren. 2021. On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning. InProceedings of the 2021 Conference of the North American Chapter of the Association for...

    2021

  32. [40]

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning.Advances in neural information processing systems33 (2020), 18661– 18673

    2020

  33. [41]

    Rafal Kocielnik, Sara Kangaslahti, Shrimai Prabhumoye, Meena Hari, R.Michael Alvarez, and Anima Anandkumar. 2023. Can you label less by using out-of- domain data? Active & transfer learning with few-shot instructions. InTransfer Learning for Natural Language Processing Worksho...

    2023

  34. [42]

    Rafal Kocielnik, Zhuofang Li, Claudia Kann, Deshawn Sambrano, Jacob Morrier, Mitchell Linegar, Carly Taylor, Min Kim, Nabiha Naqvie, and Feri Soltani. 2024. Challenges in moderating disruptive player behavior in online competitive action games.Frontiers in Computer Science6 (2...

    2024

  35. [43]

    Rafal Kocielnik, Shrimai Prabhumoye, Vivian Zhang, Roy Jiang, R Michael Al- varez, and Anima Anandkumar. 2023. Biastestgpt: Using chatgpt for social bias testing of language models.arXiv preprint arXiv:2302.07371(2023)

    2023 arXiv

  36. [44]

    Steven Kuhn. 2008. Prisoner’s Dilemma. InStanford Encyclopedia of Philosophy, Ed Zalta (Ed.). Policy Press

    2008

  37. [45]

    Egoitz Laparra, Aurelie Mascio, Sumithra Velupillai, and Timothy Miller. 2021. A review of recent work in transfer learning and domain adaptation for natural language processing of electronic health records.Yearbook of medical informatics 30, 01 (2021), 239–244

    2021

  38. [46]

    Thanh-Dung Le, Ti Ti Nguyen, and Vu Nguyen Ha. 2024. The Impact of LoRA Adapters for LLMs on Clinical NLP Classification Under Data Limitations.arXiv preprint arXiv:2407.19299(2024)

    2024 arXiv

  39. [47]

    Alyssa Lees, Vinh Q Tran, Yi Tay, Jeffrey Sorensen, Jai Gupta, Donald Metzler, and Lucy Vasserman. 2022. A new generation of perspective api: Efficient multilingual character-level transformers. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mi...

    2022

  40. [48]

    Chak Leong, Yi Cheng, Jiashuo Wang, Jian Wang, and Wenjie Li. 2023. Self- Detoxifying Language Models via Toxification Reversal. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Associat...

    2023 doi

  41. [49]

    Chak Tou Leong, Yi Cheng, Jiashuo Wang, Jian Wang, and Wenjie Li. 2023. Self-detoxifying language models via toxification reversal.arXiv preprint arXiv:2310.09573(2023)

    2023 arXiv

  42. [50]

    Willa Litvack-Miller, Daniel McDougall, and David M. Romney. 1997. The struc- ture of empathy during middle childhood and its relationship to prosocial be- havior.Genetic, social, and general psychology monographs123 3 (1997), 303–24. https://api.semanticscholar.org/CorpusID:29982230

    1997

  43. [51]

    Yanling Liu, Zhaojun Teng, Haiying Lan, Xin Zhang, and Dezhong Yao. 2015. Short-term effects of prosocial video games on aggression: an event-related potential study.Frontiers in Behavioral Neuroscience9 (2015), 193

    2015

  44. [52]

    Rahul Makhijani, Parikshit Shah, Vashist Avadhanula, Caner Gocmen, Nicolás E Stier-Moses, and Julián Mestre. 2021. Quest: Queue simulation for content moderation at scale.arXiv preprint arXiv:2103.16816(2021)

    2021 arXiv

  45. [53]

    1984.Selfishness, altruism, and rationality

    Howard Margolis. 1984.Selfishness, altruism, and rationality. University of Chicago Press

    1984

  46. [54]

    Mary L McHugh. 2012. Interrater reliability: the kappa statistic.Biochemia medica22, 3 (2012), 276–282

    2012

  47. [55]

    Aniket Nagle, Peter Wolf, Robert Riener, and Domen Novak. 2014. The use of player-centered positive reinforcement to schedule in-game rewards inreases enjoyment and performance in a serious game.International Journal of Serious Games1, 4 (2014), 35–47. https://doi.org/10.17083...

    2014 doi

  48. [56]

    Oluwafemi Oriola and Eduan Kotzé. 2020. Improved semi-supervised learning technique for automatic detection of South African abusive language on Twitter. South African Computer Journal32, 2 (2020), 56–79

    2020

  49. [57]

    Louis A Penner, John F Dovidio, Jane A Piliavin, and David A Schroeder. 2005. Prosocial behavior: Multilevel perspectives.Annu. Rev. Psychol.56 (2005), 365– 392

    2005

  50. [58]

    Susanne Poeller, Martin Johannes Dechant, Madison Klarkowski, and Regan L Mandryk. 2023. Suspecting sarcasm: how league of legends players dismiss positive communication in toxic environments.Proceedings of the ACM on Human-Computer Interaction7, CHI PLAY (2023), 1–26

    2023

  51. [59]

    Shrimai Prabhumoye, Rafal Kocielnik, Mohammad Shoeybi, Anima Anandkumar, and Bryan Catanzaro. 2021. Few-shot instruction prompts for pretrained language models to detect social biases.arXiv preprint arXiv:2112.07868(2021)

    2021 arXiv

  52. [60]

    Charan Reddy. 2022. Benchmarking bias mitigation algorithms in representation learning through fairness metrics.Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track(2022)

    2022

  53. [61]

    YCAP Reddy, P Viswanath, and B Eswara Reddy. 2018. Semi-supervised learning: A brief review.Int. J. Eng. Technol7, 1.8 (2018), 81

    2018

  54. [63]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084(2019)

    2019 arXiv

  55. [64]

    Fabian David Schmidt, Ivan Vulić, and Goran Glavaš. 2023. One For All & All For One: Bypassing Hyperparameter Tuning with Model Averaging for Cross- Lingual Transfer. InFindings of the Association for Computational Linguistics: EMNLP 2023. 12186–12193

    2023

  56. [65]

    Jens Seiffert-Brockmann, Trevor Diehl, and Leonhard Dobusch. 2018. Memes as games: The evolution of a digital discourse online.New Media & Society20, 8 (2018), 2862–2879

    2018

  57. [66]

    Amnon Shashua and Tamir Hazan. 2005. Non-Negative Tensor Factorization with Applications to Statistics and Computer Vision. InProceedings of the Self-Anchored Attention Model for Sample-Efficient Classification of Prosocial Text Chat Conference acronym ’XX, June 03–05, 2018, W...

    2005

  58. [67]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805(2023)

    2023 arXiv

  59. [68]

    Sian Tomkinson and Benn Van Den Ende. 2022. ‘thank you for your compliance’: Overwatch as a disciplinary system.Games and culture17, 2 (2022), 198–218

    2022

  60. [69]

    Robert L. Trivers. 1971. The Evolution of Reciprocal Altruism.Quarterly Review of Biology46, 1 (1971), 35–57

    1971

  61. [70]

    Lewis Tunstall, Nils Reimers, Unso Eun Seo Jo, Luke Bates, Daniel Korat, Moshe Wasserblat, and Oren Pereg. 2022. Efficient Few-Shot Learning Without Prompts. arXiv:2209.11055 [cs.CL]

    2022 arXiv

  62. [71]

    Henry Weld, Guanghao Huang, Jean Lee, Tongshu Zhang, Kunze Wang, Xinghong Guo, Siqu Long, Josiah Poon, and Caren Han. 2021. CONDA: a CONtextual Dual- Annotated dataset for in-game toxicity understanding and detection. InFindings of the Association for Computational Linguistics...

    2021

  63. [72]

    Michel Wijkstra, Katja Rogers, Regan L Mandryk, Remco C Veltkamp, and Ju- lian Frommel. 2023. Help, My Game Is Toxic! First Insights from a Systematic Literature Review on Intervention Systems for Toxic Behaviors in Online Video Games. InCompanion Proceedings of the Annual Sym...

    2023

  64. [73]

    Ziang Xiao, Xingdi Yuan, Q Vera Liao, Rania Abdelghani, and Pierre-Yves Oudeyer

  65. [74]

    GG”, “Respect

    Zachary Yang, Nicolas Grenon-Godbout, and Reihaneh Rabbany. 2023. Towards Detecting Contextual Real-Time Toxicity for In-Game Chat. InFindings of the Association for Computational Linguistics: EMNLP 2023. 9894–9906. A PROSOCIAL EXPERT KNOWLEDGE We identify the following prosoc...

    2023

  66. [2019]

    Active learning using deep Bayesian networks for surgical workflow analysis.International journal of computer assisted radiology and surgery14 (2019), 1079–1087

    2019

  67. [2023]

    InCompanion proceedings of the 28th international conference on intelligent user interfaces

    Supporting qualitative analysis with large language models: Combining codebook with GPT-3 for deductive coding. InCompanion proceedings of the 28th international conference on intelligent user interfaces. 75–78

Pith tools

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