Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CascadeMind at SemEval-2026 Task 4: A Hybrid Neuro-Symbolic Cascade for Narrative Similarity

Sebastien Kawada, Dylan Holyoak

TL;DR

The paper tackles narrative similarity assessment for SemEval-2026 Task 4 by proposing CascadeMind, a hybrid neuro-symbolic cascade that combines neural self-consistency voting with a symbolic tiebreaker. It introduces a Multi-Scale Narrative Analysis Ensemble that fuses five signals spanning lexical, semantic, structural, emotional, and event-based dimensions, with weights learned via differential evolution. The approach achieves strong performance on development data (81% accuracy) and provides insights into uncertainty estimation and the complementary roles of surface-level and structural signals. This work advances practical narrative comparison by demonstrating how selective deferral to symbolic reasoning can boost neural predictions in genuinely ambiguous cases and identifies avenues for future improvements in domain adaptation and richer event representations.

Abstract

We present a hybrid neuro-symbolic system for the SemEval-2026 Task 4 on Narrative Story Similarity. Our approach combines neural self-consistency voting with a novel Multi-Scale Narrative Analysis Ensemble that operates as a symbolic tiebreaker. The neural network component uses a large language model with multiple parallel votes, applying a supermajority threshold for confident decisions and escalating uncertain cases to additional voting rounds. When votes result in a perfect tie, a symbolic ensemble combining five narrative similarity signals (lexical overlap, semantic embeddings, story grammar structure, event chain alignment, and narrative tension curves) provides the final decision. Our cascade architecture achieves 81% accuracy on the development set, demonstrating that selective deferral to symbolic methods can enhance neural predictions on genuinely ambiguous narrative comparisons.

CascadeMind at SemEval-2026 Task 4: A Hybrid Neuro-Symbolic Cascade for Narrative Similarity

TL;DR

The paper tackles narrative similarity assessment for SemEval-2026 Task 4 by proposing CascadeMind, a hybrid neuro-symbolic cascade that combines neural self-consistency voting with a symbolic tiebreaker. It introduces a Multi-Scale Narrative Analysis Ensemble that fuses five signals spanning lexical, semantic, structural, emotional, and event-based dimensions, with weights learned via differential evolution. The approach achieves strong performance on development data (81% accuracy) and provides insights into uncertainty estimation and the complementary roles of surface-level and structural signals. This work advances practical narrative comparison by demonstrating how selective deferral to symbolic reasoning can boost neural predictions in genuinely ambiguous cases and identifies avenues for future improvements in domain adaptation and richer event representations.

Abstract

We present a hybrid neuro-symbolic system for the SemEval-2026 Task 4 on Narrative Story Similarity. Our approach combines neural self-consistency voting with a novel Multi-Scale Narrative Analysis Ensemble that operates as a symbolic tiebreaker. The neural network component uses a large language model with multiple parallel votes, applying a supermajority threshold for confident decisions and escalating uncertain cases to additional voting rounds. When votes result in a perfect tie, a symbolic ensemble combining five narrative similarity signals (lexical overlap, semantic embeddings, story grammar structure, event chain alignment, and narrative tension curves) provides the final decision. Our cascade architecture achieves 81% accuracy on the development set, demonstrating that selective deferral to symbolic methods can enhance neural predictions on genuinely ambiguous narrative comparisons.
Paper Structure (29 sections, 7 equations, 2 figures, 2 tables)

This paper contains 29 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Self-Consistency Decoding
  4. Narrative Representation Learning
  5. Ensemble Methods and Selective Prediction
  6. Narrative Theory
  7. System Architecture
  8. Neural Self-Consistency Voting
  9. Decision Logic
  10. Multi-Scale Narrative Analysis Ensemble
  11. Signal 1: Lexical Similarity (TF-IDF)
  12. Signal 2: Semantic Similarity
  13. Signal 3: Story Grammar Similarity
  14. Signal 4: Narrative Tension Curve
  15. Signal 5: Event Chain Similarity
  16. ...and 14 more sections

Figures (2)

  • Figure 1: Cascade decision process. Most cases (74%) resolve via supermajority; 21% require escalation; 5% invoke the symbolic ensemble.
  • Figure 2: Multi-Scale Narrative Ensemble with five weighted similarity signals ($\mathbf{w}^*$ optimized via differential evolution).