AI-Enabled Narrative Analytics for Persian and Kurdish

Presented at NACIL 4, University of Toronto Mississauga. May 2025

Narratives are one of humanity’s oldest technologies. They shape how we remember, imagine, warn, persuade, teach, and make sense of experience. Whether in epic stories, folktales, news reports, testimonies, or social media threads, narratives organize events into meaningful sequences.

As Labov and Waletzky famously describe it, a narrative is a recounting of things that happened, involving a sequence of events meaningfully connected through temporal and often causal relations, typically structured with a beginning, middle, and an end (Labov and Waletzky 1967, Labov 2011, Onega and Landa 2014). Toolan (2001) describes a narrative as "recounting of things spatiotemporarily distant".

Our research investigates whether large language models (LLMs) can identify, interpret, and reason about these narrative structures in underrepresented languages — specifically Persian and Sorani Kurdish.

Why Narrative Analytics Matters

Within the field of linguistics, Narrative Analysis has taken two distinct but complementary approaches:

Structural (componential) analysis

What elements make up a narrative? How are events introduced, linked, and resolved?

Functional analysis

What purpose does the narrative serve? How do speakers use narrative structure to accomplish social goals?

In computational terms, narrative analytics refers to generating a structured representation of events, including:

• Event descriptions (what happened)

• Participants (who was involved)

• Locations (where it happened)

• Temporal information (when it occurred)

• Causality (why it happened or what it led to)

• Manner and modality

Traditional approaches rely on a full NLP pipeline: tokenization, POS tagging, parsing, event extraction, temporal reasoning, and coreference resolution. But for low-resource languages, many of these components simply do not exist — and building them is time-consuming and expensive. LLMs, in contrast, have been trained on multilingual datasets and offer an opportunity to approach narrative structure holistically, without requiring each pipeline component to be engineered individually.

Our Approach: LLMs for Narrative Understanding

We evaluated several LLMs (GPT, Deepseek, LLaMA) against manually curated narrative texts in Persian and Sorani Kurdish using:

• Zero-shot prompting

• Definition-driven in-context learning

• Prompt chaining for multi-step narratives

• Schema-inspired output formats

• Causal and temporal reasoning prompts

For example, models were asked to:

• identify events and participants

• order events temporally

• infer implicit relations

• detect causality

• resolve coreference across sentences

• assemble full narrative chains

Prompt chaining was especially effective: one prompt identifies events, another links them, another infers temporal relations, and so on.

Sample Analysis: Persian and Kurdish

Here is a short Sorani Kurdish narrative and the English gloss used during evaluation:

لە شەوی ١٥ی ئەیلوولی ٢٠٢٣دا، تەقینەوەیەک لە نێو بازاڕی سلێمانی ڕوویدا.ڕەزام حەمە، کەسێکی فرۆشیار، لە کاتی تەقینەوەکە بریندار بوو.هاوڵاتیان بەخێرایی کەسی بریندارەکان گەیاندن بۆ نەخۆشخانە.ڕۆژی دوای ئەو ڕووداوە، حکومەت ڕاگەیاند کە تاقیکردنەوە دەستپێدەکات.تاکو ئێستا، هێشتا کەسێک بەتاوانبار نەناسراوە.

On the night of September 15, 2023, an explosion occurred in the Sulaimani bazaar. Rezam Hama, a vendor, was injured during the explosion. Civilians quickly transported the injured to the hospital. The day after the incident, the government announced that an investigation would begin. As of now, no one has been identified as responsible.

LLMs were able to:

• identify the explosion as the initial event

• track the injury and community response

• infer the implicit causal chain (explosion → injury → hospital → investigation)

• recognize the narrative tension created by “no one identified”

Evaluation Findings

We evaluated the results against a manually annotated test set of about 20 documents per language (~2500 words) containing the expected output. Overall, the LLMs performed well out-of-the-box. We were able to improve the results even further by tuning the systems using few-shot and chain-of-thought prompting.

✅ Strengths

• Event identification was reliable for Persian, including nested events detection (e.g., relative clauses, nominalization)

• Implicit reasoning was surprisingly strong: models inferred agents of the action, as well as causes or consequences even when unstated.

• Discontinuous constructions were often correctly interpreted.

• Cross-document inference (connecting earlier and later events) worked better than expected.

  
  

⚠️ Challenges

• Inconsistency in Sorani Kurdish, likely due to limited training data.

• Coreference degraded across longer narratives.

• Temporal relation classification (BEFORE/AFTER/CAUSES) was variable.

• Although LLMs were able to interpret conditional events correctly, they had difficulty with counterfactuals.

• Cultural knowledge gaps led to misinterpretation of euphemisms, honorifics, or narrative softening.

• Hallucinated details based on historical knowledge, where LLMs leverage historical knowledge about an event to fill in elements of the narrative that are not explicitly stated in the text (e.g., reference to reform in the 1960s in a Persian document was linked directly to the Pahlavi government's reforms, even though no specific country was mentioned).

• Lack of schema alignment (e.g., TimeML) makes systematic evaluation difficult.

Compared to traditional TimeML-style extraction work in Persian (Yaghoobzadeh et al. 2012; Eshaghzadeh et al. 2013), LLMs offer:

• greater flexibility

• better reasoning about implied events

• multilingual adaptability

• more robust in cross-sentence and cross-document analysis

But they lack the precision and standardization of classical NLP pipelines.

Key Insights

• LLMs offer holistic narrative analysis, capturing causality, temporality, and implicit elements.

• This is especially valuable for low-resource languages like Kurdish, where pipeline components are missing.

• Traditional NLP provides precision and schema alignment, but lacks multilingual and inferential flexibility.

• A hybrid approach, combining LLM reasoning with structured extraction frameworks, may be ideal.

• Significant room remains for improvement, especially for Kurdish.

• Building annotated datasets for narrative tasks is essential for benchmarking and future research.

Use Cases

Narrative analytics for Persian and Kurdish has real-world impact:

✅ Media monitoring & disinformation research

Timeline reconstruction across news and social media.

✅ Human rights documentation

Extracting sequences of events from testimonies or reports.

✅ Educational tools

Teaching temporal structure, narrative sequencing, and comprehension.

✅ Cultural research

Studying narrative framing in Persian and Kurdish traditions.

✅ Dataset & benchmark creation

Informing the next generation of narrative analysis tasks in low-resource languages.

Next Steps

Based on this research, we plan to:

• Develop gold-standard narrative datasets for Sorani Kurdish and expand the annotated data for Persian.

• Explore schema-constrained prompting to align outputs with TimeML.

• Pilot a modular hybrid pipeline combining LLM inference with rule-based components.

• Extend narrative analytics to include motivation, discourse markers, and evidentiality.

• Apply this work to media ecosystems, historical corpora, and cultural narrative studies.

Conclusion

Narratives reveal how communities understand their world. For languages with rich oral and written traditions, such as Persian and Kurdish, narrative analytics opens new paths for linguistic research, cultural preservation, and multilingual AI.

LLMs are not perfect readers, but they offer a powerful starting point for analyzing narrative structure in underrepresented languages. With careful methodology, cultural insight, and hybrid modeling, we can build AI tools that support researchers, educators, journalists, and communities across the Iranian linguistic landscape.

Download Presentation