arxiv: 2605.00776 · v1 · submitted 2026-05-01 · 💻 cs.CL · cs.AI

Recognition: unknown

Directed Social Regard: Surfacing Targeted Advocacy, Opposition, Aid, Harms, and Victimization in Online Media

Scott Friedman , Ruta Wheelock , Sonja Schmer-Galunder , Drisana Iverson , Jake Vasilakes , Joan Zheng , Jeffrey Rye , Vasanth Sarathy

show 1 more author

Christopher Miller

Authors on Pith no claims yet

Pith reviewed 2026-05-09 18:53 UTC · model grok-4.3

classification 💻 cs.CL cs.AI

keywords directed social regardspan-level sentimentmulti-valence analysismoral framingtarget detectiononline mediatransformer modelssocial science annotation

0 comments

The pith

A pair of transformer models first identifies targets in a message then scores every span on three axes of regard to capture mixed advocacy and opposition.

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

The paper introduces Directed Social Regard to analyze online media for pro-social and anti-social sentiments that point at specific targets within the same message. Existing sentiment tools report only overall positive or negative tone and cannot name the objects of those feelings or handle opposing sentiments side by side. The new approach collects annotated data, trains transformers to locate sentiment-bearing spans, and rates each span along three continuous axes drawn from moral disengagement and moral framing theories. When applied to six existing social-science datasets the outputs align with their labels and topics, suggesting the method can expose patterns of targeted help, harm, blame, and victimization that standard tools miss.

Core claim

Directed Social Regard consists of two transformer-based models that detect span-level targets of sentiment in a message and then assign scores to all spans within context along three (-1, 1) axes of regard motivated by social science theories of moral disengagement and moral framing.

What carries the argument

A pair of transformer-based models: the first detects span-level targets of sentiment and the second scores every span on three continuous regard axes.

If this is right

The validated model can be run on any collection of online posts or articles to surface which entities receive advocacy, opposition, aid, or harm.
Correlations with existing dataset labels indicate that the three-axis scores track recognized social-science constructs such as moral framing.
The span-level output allows a single message to be decomposed into multiple directed sentiments rather than reduced to one overall polarity.
The annotation strategy and architecture provide a reusable template for building larger DSR datasets in new domains.
Application to influence operations and political rhetoric becomes possible because both positive and negative directed sentiments can be reported together.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method could be tested on streaming social-media data to track how victimization or advocacy narratives shift over time around a single event.
Combining DSR outputs with topic models might reveal which issues attract the most mixed or contradictory regard in public discourse.
The three-axis representation might serve as input features for downstream tasks such as detecting coordinated inauthentic behavior that mixes praise and blame.
If the axes generalize beyond the validation sets, they could support comparative studies of moral language across languages or platforms.

Load-bearing premise

The three chosen axes of regard are sufficient to represent the targeted pro-social and anti-social sentiments that appear in online media.

What would settle it

If DSR scores show no reliable correlation with the labels or topics already present in the six third-party online-media datasets, or if human annotators cannot consistently apply the three axes to the same messages.

read the original abstract

The language in online platforms, influence operations, and political rhetoric frequently directs a mix of pro-social sentiment (e.g., advocacy, helpfulness, compassion) and anti-social sentiment (e.g., threats, opposition, blame) at different topics, all in the same message. While many natural language processing (NLP) tools classify or score a text's overall sentiment as positive, neutral, or negative, these tools cannot report that positive and negative sentiments coexist, and they cannot report the target of those sentiments. This paper presents the Directed Social Regard (DSR) approach to multi-dimensional, multi-valence sentiment analysis, comprised of a pair of transformer-based models that (1) detects span-level targets of sentiment in a message and then (2) scores all spans within the message context along three (-1, 1) axes of regard that are motivated by social science theories of moral disengagement and moral framing. We present a data collection and annotation strategy for DSR dataset construction, a transformer-based architecture for span-level scoring, and a validation study with promising results. We apply the validated DSR model on six third-party datasets of online media and report meaningful correlations between DSR outputs and the labels and topics in these pre-existing social science datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DSR pairs span-level target detection with three social-regard axes for mixed-sentiment analysis in online text, but the validation details are too thin to judge reliability.

read the letter

The main point is that this paper gives a concrete way to score targeted pro-social and anti-social sentiments at the span level instead of treating a whole message as one positive or negative blob. It uses three axes drawn from moral disengagement and framing work: advocacy versus opposition, aid versus harms, and victimization. They describe building an annotated dataset, training a pair of transformer models to find targets then score them, and checking the outputs against six existing social-science datasets for correlations. That setup is new enough compared with standard sentiment tools and directly tackles the mixed-valence problem the abstract flags. The external-dataset step is also a reasonable move for showing the outputs line up with real labels and topics. The validation study is called promising, which at least shows they tried to ground the work rather than just release another model. The soft spots sit in the missing numbers. No accuracy figures, no inter-annotator agreement, no error analysis, and no check on whether the three axes leave out sentiments that matter in online media. Without those, it is hard to know if the correlations are robust or if the axes are jointly sufficient. The concern about exhaustiveness is real given what is shown. This is for NLP people and computational social scientists who analyze influence operations, political rhetoric, or online discourse and need finer-grained tools than off-the-shelf sentiment classifiers. Readers who want to adapt the framework or test new axes will find the annotation strategy and architecture useful even if they disagree with the exact three dimensions. I would send it to peer review. The core idea and the external-validation attempt are worth referee time, but the authors will need to supply the quantitative results and some failure cases before it can be assessed properly.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Directed Social Regard (DSR) framework consisting of a pair of transformer-based models that first detect span-level targets of sentiment in online media messages and then score those spans along three (-1, 1) axes (advocacy/opposition, aid/harms, victimization) motivated by moral disengagement and moral framing theories. It describes a data collection and annotation strategy for constructing DSR datasets, presents the model architecture for span-level scoring, reports a validation study with promising results, and applies the model to six third-party datasets of online media, finding meaningful correlations between DSR outputs and the labels/topics in those pre-existing social science datasets.

Significance. If the validation results prove robust upon provision of quantitative metrics and the three axes prove sufficient for capturing targeted sentiments, the work could meaningfully advance multi-valence, target-specific sentiment analysis beyond standard positive/negative classifiers. This would offer a practical bridge between NLP methods and social science theories for studying directed advocacy, harms, and victimization in online discourse, with potential utility for analyzing influence operations and political rhetoric.

major comments (2)

Abstract and validation study section: The central claim that the DSR model produces outputs that meaningfully correlate with labels in six external datasets rests on the validation study, yet the manuscript provides no quantitative metrics (e.g., inter-annotator agreement, model F1 scores, Pearson correlations, or error analysis) to support the 'promising results' assertion; without these, the reliability of the span-scoring outputs and downstream correlations cannot be assessed.
Abstract and theory-motivation section: The approach assumes the three specific axes are jointly sufficient and appropriate for capturing the targeted pro-social and anti-social sentiments present in online media, but the manuscript offers no systematic analysis, coverage study, or evidence that sentiments outside these axes are negligible or that the axes are orthogonal; this assumption is load-bearing for the claim of meaningful correlations with third-party datasets.

minor comments (2)

Abstract: Consider adding at least one concrete performance number or correlation coefficient from the validation study to give readers an immediate sense of the reported 'promising results'.
The description of the transformer architecture for span-level scoring could clarify how the two models in the pair interact (e.g., whether target detection is a prerequisite step or jointly trained).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments on our manuscript. We respond to each major comment below and indicate the revisions we will make to strengthen the paper.

read point-by-point responses

Referee: Abstract and validation study section: The central claim that the DSR model produces outputs that meaningfully correlate with labels in six external datasets rests on the validation study, yet the manuscript provides no quantitative metrics (e.g., inter-annotator agreement, model F1 scores, Pearson correlations, or error analysis) to support the 'promising results' assertion; without these, the reliability of the span-scoring outputs and downstream correlations cannot be assessed.

Authors: We acknowledge the need for quantitative metrics to support the claims in the validation study. Although the manuscript describes the validation and reports meaningful correlations with third-party datasets, it does not include specific numerical values for inter-annotator agreement, model performance scores, or correlation coefficients. We will revise the validation study section to include these metrics (e.g., IAA, F1, Pearson r) along with an error analysis to allow readers to fully assess the reliability of the DSR outputs and the downstream findings. revision: yes
Referee: Abstract and theory-motivation section: The approach assumes the three specific axes are jointly sufficient and appropriate for capturing the targeted pro-social and anti-social sentiments present in online media, but the manuscript offers no systematic analysis, coverage study, or evidence that sentiments outside these axes are negligible or that the axes are orthogonal; this assumption is load-bearing for the claim of meaningful correlations with third-party datasets.

Authors: The three axes are selected based on moral disengagement and moral framing theories, which provide a principled basis for focusing on these dimensions of targeted regard in online media. The manuscript does not include a systematic coverage study or orthogonality analysis. We will add a subsection in the theory-motivation section discussing the theoretical justification for these axes being sufficient for the phenomena under study and report pairwise correlations among the axes from the DSR dataset to evaluate their independence. A comprehensive coverage study of all possible sentiments would require additional data collection beyond the current scope, but we believe the theory-driven approach supports the observed correlations. revision: partial

Circularity Check

0 steps flagged

No circularity: new annotation strategy and span-scoring model are empirically constructed, not derived from fitted inputs

full rationale

The paper presents a data collection/annotation pipeline, a transformer architecture for detecting targets and scoring spans on three (-1,1) regard axes, a validation study, and downstream correlations on six external datasets. No equations, parameter fits, or self-citations are shown that reduce any output to the inputs by construction. The axes are motivated by external social-science theories rather than defined circularly from the model's own predictions. The validation and correlation results are independent empirical checks, not tautological renamings or fitted-input predictions. This is a standard empirical NLP contribution with no load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim depends on the assumption that the three regard axes drawn from moral disengagement and framing theories are valid for the domain and that transformer models can reliably learn span-level scoring from the described annotation process. No specific numerical free parameters are named in the abstract beyond standard model training.

free parameters (1)

transformer model parameters
Weights and hyperparameters fitted during training of the two models; standard for any neural approach but not enumerated.

axioms (1)

domain assumption The three (-1, 1) axes of regard motivated by moral disengagement and moral framing theories adequately represent targeted advocacy, opposition, aid, harms, and victimization in online messages.
Invoked in the abstract as the motivation for the scoring dimensions.

invented entities (1)

Directed Social Regard (DSR) model pair no independent evidence
purpose: To perform span-level target detection followed by multi-axis regard scoring.
Newly proposed architecture in this work.

pith-pipeline@v0.9.0 · 5558 in / 1607 out tokens · 41558 ms · 2026-05-09T18:53:09.931920+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 5 canonical work pages · 2 internal anchors

[1]

Nature human behaviour7(6), 917–927 (2023)

Brady, W.J., McLoughlin, K.L., Torres, M.P., Luo, K.F., Gendron, M., Crockett, M.: Overperception of moral outrage in online social networks inflates beliefs about intergroup hostility. Nature human behaviour7(6), 917–927 (2023)

2023
[2]

arXiv preprint arXiv:2203.03608 (2022)

Guo, S., Burghardt, K., Rao, A., Lerman, K.: Emotion regulation and dynam- ics of moral concerns during the early covid-19 pandemic. arXiv preprint arXiv:2203.03608 (2022)

work page arXiv 2022
[3]

In: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, vol

Schmer-Galunder, S., Wheelock, R., Jalan, Z., Chvasta, A., Friedman, S., Saltz, E.: Annotator in the loop: A case study of in-depth rater engagement to create a prosocial benchmark dataset. In: Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, vol. 7, pp. 1319–1328 (2024)

2024
[4]

In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp

Cai, H., Ma, H., Yu, J., Xia, R.: A joint coreference-aware approach to document- level target sentiment analysis. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 12149– 12160 (2024)

2024
[5]

arXiv preprint arXiv:2509.11604 (2025)

Hossain, M.M., Hossain, M.S., Chaki, S., et al.: Dynamic span interaction and graph-aware memory for entity-level sentiment classification. arXiv preprint arXiv:2509.11604 (2025)

work page arXiv 2025
[6]

In: Proceedings of the 14th Workshop on Computa- tional Approaches to Subjectivity, Sentiment, & Social Media Analysis, pp

Rønningstad, E., Klinger, R., Øvrelid, L., Velldal, E.: Entity-level sentiment: More than the sum of its parts. In: Proceedings of the 14th Workshop on Computa- tional Approaches to Subjectivity, Sentiment, & Social Media Analysis, pp. 84–96 (2024)

2024
[7]

Journal of moral education31(2), 101–119 (2002)

Bandura, A.: Selective moral disengagement in the exercise of moral agency. Journal of moral education31(2), 101–119 (2002)

2002
[8]

Journal of personality and social psychology71(2), 364 (1996)

Bandura, A., Barbaranelli, C., Caprara, G.V., Pastorelli, C.: Mechanisms of moral 28 disengagement in the exercise of moral agency. Journal of personality and social psychology71(2), 364 (1996)

1996
[9]

In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

Li, X., Bing, L., Lam, W., Shi, B.: A unified model for opinion target extraction and target sentiment prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, pp. 6714–6721 (2019)

2019
[10]

In: Proceedings of EMNLP, pp

Xu, H., Liu, B., Shu, L., Yu, P.S.: Position-aware tagging for aspect sentiment triplet extraction. In: Proceedings of EMNLP, pp. 2339–2349 (2020)

2020
[11]

IEEE Computational Intelligence Magazine15(1), 64–75 (2020)

Akhtar, M.S., Ekbal, A., Cambria, E.: How intense are you? predicting intensities of emotions and sentiments using stacked ensemble [application notes]. IEEE Computational Intelligence Magazine15(1), 64–75 (2020)

2020
[12]

In: Proceedings of SemEval, pp

Pontiki, M., Galanis, D., Papageorgiou, H., Androutsopoulos, I., Manandhar, S., AL-Smadi, M., Al-Ayyoub, M., Zhao, Y., Qin, B., De Clercq, O.,et al.: Semeval- 2016 task 5: Aspect based sentiment analysis. In: Proceedings of SemEval, pp. 19–30 (2016)

2016
[13]

In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

Yang, J., Yang, R., Wang, C., Xie, J.: Multi-entity aspect-based sentiment anal- ysis with context, entity and aspect memory. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32 (2018)

2018
[14]

In: Proceed- ings of the Sixth Workshop on Online Abuse and Harms (WOAH), pp

Zheng, J., Friedman, S., Schmer-Galunder, S., Magnusson, I., Wheelock, R., Got- tlieb, J., Gomez, D., Miller, C.: Towards a multi-entity aspect-based sentiment analysis for characterizing directed social regard in online messaging. In: Proceed- ings of the Sixth Workshop on Online Abuse and Harms (WOAH), pp. 203–208 (2022). Association for Computational L...

2022
[15]

In: Proceedings of SemEval, pp

Mohammad, S.M., Kiritchenko, S., Sobhani, P., Zhu, X., Cherry, C.: Semeval- 2016 task 6: Detecting stance in tweets. In: Proceedings of SemEval, pp. 31–41 (2016)

2016
[16]

In: Proceedings of ACL, pp

Xu, C., He, W., Lu, W., Xu, Q.: Multi-target stance detection via a multi-task learning framework incorporating stance lexicons. In: Proceedings of ACL, pp. 3829–3838 (2019)

2019
[17]

In: Proceedings of NAACL, pp

Hazarika, D., Zimmermann, R., Mihalcea, R., Zimmermann, R., Zimmermann, R.: Conversational stance prediction using a sequential matching network. In: Proceedings of NAACL, pp. 1432–1442 (2021)

2021
[18]

In: The International FLAIRS Conference Proceedings, vol

Mather, B., Dorr, B.J., Rambow, O., Strzalkowski, T.: A general framework for domain-specialization of stance detection: a covid-19 response use case. In: The International FLAIRS Conference Proceedings, vol. 34 (2021)

2021
[19]

arXiv preprint arXiv:2203.10659 (2022)

Mather, B., Dorr, B.J., Dalton, A., Beaumont, W., Rambow, O., Schmer- Galunder, S.M.: From stance to concern: Adaptation of propositional analysis to 29 new tasks and domains. arXiv preprint arXiv:2203.10659 (2022)

work page arXiv 2022
[20]

In: Proceedings of the 55th Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers), pp

He, L., Lee, K., Lewis, M., Zettlemoyer, L.: Deep semantic role labeling: What works and what’s next. In: Proceedings of the 55th Annual Meeting of the Asso- ciation for Computational Linguistics (Volume 1: Long Papers), pp. 473–483 (2017)

2017
[21]

Brown, P., Levinson, S.C.: Politeness: Some Universals in Language Usage vol. 4. Cambridge university press, ??? (1987)

1987
[22]

Current directions in psychological science9(3), 75–78 (2000)

Bandura, A.: Exercise of human agency through collective efficacy. Current directions in psychological science9(3), 75–78 (2000)

2000
[23]

Handbook of theories of social psychology2, 313–343 (2012)

Jost, J.T., Toorn, J.: System justification theory. Handbook of theories of social psychology2, 313–343 (2012)

2012
[24]

Simmel, G.: The sociology of conflict. i. American journal of sociology9(4), 490– 525 (1904)

1904
[25]

Vintage, ??? (2012)

Haidt, J.: The Righteous Mind: Why Good People Are Divided by Politics and Religion. Vintage, ??? (2012)

2012
[26]

Guilford Press, New York (1995)

Weiner, B.: Judgments of Responsibility: A Foundation for a Theory of Social Conduct. Guilford Press, New York (1995)

1995
[27]

Oxford University Press, New York (2011)

Batson, C.D.: Altruism in Humans. Oxford University Press, New York (2011)

2011
[28]

Harvard University Press, Cambridge, MA (1982)

Gilligan, C.: In a Different Voice: Psychological Theory and Women’s Develop- ment. Harvard University Press, Cambridge, MA (1982)

1982
[29]

In: Advances in Experimental Social Psychology vol

Graham, J., Haidt, J., Koleva, S., Motyl, M., Iyer, R., Wojcik, S.P., Ditto, P.H.: Moral foundations theory: The pragmatic validity of moral pluralism. In: Advances in Experimental Social Psychology vol. 47, pp. 55–130. Elsevier, ??? (2013)

2013
[30]

In: Proceedings of the International AAAI Conference on Web and Social Media, vol

Baumgartner, J., Zannettou, S., Keegan, B., Squire, M., Blackburn, J.: The pushshift reddit dataset. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 14, pp. 830–839 (2020)

2020
[31]

In: Proceedings of the Annual Meeting of the Cognitive Science Society, vol

Friedman, S.E., Magnusson, I., Schmer-Galunder, S., Wheelock, R., Gottlieb, J., Miller, C.,et al.: Toward transformer-based nlp for extracting psychosocial indicators of moral disengagement. In: Proceedings of the Annual Meeting of the Cognitive Science Society, vol. 43 (2021)

2021
[32]

Social Psychological and Personality Science11(8), 1057–1071 (2020) 30

Hoover, J., Portillo-Wightman, G., Yeh, L., Havaldar, S., Davani, A.M., Lin, Y., Kennedy, B., Atari, M., Kamel, Z., Mendlen, M.,et al.: Moral foundations twitter corpus: A collection of 35k tweets annotated for moral sentiment. Social Psychological and Personality Science11(8), 1057–1071 (2020) 30

2020
[33]

PsyArXiv

Kennedy, B., Atari, M., Davani, A.M., Yeh, L., Omrani, A., Kim, Y., Coombs, K., Havaldar, S., Portillo-Wightman, G., Gonzalez, E., et al.: The gab hate corpus: A collection of 27k posts annotated for hate speech. PsyArXiv. July18(2018)

2018
[34]

Computers, Materials & Continua 66(2) (2021)

Ul Rehman, Z., Abbas, S., Khan, M.A., Mustafa, G., Fayyaz, H., Hanif, M., Saeed, M.A.: Understanding the language of isis: An empirical approach to detect radical content on twitter using machine learning. Computers, Materials & Continua 66(2) (2021)

2021
[35]

In: Proceedings of the International AAAI Conference on Web and Social Media, vol

Ribeiro, M.H., Blackburn, J., Bradlyn, B., De Cristofaro, E., Stringhini, G., Long, S., Greenberg, S., Zannettou, S.: The evolution of the manosphere across the web. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 15, pp. 196–207 (2021)

2021
[36]

In: Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pp

Pradhan, S., Elhadad, N., Chapman, W., Manandhar, S., Savova, G.: Semeval- 2014 task 7: Analysis of clinical text. In: Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pp. 54–62 (2014)

2014
[37]

DeBERTa: Decoding-enhanced BERT with Disentangled Attention

He, P., Liu, X., Gao, J., Chen, W.: Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654 (2020)

work page internal anchor Pith review arXiv 2006
[38]

He, P., Gao, J., Chen, W.: DeBERTaV3: Improving DeBERTa using ELECTRA- Style Pre-Training with Gradient-Disentangled Embedding Sharing (2021)

2021
[39]

In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pp

Friedman, S., Zheng, J., Steinmetz, H.: Debiasing multi-entity aspect-based sen- timent analysis with norm-based data augmentation. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pp. 4456–4461 (2024)

2024
[40]

LoRA: Low-Rank Adaptation of Large Language Models

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

work page internal anchor Pith review Pith/arXiv arXiv 2021
[41]

In: Proceedings of the International AAAI Conference on Web and Social Media, vol

Davidson, T., Warmsley, D., Macy, M., Weber, I.: Automated hate speech detec- tion and the problem of offensive language. In: Proceedings of the International AAAI Conference on Web and Social Media, vol. 11, pp. 512–515 (2017) 6 Appendix* 31 T able 8Referring phrases used across datasets to categorize Character spans into groups. Category (Display) Lemma...

2017