CITD@UIT at SemEval-2026 Task 4: Structured Reasoning and Metric Specialization for Narrative Similarity

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

Summary

CITD@UIT presented a dual-track approach for SemEval-2026 Task 4 on narrative similarity, involving triple-wise classification (Track A) and narrative representation (Track B) through failure-driven data enrichment. For Track A, they explored reasoning strategies including hybrid Cross-Encoder–LLM arbitration and DSPy-based decomposition, with a multi-stage pairwise reasoning pipeline using enforced moral agency hierarchies. Their final Gemini 2.5 Pro/Flash system achieved 77.39% on development and 69.25% on test data, ranking 17th among 46 teams. For Track B, they proposed BGE-M3 (LoRA), an instruction-guided dense representation model trained with Multiple Negatives Ranking Loss (MNRL), specialized using adversarial samples from Track A failures. This achieved 68.75% in official evaluation, ranking 6th among 26 teams. Analysis indicated narrative similarity relies more on outcome alignment and moral trajectory than lexical overlap.

Key takeaway

For NLP Engineers developing systems for complex narrative similarity tasks, you should prioritize explicit reasoning and specialized metric spaces. Consider integrating moral agency hierarchies into your reasoning pipelines and employing failure-driven data enrichment to specialize embedding models. This approach, demonstrated by CITD@UIT, can significantly improve performance by focusing on outcome alignment and moral trajectory rather than just lexical overlap.

Key insights

Narrative similarity benefits from explicit reasoning and specialized metric spaces, prioritizing moral trajectory over lexical overlap.

Principles

• Narrative similarity depends on outcome alignment and moral trajectory.
• Failure-driven data enrichment specializes embedding spaces effectively.
• Explicit reasoning and task-specific metric specialization are complementary.

Method

A dual-track approach combining a multi-stage pairwise reasoning pipeline with moral agency hierarchies (Track A) and an instruction-guided BGE-M3 (LoRA) model specialized with adversarial samples (Track B).

In practice

• Incorporate moral agency hierarchies into narrative reasoning pipelines.
• Synthesize adversarial samples from failure cases to specialize embedding models.
• Utilize instruction-guided dense representation models like BGE-M3 (LoRA).

Topics

• SemEval-2026 Task 4
• Narrative Similarity
• Large Language Models
• BGE-M3
• Data Enrichment
• Moral Agency

Best for: 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.
• 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.
• Deep Dive into Content Faithfulness: A new metric for ensuring text accuracy — Content faithfulness is paramount for AI agents, as even small errors compromise downstream decisions.
• Characterizing Narrative Content in Web-scale LLM Pretraining Data — Narrative structure in web-scale LLM pretraining data is measurable and unevenly distributed, impacting model reasoning.
• Impact of enriched meaning representations for language generation in dialogue tasks: A comprehensive exploration of the relevance of tasks, corpora and metrics — Enriched Meaning Representations improve NLG quality, especially for complex tasks and small, variable datasets.
• OPERA: Aligning Open-Ended Reasoning via Objective Perplexity-based Reinforcement Learning — OPERA aligns LLMs for open-ended tasks using intrinsic perplexity-based rewards, overcoming external judge biases.

Open in AIssential →

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