Skip to content

Intent detection and classification

Intent detection and classification involve computationally identifying the underlying goals and motivations behind information content. This differs from traditional fake-news detection (binary real/fake) by asking why false information is being spread and what goal it serves.

Problem framing

Intent detection can be framed as:

  • Multiclass classification: Given a text, assign it one or more intent labels from a predefined taxonomy (e.g., political, commercial, ideological).
  • Multilabel classification: Recognize that a single document may serve multiple intents simultaneously.
  • Reasoning task: Generate structured reasoning about the text's intent before making a classification decision.

Challenges

Distinguishing intent from topic: A text about vaccines may be anti-vaccination misinformation (intent: promote anti-scientific views) or legitimate medical debate. Distinguishing requires understanding the author's motivations, not just the topic.

Limited training data: Unlike generic misinformation datasets (millions of examples), intent-annotated datasets are rare and small. MALINT (1,600 articles) and the Polish disinformation dataset (Modzelewski et al. 2024) are among the few.

Subjectivity and disagreement: Annotators must infer intent from textual evidence. Disagreement is high for borderline cases (e.g., a narrative with both legitimately political and destabilizing motivations).

Language and cultural variation: Intent categories that are salient in English (e.g., undermining electoral integrity) may not translate directly to non-Western political contexts.

LLM-specific challenges: Large language models can perform intent detection in zero-shot settings (given only the task definition and examples), but their reasoning may be opaque and prone to confabulation. Fine-tuning on intent-annotated data is expensive and dataset-specific.

Approaches

Supervised classification with SLMs: Fine-tune BERT-family models (DeBERTa, RoBERTa) on annotated datasets. Modzelewski et al. (2026) show that fine-tuned models achieve 82% weighted F₁ on multilabel intent detection, outperforming logistic regression baselines.

Zero-shot LLM-based reasoning: Prompt large language models (GPT-4, Llama, Gemma) to reason about intent without fine-tuning. Effectiveness varies by model size and prompt design; chain-of-thought prompting that first generates intent analysis before classification improves accuracy (~20% improvement in zero-shot disinformation detection in Modzelewski et al. 2026).

Intent-augmented inoculation: Combine intent detection with psychological inoculation principles. Generate structured intent analysis (threat + refutational preemption) to improve downstream disinformation detection.

Multimodal intent detection: Extend intent classification to images, videos, and social-media posts. Some preliminary work exists on identifying propaganda intent in images and multimedia content.

Evaluation

Intent detection is typically evaluated using:

  • F₁ score per intent class (for imbalanced datasets with per-intent counts)
  • Macro/Micro F₁ (averaging across classes)
  • Weighted F₁ (weighting by class frequency)
  • Multilabel evaluation metrics (Hamming loss, subset accuracy for simultaneously correct intents)

Agreement metrics (Cohen's kappa, Krippendorff's alpha) measure annotation reliability.

Applications

Enhanced fact-checking: By understanding the intent behind a claim, fact-checkers can target rebuttals more precisely.

Detection system improvement: Intent features improve disinformation detection in zero-shot LLM settings.

Content moderation: Platforms could use intent signals to prioritize removal of coordinated inauthentic behavior over isolated low-quality content.

Research and analysis: Intent classification enables analysis of narratives, campaigns, and disinformation evolution.

Key papers

  • Bridging Thoughts and Words: Graph-Based Intent-Semantic Joint Learning for (2025) — InSIDE: Graph-based intent-semantic joint learning for fake news detection. Proposes coarse-to-fine intent modeling: high-level conceptual intents (belief, desire, outcome) initialized via frozen LLM and prompts; fine-grained learnable intent nodes capture narrative-specific manifestations. Dynamic alignment with semantic signals via pseudo nodes. Validates adaptability across four intent frameworks. Demonstrates intent is a stable, complementary signal to shifting semantic patterns.
  • Modzelewski et al. (2026) — Intent classification with SLMs and LLMs: Evaluates 12 language models on binary and multilabel intent classification using the MALINT dataset. Shows DeBERTa V3 and RoBERTa outperform other SLMs; fine-tuned models substantially outperform zero-shot LLMs on multilabel tasks. Intent-augmented reasoning improves downstream disinformation detection by ~20%.
  • Prior work on intent-based disinformation detection in Polish language establishes taxonomy and annotation methodology.