Directed Social Regard: Surfacing Targeted Advocacy, Opposition, Aid, Harms, and Victimization in Online Media
| Authors | Scott Friedman et al. |
| Year | 2026 |
| Field | NLP |
| arXiv | 2605.00776 |
| Download | |
| Categories | cs.CL, cs.AI |
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.
Engineering Breakdown
Plain English
This paper addresses a critical gap in sentiment analysis: most NLP tools score text as simply positive, negative, or neutral, but they miss the real-world complexity where a single message contains mixed sentiments directed at different targets. The authors present Directed Social Regard (DSR), a two-stage transformer-based system that first identifies span-level targets of sentiment within text, then scores the sentiment directed at each target independently. This enables nuanced analysis of messages common in online discourse, political rhetoric, and influence operations—where someone might express support for one group while criticizing another, all in the same sentence. The approach moves beyond global document-level sentiment classification to target-aware, multi-valence sentiment detection.
Core Technical Contribution
The core novelty is decomposing multi-valence sentiment analysis into two tractable transformer tasks: (1) target span extraction—identifying what entities or concepts the sentiment is directed toward, and (2) sentiment scoring per span—assigning valence to each target independently rather than the entire message. This two-stage architecture elegantly sidesteps the complexity of joint multi-label sentiment prediction by treating it as a sequence labeling problem followed by classification, allowing reuse of well-established transformer patterns. Prior work either collapsed multiple sentiments into a single global score or required complex multi-task learning setups; DSR's pipeline approach is simpler to implement, train, and debug while maintaining interpretability—the output explicitly shows what was said about whom. The method is specifically designed for real-world messy text where sentiment targets are overlapping, nested, or numerous within a single message.
How It Works
The DSR system operates in two sequential stages on transformer model inputs. Stage 1 takes a raw text input and uses a transformer encoder (likely BERT-family or similar) with token-level classification heads to identify sentiment target spans—this is a BIO or similar tagging scheme that marks which token sequences are targets of sentiment. Stage 2 then extracts all identified spans from stage 1 and, for each span, feeds a context window (the span plus surrounding context) into a second transformer model that predicts sentiment polarity—typically a multi-class classification (positive/negative/neutral/mixed) or regression score. Both stages leverage transformer attention mechanisms to capture long-range dependencies; the first model learns what linguistic patterns indicate sentiment targets (e.g., proper nouns, entity mentions, topic references), while the second learns how sentiment words and phrasing relate to each target. The pipeline design allows stage 2 to be run in parallel over all detected spans, and outputs are then aggregated into a structured representation (target → sentiment value mapping) that preserves which sentiments attach to which targets.
Production Impact
For teams building content moderation, influence analysis, or political discourse systems, DSR directly solves a major classification problem: today's systems flag messages as 'toxic' or 'positive' holistically, missing that a message praising one group while attacking another requires different handling. In production, you'd integrate DSR as a preprocessing/enrichment step in your classification pipeline—feeding flagged messages into DSR to get target-aware sentiment breakdowns, then routing different targets to different moderation queues or feeding richer signals to downstream models. This is especially valuable for narrative analysis, disinformation detection, and polarization measurement where capturing the full sentiment landscape within a message is crucial. Trade-offs include: (1) inference latency—two sequential transformer passes per message is ~2x a single-pass baseline, (2) data requirements—you need training data with span-level target and per-target sentiment annotations, which is more expensive to collect than document-level labels, (3) compute cost—two model instances or larger batch sizes to amortize inference, (4) engineering complexity—your pipeline now has failure modes in stage 1 (missed targets) that cascade into stage 2, requiring careful monitoring.
Limitations and When Not to Use This
The paper's abstract doesn't detail the training data or evaluation metrics, but the approach inherently struggles with (1) ambiguous or implicit targets—if sentiment is directed at an idea or concept without explicit mention (e.g., 'this is terrible' with no stated target), stage 1 may fail to extract anything, (2) nested or overlapping targets—complex sentences where one target is a subset of another, or where sentiment spans multiple discontinuous targets, strain the span extraction model and require careful handling in the tagging scheme, (3) domain shift—a model trained on political discourse may not generalize to product reviews or clinical notes where sentiment targets have different linguistic patterns. The two-stage pipeline is greedy and non-iterative; errors in stage 1 (false negatives in span detection) cannot be corrected by stage 2, so performance is bounded by recall in target identification. Additionally, this approach assumes sentiments are compositional and can be scored independently per target, which breaks down in cases of sarcasm, context-dependent flipping, or interdependent sentiments where criticizing one target implicitly endorses another.
Research Context
This work builds on decades of sentiment analysis research (lexicon-based, machine learning, then deep learning approaches) but specifically addresses the multi-target, mixed-sentiment setting that prior academic work often simplified away. It likely relates to aspect-based sentiment analysis (ABSA), which similarly decomposes sentiment by target, but extends that to political and social contexts where targets are less formal and more semantically fluid. The paper also engages with research on online discourse analysis, computational propaganda detection, and influence operations—areas where naive sentiment scoring has proven inadequate for researchers and platforms. DSR opens research directions in interpretable NLP (the explicit target-sentiment mapping is inherently more interpretable than black-box scores), multi-hop reasoning over sentiment (chaining multiple targets), and cross-lingual sentiment analysis where target extraction patterns may differ sharply across languages.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::