💬 NLP

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

This paper introduces Directed Social Regard (DSR), a novel multi-dimensional sentiment analysis framework utilizing transformer-based models to simultaneously identify sentiment targets and score them along three social-regard axes, thereby capturing the coexistence of pro- and anti-social sentiments in online media that traditional tools miss.

Original authors: Scott Friedman, Ruta Wheelock, Sonja Schmer-Galunder, Drisana Iverson, Jake Vasilakes, Joan Zheng, Jeffrey Rye, Vasanth Sarathy, Christopher Miller

Published 2026-05-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Scott Friedman, Ruta Wheelock, Sonja Schmer-Galunder, Drisana Iverson, Jake Vasilakes, Joan Zheng, Jeffrey Rye, Vasanth Sarathy, Christopher Miller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a crowded, noisy town square. In this square, people are shouting messages at each other. Sometimes, a single person shouts a sentence that is a mix of love, hate, anger, and pity all at once.

For example, someone might yell: "How can they do this to the children??"

Older tools for analyzing language (like basic sentiment analyzers) are like a simple thermometer. They look at that whole sentence and say, "This is Negative." That's it. They miss the nuance. They don't tell you who is being hated, who is being pitied, or what action is being condemned.

This paper introduces a new, much smarter tool called Directed Social Regard (DSR). Think of DSR not as a thermometer, but as a high-tech spotlight with three different colored lenses.

The Three Lenses of the Spotlight

Instead of just saying "Good" or "Bad," the DSR spotlight shines on specific words (or "spans") in a sentence and asks three specific questions for each one:

The "Oppose vs. Advocate" Lens: Is the speaker trying to stop this person/thing, or are they cheering for them?
- In our example: The speaker is Opposing "them" and Advocating for "the children."
The "Harmful vs. Helpful" Lens: Does the speaker think this person/thing is causing damage or providing help?
- In our example: "Them" and the action "do this" are seen as Harmful.
The "Victimized vs. Aided" Lens: Is the speaker seeing this person/thing as a victim who needs saving, or as someone who is being helped?
- In our example: The "children" are seen as Victimized.

How the Tool Works (The "Brain" and the "Hands")

The researchers built a two-part machine learning system to do this:

The "Hands" (Span Recognition): First, the system reads the text and acts like a highlighter. It finds the important parts. It highlights "they" as a character, "children" as a character, and "do this" as a topic. It knows the difference between a person and an event.
The "Brain" (Regard Scoring): Once the parts are highlighted, the "Brain" assigns a score to each one on a scale from -1 (strongly negative) to +1 (strongly positive) for each of the three lenses mentioned above.

Why This Matters (The "Why")

The paper explains that this isn't just about fancy math; it's based on real human psychology.

Moral Disengagement: Sometimes, people justify hurting others by saying, "They deserved it." The DSR tool can catch this complex mix: it sees that the speaker views a group as both "Opposed" (we hate them) and "Harmful" (they are bad), which is a way humans justify bad behavior.
Group Dynamics: It helps us see how groups talk about "Us" vs. "Them." For instance, in one dataset they studied (a group called the "Manosphere"), the tool found that the speakers often viewed themselves as victims ("Me/Us") while viewing women as the "Harmful" ones. In another dataset (about the #MeToo movement), the tool flipped that script, showing women as "Victimized" and men as "Harmful."

The Proof (Did it work?)

The researchers tested their tool in two ways:

The Test Drive: They fed it thousands of social media posts they had manually labeled by humans. The tool's "highlighting" was about 90-96% accurate, and its "scoring" matched human opinions very closely.
The Reality Check: They took the tool and ran it on existing datasets from other researchers (like datasets about hate speech, political movements, and online debates).
- Result: The tool's scores lined up perfectly with what human experts had already found. For example, when a dataset was labeled "Moral Outrage," the DSR tool correctly showed high "Opposition" scores toward the targets of that outrage.

The Bottom Line

This paper presents a new way to listen to online conversations. Instead of just hearing "This is a bad tweet," the DSR tool lets us hear: "This tweet is attacking Group A for hurting Group B, while expressing deep sympathy for Group B."

It's a tool that helps us understand the complex, layered, and often contradictory feelings people express when they type on their phones. The researchers have built the engine, trained it on a massive dataset of 1,800+ posts, and proven it works better than older, simpler methods at spotting who is being blamed, who is being saved, and who is being cheered on.

1. Problem Statement

Current Natural Language Processing (NLP) tools for sentiment analysis typically classify an entire text as positive, negative, or neutral, or assign a single intensity score. These approaches fail to address two critical complexities in online media:

Co-existing Sentiments: A single message often contains mixed sentiments (e.g., advocating for a victim while opposing an aggressor) that are lost when aggregated into a single score.
Target Ambiguity: Standard tools cannot identify who or what the sentiment is directed toward (the target), nor can they distinguish between different types of social relationships (e.g., is the target being helped, harmed, opposed, or victimized?).

Existing frameworks like Targeted Sentiment Analysis (TSA), Aspect-Based Sentiment Analysis (ABSA), and Stance Detection often require pre-defined targets, rely on categorical labels, or lack multidimensional evaluation. There is a need for a framework that can automatically extract targets and score them along continuous, theoretically grounded dimensions of social regard.

2. Methodology: The Directed Social Regard (DSR) Framework

The authors propose Directed Social Regard (DSR), a multi-dimensional, multi-valence sentiment analysis approach grounded in social science theories (specifically Moral Disengagement and Moral Framing).

A. Theoretical Dimensions

DSR defines sentiment not as a single polarity but as three independent, bipolar axes (ranging from -1 to +1):

Opposed–Advocated: Whether the author desires to hinder or defend a target.
Harmful–Helpful: Whether the author deems the target a source of harm or help to others.
Victimized–Aided: Whether the author views the target as being harmed or helped by others.

These dimensions allow for nuanced representations, such as a target being simultaneously "Opposed" and "Harmful" (justifying aggression) or "Advocated" and "Victimized" (expressing empathy).

B. Dataset Construction

Data Sources: The authors collected 1,838 texts from eight diverse Publicly Accessible Information (PAI) sources, including Reddit, New York Times comments, Gab, and specific corpora on hate speech, the "Manosphere," and political events (e.g., #MeToo, French elections).
Annotation: Human annotators identified spans (contiguous word sequences representing characters or topics) and scored them on the three DSR axes using a slider interface (-1.0 to 1.0).
Scale: The final dataset contains 13,893 spans with 26,779 averaged intensity scores. Inter-annotator agreement (Krippendorff's $\alpha$ ) was high (0.71–0.91).

C. Model Architecture

The DSR system employs a two-stage transformer-based pipeline:

Span Recognition (Token Classification):
- Uses a fine-tuned Transformer (e.g., DeBERTa-v3) to identify and classify spans as either Characters (people/groups) or Topics (events/concepts).
- Output: A list of detected spans with their types.
Regard Scoring (Span-Level Intensity):
- Takes the identified spans and the full text context.
- Uses a Transformer encoder (DeBERTa-v3) to generate contextual embeddings.
- Applies a Context Pooler to normalize span vectors.
- Passes vectors through linear layers to predict three continuous regression scores (-1 to 1) for each span: Opposed–Advocated, Harmful–Helpful, and Victimized–Aided.
- Note: The model is trained using Mean Squared Error (MSE) loss.

3. Key Contributions

Novel Framework: Introduction of DSR, which moves beyond categorical sentiment to continuous, multidimensional, and target-specific social regard.
New Dataset: Creation of a large-scale, high-quality annotated dataset for training and evaluating span-level, multi-dimensional social sentiment.
Theoretical Integration: Successfully operationalizing social science theories (Bandura's Moral Disengagement, Brown & Levinson's Politeness) into NLP tasks.
Zero-Shot Target Discovery: Unlike Stance Detection, DSR does not require pre-specified targets; it automatically discovers salient spans in the text.

4. Experimental Results

A. Span Recognition Performance

The best-performing model (DeBERTa-v3-large-mnli) achieved an F1 score of 0.96 for character spans and 0.74 for topic spans under strict matching criteria.
Token-level metrics were higher (F1 ~0.92), indicating the model effectively identifies the boundaries of sentiment targets.

B. Regard Scoring Performance

The scoring model demonstrated strong correlation with human annotations.
Opposed–Advocated: $R^2 = 0.85$ , RMSE = 0.20.
Victimized–Aided: $R^2 = 0.67$ , RMSE = 0.10.
Harmful–Helpful: $R^2 = 0.57$ , RMSE = 0.12.
Comparison: The specialized DeBERTa models significantly outperformed large language models (LLMs) like Llama 3.2 (fine-tuned via LoRA) and GPT-4o (few-shot), suggesting that purpose-built regression models are more effective for this specific numerical prediction task than generative LLMs.

C. Application Studies (Validation on External Datasets)

The authors applied the trained DSR model to six external datasets to validate its utility:

Bridging Benchmark: DSR scores correlated strongly with document-level labels. High "Moral Outrage" correlated with high opposition toward 2nd/3rd person targets, while "Respect" correlated with advocacy.
Moral Foundations Twitter Corpus (MFTC): DSR attributes aligned with Moral Foundations theory. "Virtues" (e.g., Care) correlated with "Helpful/Aided" scores, while "Vices" (e.g., Harm) correlated with "Harmful/Opposed" scores.
Corpus Comparison:
- Manosphere vs. Hate Speech: The Manosphere corpus showed high "Advocate for You" and "Victim|Me/Us" (self-victimization), whereas the Hate Speech corpus showed high "Advocate Harm."
- MeToo: Showed high "Victim|Female" and "Harmful|Male," aligning with movement themes.
- ALM vs. Baltimore: Distinguished opposing views on law enforcement; "All Lives Matter" showed advocacy for police, while "Baltimore" protests showed opposition.
Pairwise Analysis: The model successfully extracted micro-narratives (e.g., "Them harms Us") by visualizing co-occurring harm/help pairs, revealing structural patterns in community discourse.

5. Significance and Implications

Granularity: DSR provides a granular view of online discourse, revealing how pro-social and anti-social sentiments coexist within single messages directed at different targets.
Social Science Utility: It offers a quantitative tool for researchers to study moral disengagement, group identity formation, and the mechanics of online radicalization without relying on pre-defined keywords.
Platform Safety: The framework can help platforms identify complex harmful content (e.g., advocating for harm against a specific group) that might be missed by simple negative sentiment filters.
Future Directions: The authors suggest using DSR for real-time monitoring of influence operations, understanding cultural shifts in moral framing, and improving the explainability of AI-driven content moderation.

In conclusion, the paper presents a robust, theoretically grounded NLP framework that significantly advances the state of the art in sentiment analysis by capturing the direction, intensity, and multidimensionality of social regard in online media.