Actor-Action-Target Frame Extraction Model

nlpmethodologycore-architecture

Context

We needed a structured representation for narrative content that could be extracted automatically from unstructured Telegram messages.

Decision

Adopt the Actor-Action-Target (AAT) triple as the core analytical unit, extracted using a combination of NER and semantic role labeling.

Alternatives Considered

Topic modeling (LDA/BERTopic)

Pros
  • Well-established methodology
  • Lower computational cost
Cons
  • Loses actor and target information
  • Cannot track specific narrative frames

Full knowledge graph extraction

Pros
  • Richer representation
  • More relationship types
Cons
  • Much higher error rates on noisy Telegram text
  • Slower processing speed

Manual annotation only

Pros
  • Highest accuracy
  • Nuanced understanding
Cons
  • Cannot scale to millions of messages
  • Expensive and slow

Reasoning

AAT frames capture the essential structure of narrative claims (who does what to whom) while being simple enough for reliable automated extraction. The triple format maps directly to how disinformation frames reality and enables quantitative tracking of narrative patterns at scale.

All decisions