📄 ChronoLearn: A GRAG LLM-Based System for Structuring and Exploring Historical Narratives

Summary

ChronoLearn targets a gap in history learning. Material on Arabic historical narratives, especially Palestinian and Jordanian accounts of the Nakba, is scattered, heavy on text and hard to explore. Chat-style LLM tools add their own problems: they cite no sources and can hallucinate.

The authors build a pipeline in three parts:

• Extraction: an ETL-style NLP stage pulls entities and relations from user-supplied Arabic texts.
• Graph building: a schema guides how those extractions become a knowledge graph.
• Generation: graph retrieval is coupled with LLM generation (GRAG) to produce narratives grounded in the graph.

The test material is a small, deliberately mixed set of three sources: refugee testimony, an encyclopedic article on a folk song, and a scholarly essay on memory. On it they compare two generation strategies: topic-guided generation from full text, and generation grounded in KG triples.

Neither strategy wins outright. Topic conditioning yields more fluent text that overlaps more with the references. Triple grounding gives better factual precision and covers more entities.

An interactive platform with graph views and storylines accompanies the work. Its value lies in showing structured, explainable access to low-resource Arabic historical content.

Key points

• Data: three heterogeneous sources, chosen to span event-based, cultural and memory-based narrative types.
  • Zayed (2013) on inter-generational refugee narratives.
  • The Arabic Wikipedia article on the song Jafra.
  • Abu-Lughod (2007) on memory and postmemory.
• Method A (topic-guided) leads on surface and semantic similarity:
  • BERTScore 0.703 vs 0.691.
  • ROUGE-L 0.097 vs 0.033.
• Method B (KG triple grounding) leads on entity coverage:
  • Recall@5 0.174 vs 0.130.
• HITS@5 is 1.00 for both methods. The authors still note that low Recall@5 means many relevant entities are missed.
• Event linking is reported as linked_named_events: 0. The component is unfinished and is not used during generation or evaluation. Future work covers temporal reasoning, named-event linking and human-in-the-loop validation.

Interesting ideas

• I1. The paper runs two generation arms on the same input: one conditioned on the whole text, one constrained by structured triples. It reports the trade-off between fluency and factual grounding rather than declaring one winner, which makes the cost of grounding visible.
• I2. Generated text is scored with ranking metrics over entities (HITS@k, Recall@k) as well as overlap metrics like ROUGE-L and BERTScore. This shows that a saturated hit rate can hide poor coverage.
• I3. An unfinished pipeline stage is exposed as an explicit zero counter rather than silently skipped, so readers know which constraints were actually active.
• I4. The test set is stratified by narrative type: factual events, symbolic and cultural content, and memory-driven testimony. Extraction difficulty differs sharply across these types, so stratified results are more informative than a pooled score.
• I5. A fixed schema guides the conversion of extracted entities and relations into a graph. This keeps the output canonical and queryable instead of a loose pile of triples.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): when generating from the meaning graph, compare graph-only realisation against a text-conditioned baseline. Track fluency and fact fidelity as separate axes.
• I2 (learn): for the text → graph → text compressor, measure what share of source entities and roles survive the round trip, not just whether any one survives.
• I3 (learn): every evaluation report could list each explainable stage with an activity count, so a dead component shows up instead of quietly inflating or deflating scores.
• I4 (answer): split QA evaluation into factual school texts versus fables and tales with implicit or figurative content, and report each separately.
• I5 (world): map frame outputs onto a fixed schema of entities, states and events before they enter the world model, so downstream readers query one consistent structure.