MarSan at SemEval-2026 Task 4: Narrative Similarity via Sentence-BERT Metric Learning with Triple-Derived Losses

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Expert, medium

Summary

The "MarSan" system, developed by Najafi, Tavan, and Colreavy for SemEval-2026 Task 4, addresses Narrative Story Similarity and Narrative Representation Learning (NSNRL). This system employs a unified representation-learning approach centered on a Sentence-BERT bi-encoder. The bi-encoder is trained using triple-derived metric learning objectives, which integrate in-batch contrastive learning with explicit triplet and margin-ranking constraints. For Track A, which requires determining the narratively closer candidate from a triple (anchor and two candidates), the system uses direct cosine comparison of embeddings. For Track B, focused on outputting story embeddings that maintain narrative similarity ordering via cosine distances, it provides normalized story vectors from the same encoder. MarSan achieved 65.00% accuracy on Track A and 65.50% on Track B, demonstrating that a single, well-aligned bi-encoder can perform competitively and efficiently across both narrative similarity tasks.

Key takeaway

For Machine Learning Engineers developing narrative understanding systems, you should consider a unified Sentence-BERT bi-encoder approach. This method, trained with triple-derived metric learning, offers competitive accuracy (65.00%-65.50%) for both comparative narrative similarity and embedding generation, while maintaining computational efficiency. Evaluate this architecture to streamline your model deployment for similar story analysis tasks.

Key insights

A single Sentence-BERT bi-encoder with triple-derived metric learning effectively solves narrative similarity and representation tasks.

Principles

• Metric learning improves narrative similarity.
• Bi-encoders offer competitive efficiency.

Method

A Sentence-BERT bi-encoder is trained with in-batch contrastive learning, explicit triplet, and margin-ranking constraints to generate story embeddings for cosine distance comparison.

In practice

• Apply Sentence-BERT for story embedding generation.
• Use triple-derived losses for narrative tasks.
• Evaluate bi-encoders for efficiency.

Topics

• Narrative Similarity
• Sentence-BERT
• Metric Learning
• Bi-encoder Models
• SemEval-2026 Task 4
• Representation Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• ttda704 at SemEval-2026 Task 4: Modeling Narrative Structures via Pseudonymization and Multi-View Sentence Alignment — Using contrastive learning with fine-tuned sentence transformers and multi-view alignment effectively models narrative structures and similarity.
• Rerankers Aren’t Magic Either: When the Cross-Encoder Layer Is Worth the Cost — Cross-encoder rerankers offer inconsistent performance gains over strong embeddings, often failing on complex query types.
• MiqraBERT: Regression-Based Sentence-BERT Finetuning for Biblical Hebrew Parallel Detection — Finetuning a BERT model with regression for semantic similarity effectively detects textual reuse in ancient languages beyond lexical overlap.
• Narrative Knowledge Weaver: Narrative-Centric Retrieval-Augmented Reasoning for Long-Form Text Understanding — Narrative Knowledge Weaver (NKW) enhances RAG by modeling evolving story worlds through structured narrative assets and multi-channel evidence assembly.
• Which sentence is doing the most work in your favourite novel ? I tried to find out. — Machine learning can identify semantically distinctive sentences in texts, revealing structural or narrative pivots.
• A Self Consistency Based Reranking for Narrative Question Answering — Semantic agreement among multiple generated answers improves Narrative Question Answering robustness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.