Narrative Analytics
Uncovering timelines, events, and hidden meanings in Persian and Kurdish narratives using LLMs.
Overview
This project investigates how large language models (LLMs) handle narrative understanding in low-resource languages, specifically Persian and Sorani Kurdish. Focusing on timeline construction, implicit event detection, and cultural nuance, we explore how well LLMs can replicate human-like comprehension in these languages—despite limited training data.
As geopolitical and social narratives increasingly shape online discourse, our goal is to evaluate whether modern AI systems can meaningfully parse narrative structure, bias, and implicit information in underrepresented language contexts.
Methodology
We apply both zero-shot and fine-tuned LLMs to a curated multilingual corpus of news articles, social media threads, and cultural texts in Persian and Kurdish. Key methods include:
- Prompt-based evaluation of event detection, coreference, and implicit relation resolution
- Temporal expression identification and timeline assembly
- Entity-event-timeline linking
- Gold-standard test set creation for benchmarking narrative comprehension
- Cross-model comparison (GPT, LLaMA, Claude, etc.)
Evaluation blends human annotation, linguistic heuristics, and automated metrics for coherence, completeness, and cultural appropriateness.
Preliminary Results
- LLMs show strong performance in surface-level event identification, but struggle with nested timelines and implicit causality.
- Persian fares better than Sorani due to more available training data.
- Models often miss culturally-specific cues like euphemisms or honorifics that signal narrative shifts.
- Prompt engineering significantly boosts performance, particularly when anchoring temporal and narrative structure in input context.
Use Case
This research has applications in:
- Media monitoring and disinformation detection in Persian/Kurdish spaces
- Timeline generation for human rights documentation and event-based archiving
- Multilingual Q&A and summarization tools for humanitarian or journalistic work
- Curriculum design for teaching temporal reasoning and narrative analysis in NLP
It also supports the creation of benchmark datasets and evaluation frameworks for low-resource narrative understanding tasks.
Team
Karine Megerdoomian, E. Garcia
Latest publication or presentation
(if available)