Narrative Team at SemEval-2026 Task 4: Two-Stage Contrastive Learning for Narrative Similarity Assessment

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

The Narrative Team at SemEval-2026 Task 4 introduced a unified two-stage contrastive learning framework utilizing a RoBERTa-large encoder for narrative similarity assessment. The first stage involves contrastive pre-training on synthetic triplets to acquire general narrative similarity patterns. Subsequently, the second stage fine-tuned the model using a ranking-based objective specifically designed for Track A. This architecture supports both binary similarity classification for Track A and narrative embedding generation for Track B without requiring any architectural modifications. The system achieved an accuracy of 0.64 on Track A and 0.69 on Track B, demonstrating superior performance compared to single-stage baselines. This approach highlights that integrating synthetic contrastive supervision with task-specific ranking effectively produces stable and reusable narrative representations.

Key takeaway

For NLP Engineers developing narrative similarity systems, consider implementing a two-stage contrastive learning approach. Your models can achieve stable and reusable narrative representations by combining synthetic triplet pre-training with task-specific ranking objectives. This method supports both binary classification and embedding generation, potentially improving accuracy on tasks like SemEval-2026 Track A (0.64) and Track B (0.69) without architectural changes.

Key insights

Two-stage contrastive learning with synthetic data and ranking improves narrative similarity assessment.

Principles

• Synthetic data enhances general similarity learning.
• Task-specific ranking refines representations.
• Unified encoders support multiple tasks.

Method

A RoBERTa-large encoder undergoes Stage 1 contrastive pre-training on synthetic triplets, then Stage 2 fine-tuning with a ranking-based objective for specific task adaptation.

In practice

• Apply two-stage contrastive learning for text similarity.
• Use synthetic triplets for initial model pre-training.
• Fine-tune with ranking for specific classification tasks.

Topics

• Narrative Similarity
• Contrastive Learning
• RoBERTa-large
• SemEval-2026
• Text Embeddings
• Fine-tuning

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.
• SemEval-2026 Task 4: Narrative Story Similarity and Narrative Representation Learning — The NSNRL task operationalizes narrative similarity for evaluation and identifies areas for automated system improvement.
• Retell, Reward, Repeat: Reinforcement Learning for Narrative Theory-Informed Story Retelling — Reinforcement learning with narrative theory-informed AI feedback improves story diversity and human narrative alignment.
• A Self Consistency Based Reranking for Narrative Question Answering — Semantic agreement among multiple generated answers improves Narrative Question Answering robustness.
• The Case for Multi-Vector Contrastive Pre-Training, Eliminating Approximate Nearest Neighbor Search in Large-Scale Recommendation, and More! — Innovations in recommendation systems focus on efficiency, scalability, and quality through novel training, indexing, and architectural approaches.
• Solving the Embedding Step in NLP Projects — Effective embeddings are crucial for NLP model generalization and reliable downstream task performance.

Open in AIssential →

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