Neuro-Symbolic Narrative Similarity (HeteroGNN + Text Fusion)
I built a neuro-symbolic narrative-similarity system that combines a HeteroGNN over LLM-extracted story structure with Gemini text embeddings. I evaluated the interpretable additive fusion on pairwise and five-language cross-lingual data.
72.5% pairwise accuracy and 95% precision on 500 story-comparison test cases spanning five languages
I used GPT-OSS 20B to extract themes, actions, outcomes, and plot type into schema-validated JSON, then formed a heterogeneous graph with four node types. I combined the graph representation with Gemini text embeddings through a simple weighted average.
The system reached 72.5% accuracy on pairwise narrative similarity and 95% precision on 500 story-comparison test cases from the Tell-Me-Again dataset, spanning German, English, Spanish, French, and Italian. An angle analysis showed the graph and text signals carry largely independent information, so each one's contribution to the final similarity score can be measured exactly. I trained the GNN on 1,900 synthetic triplets, so generalization to substantially different narrative domains remains untested.