Unlike traditional text classification, which assigns documents to explicit, surface-level topics (such as "sports" or "finance"), identifying manipulative or persuasive narratives requires models to capture implicit ideological framing, agenda-driven reporting, and subtle rhetorical structures.
Compounding this difficulty, real-world narratives are inherently complex. They are rarely isolated; instead, they form nested, multi-level taxonomies where a broad parent narrative (e.g., Discrediting Climate Science) branches into specific sub-narratives (e.g., Claiming data is manipulated), which are further defined by the specific roles assigned to entities (e.g., Protagonists, Antagonists, and Victims). Furthermore, these narratives propagate globally across multiple languages simultaneously.
To tackle this multi-label, multi-class, and multilingual challenge, recent research—heavily catalyzed by benchmarks like SemEval-2025 Task 10—has converged on a set of distinct architectural paradigms designed to handle narrative complexity.
Stagewise Reasoning: Deconstructing the Hierarchy
When a classification model is forced to choose from dozens of fine-grained sub-narratives simultaneously, it often suffers from performance degradation due to context fragmentation and label confusion. To combat this, a prominent direction in the literature leverages hierarchical, stagewise reasoning to mirror human cognitive processing.
Instead of predicting the entire taxonomic path in a single pass, models break the problem down into sequential, isolated decisions.
[Raw News Text]
│
▼
┌───────────────────────────┐
│ Stage 1: Domain Detection │ ──► (e.g., Russia-Ukraine War)
└───────────────────────────┘
│
▼
┌───────────────────────────┐
│ Stage 2: Main Narrative │ ──► (e.g., Economic Sanctions Impact)
└───────────────────────────┘
│
▼
┌───────────────────────────┐
│ Stage 3: Sub-Narrative │ ──► (e.g., Sanctions hurt the West more)
└───────────────────────────┘
In GateNLP at SemEval-2025 Task 10: Hierarchical Three-Step Prompting for Multilingual Narrative Classification, Iknoor Singh et al. introduce "H3Prompt," a three-step prompting strategy for Large Language Models (LLMs). By isolating the domain, the main narrative, and the sub-narrative into distinct steps, the model's attention is constrained to a highly relevant subset of choices at each layer, drastically reducing the search space.
Similarly, Qiangyu Tan et al. propose the Tree-guided Stagewise Classifier (TGSC) in YNUzwt at SemEval-2025 Task 10: Tree-guided Stagewise Classifier for Entity Framing and Narrative Classification. This framework relies on scaffolding theory, guiding the LLM through a structured tree path to ensure that fine-grained sub-narrative assignments remain logically consistent with their parent classifications.
By structuring the task as a flow of thoughts, as explored by Du Py et al. in DUTtask10 at SemEval-2025 Task 10: ThoughtFlow: Hierarchical Narrative Classification via Stepwise Prompting, models generate explicit intermediate reasoning chains. This step-by-step decomposition ensures that subtle, implicit cues embedded within lengthy news articles are not lost in a single, monolithic prediction step.
The Paradigm Hybrid: Prompting Meets Fine-Tuning
A recurring tension in this research space lies between the flexibility of generative prompting and the reliability of discriminative fine-tuning.
- Large Language Models (LLMs) excel at zero-shot reasoning and capturing abstract, implicit framing, but they can struggle with strict multi-label constraints, hallucinated labels, and high computational costs.
- Smaller Language Models (SLMs) like BERT or BART provide stable, high-accuracy multi-label classification, but they require substantial amounts of high-quality labeled data to grasp nuanced narrative structures.
To resolve this, researchers frequently employ hybrid pipelines that use LLMs as data generators or reasoners, and SLMs as the final execution engine. For instance, the authors of ThoughtFlow use an LLM to generate structured "thought processes," which are then distilled into the training data of a smaller, fine-tuned model to boost its performance on low-resource languages like Hindi.
In bbStar at SemEval-2025 Task 10: Improving Narrative Classification and Explanation via Fine Tuned Language Models, Rishit Tyagi et al. implement a two-stage architecture: a fine-tuned BERT model handles the heavy lifting of multi-label narrative classification, while a retrieval-augmented LLM operating under a ReACT framework generates grounded, transparent explanations based on a taxonomy knowledge base.
Conversely, when evaluating pure approaches, Vineet Saravanan et al. demonstrate in cocoa at SemEval-2025 Task 10: Prompting vs. Fine-Tuning: A Multilevel Approach to Propaganda Classification that while LLMs offer zero-shot flexibility, carefully fine-tuned transformer models consistently outperform them in raw accuracy and reliability when mapping complex propaganda techniques and entity roles.
Overcoming the Multilingual Bottleneck
Because manipulative narratives are frequently deployed across borders, models must generalize across diverse languages. The literature reveals two competing strategies for handling this multilingual requirement: native multilingual embeddings versus machine translation pipelines.
Surprisingly, the simpler "Translate-then-Classify" pipeline frequently emerges as the more robust solution.
[Multilingual Input Text] ──► [Machine Translation] ──► [Monolingual English SLM] ──► [Narrative Labels]
In Irapuarani at SemEval-2025 Task 10: Evaluating Strategies Combining Small and Large Language Models for Multilingual Narrative Detection, Gabriel Assis et al. systematically compared native multilingual models against a translation-based approach. They discovered that translating all source texts into English and subsequently applying a monolingual SLM yielded the most consistent and generalizable performance across languages.
This phenomenon occurs because high-resource monolingual models capture subtle stylistic framing and rhetorical nuances far more effectively than multilingual models, whose embedding spaces are stretched thin across dozens of syntax structures. This strategy is echoed by Jesus M. Fraile-Hernandez et al. in UNEDTeam at SemEval-2025 Task 10: Zero-Shot Narrative Classification, where translating text to English allowed zero-shot LLMs to better navigate context management and accurately assign labels from complex taxonomies.
Alternative Representations: Hyperbolic Geometry and Retrieval
Beyond standard prompting and fine-tuning, some researchers explore alternative mathematical representations to natively capture the geometry of narrative taxonomies.
Hyperbolic Embeddings
Standard Euclidean space expands polynomially, making it poorly suited for representing hierarchical trees, which expand exponentially as you move from root to leaves. Forcing a tree structure into Euclidean space introduces severe distortion.
To address this, Christopher Brückner et al. explore hyperbolic geometry in Hierarchical Classification of Propaganda Techniques in Slavic Texts in Hyperbolic Space. Because the volume of hyperbolic space grows exponentially with its radius, it naturally accommodates hierarchical structures, allowing the model to preserve the nested relationships between broad propaganda categories and fine-grained techniques without losing structural information.
Retrieval-Based Classification
Rather than treating narrative detection as a traditional parametric classification task, Arjumand Younus et al. reframe it as a retrieval task in nlptuducd at SemEval-2025 Task 10: Narrative Classification as a Retrieval Task through Story Embeddings.
By mapping news articles into a shared "story embedding" space, the system bypasses explicit label boundaries entirely. Instead, it leverages semantic similarity to transfer narrative labels from known, annotated articles to new, unseen texts. This non-parametric approach proves highly effective at capturing the fluid, overlapping, and evolving nature of online disinformation campaigns.
Key Foundational Work
GateNLP at SemEval-2025 Task 10: Hierarchical Three-Step Prompting for Multilingual Narrative Classification (2025) — contributed 2 concept elements to this direction.
DUTtask10 at SemEval-2025 Task 10: ThoughtFlow: Hierarchical Narrative Classification via Stepwise Prompting (2025) — contributed 2 concept elements to this direction.
bbStar at SemEval-2025 Task 10: Improving Narrative Classification and Explanation via Fine Tuned Language Models (2025) — contributed 1 concept elements to this direction.