💬 NLP

ttda704 at SemEval-2026 Task 4: Modeling Narrative Structures via Pseudonymization and Multi-View Sentence Alignment

The paper presents the ttda704 team's approach to SemEval-2026 Task 4, which employs contrastive learning with fine-tuned sentence transformers to model narrative similarity through two distinct pipelines: a single-view method for full narrative encoding and a multi-view method for aligning specific thematic, plot, and outcome components.

Original authors: Tai Tran Tan, An Dinh Thien

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tai Tran Tan, An Dinh Thien

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer how to understand stories, not just by counting words, but by understanding the soul of the plot. That is exactly what the team behind "ttda704" did for a competition called SemEval-2026.

Here is a simple breakdown of their approach, using everyday analogies.

The Big Problem: The "Name Game" Trap

Usually, when computers compare two stories, they get tricked by names. If Story A is about a boy named Tim who fights a dragon, and Story B is about a girl named Sarah who fights a dragon, a simple computer might think they are totally different because "Tim" and "Sarah" are different words.

But a human knows: Wait, both stories are about a hero fighting a dragon. They are actually very similar.

The authors wanted the computer to ignore the names and focus on the structure of the story.

The Solution: Two Different Pipelines

The team built two different "machines" (pipelines) to solve two different parts of the puzzle.

1. Track A: The "Blindfolded" Detective (Single-View)

The Goal: You are given a "Anchor" story and two other stories. You have to pick which one is more similar to the Anchor.

The Trick: Pseudonymization (The "Blindfold")
Before the computer reads the story, the team plays a game of "Mad Libs." They take every specific name (like "Tim," "The Dragon," "London") and replace them with generic labels like Character_A, Monster_B, and Place_1.

Why? This forces the computer to stop looking at who is doing things and start looking at what is happening. It's like listening to a story where everyone is wearing a mask; you can only judge the plot by the actions, not the faces.

The Engine: Smart Freezing
They used a powerful pre-trained brain (a sentence transformer) to read these masked stories. To keep the brain from "forgetting" how to speak English while learning this new task, they used Smart Layer Freezing.

Analogy: Imagine a master chef (the pre-trained model) learning to cook a specific new dish. You tell them, "Don't change how you chop onions or boil water (the bottom layers), but you can invent new sauces for this specific dish (the top layers)." This stops them from overthinking and making mistakes.

The Result: This method worked well, helping the team rank 14th out of 42 teams.

2. Track B: The "Three-Legged Stool" (Multi-View)

The Goal: Instead of just picking a winner, the team had to turn every story into a unique "fingerprint" (a vector) so that similar stories end up close together in a digital space.

The Trick: Breaking the Story Apart
The team realized that a story isn't just one thing. It has three distinct parts, like the legs of a stool:

The Theme: The big idea (e.g., "Revenge" or "Coming of Age").
The Plot: The sequence of events (e.g., "Hero gets hurt, then fights back").
The Outcome: How it ends (e.g., "Hero wins" or "Hero learns a lesson").

The Process:

Extraction: They used a super-smart AI (LLM) to read the story and pull out these three specific parts, writing them down like a summary.
Alignment: They created a special system that learns to compare stories based on all three legs at once.
- Analogy: Imagine trying to match two people by their face, their voice, and their walk. If you only look at the face, you might get it wrong. But if you look at the face, voice, and walk together, you are much more accurate.
Self-Correction: They added a rule to make sure the "Theme," "Plot," and "Outcome" parts all agree with each other, creating a unified story fingerprint.

The Result: This multi-view approach helped them rank 6th out of 25 teams.

The Secret Sauce: Synthetic Data

The team didn't have enough real human-labeled stories to train their models perfectly. So, they used AI to invent thousands of fake story examples (synthetic data) to practice on.

Analogy: It's like a pilot training in a flight simulator. They don't need to fly a real plane in a storm to learn; they can practice on a computer simulation first. This allowed them to train their models efficiently without needing a massive library of human-graded stories.

Summary of What They Claimed

They didn't use expensive, slow AI to generate answers during the test. Instead, they built a fast, efficient system that learns patterns first and then applies them quickly.
They proved that removing names (pseudonymization) helps computers understand the structure of a story better than just reading the raw text.
They proved that breaking a story into Theme, Plot, and Outcome creates a much better "fingerprint" for comparing stories than treating the story as one big block of text.

In short, they taught a computer to ignore the "names" and the "noise" to focus entirely on the shape and flow of the story.

Technical Summary: ttda704 at SemEval-2026 Task 4

Problem Statement
The paper addresses SemEval-2026 Task 4, which challenges participants to model narrative similarity and representation learning. The core difficulty lies in distinguishing between surface-level lexical overlap and deep narrative structure. As noted by Hatzel and Biemann (2024a), effective story comparison requires capturing latent dimensions such as abstract themes, the progression of the plot (course of action), and final outcomes, rather than relying on simple keyword matching or event extraction. The authors highlight that while generative Large Language Models (LLMs) can reason through these ambiguities, their computational cost and latency make them impractical for scalable retrieval and embedding tasks. Consequently, the goal is to develop an efficient, embedding-based system that captures structural narrative nuances without relying on expensive inference-time generation.

Methodology
The authors propose a contrastive learning framework utilizing fine-tuned sentence transformers, implemented through two distinct pipelines corresponding to the task tracks:

Track A (Single-View Classification): This pipeline focuses on determining which of two candidate stories is narratively closer to an anchor story.
- Pseudonymization: To mitigate spurious correlations caused by specific entity names (e.g., character or location overlap), the authors apply a referential normalization step. Using fastcoref and spaCy, they map repeated mentions of the same entity to consistent generic placeholders (e.g., "Character_A") while preserving within-story coreference. This forces the model to attend to semantic progression rather than lexical identity.
- Architecture & Training: The system uses the all-mpnet-base-v2 model as a backbone. To prevent overfitting and preserve pre-trained linguistic knowledge, a "Smart Layer Freezing" strategy is employed: the embedding parameters and the bottom 40% of transformer layers are frozen, while the top layers are fine-tuned. The model is trained using triplet-margin loss on synthetic data.
Track B (Multi-View Representation Learning): This pipeline generates vector embeddings where cosine similarity reflects narrative affinity.
- Multi-View Decomposition: Instead of encoding the full text as a single vector, the authors decompose each story into three distinct views using LLMs (specifically o3-mini): Theme (central controlling idea), Plot (chronological sequence of state-changing events), and Outcome (resolution).
- Contrastive Learning: The model employs view-specific projection heads to map these components into orthogonal embedding spaces. A contrastive loss function ( $L_{con}$ ) is applied to each view to maximize the similarity between matching narrative facets and minimize it for non-matching ones.
- Self-Supervised Alignment: To ensure coherence across the different views, an alignment loss ( $L_{align}$ ) is introduced. This loss minimizes the distance between the fused embedding (a weighted sum of the four views: full text, theme, plot, outcome) and the average of the individual view embeddings, encouraging a unified joint space.
- Fusion: At inference, embeddings from the four components are fused using fixed weights (0.5 for full text, 0.1 for theme, 0.2 for plot, 0.2 for outcome) followed by L2 normalization.

Key Contributions
The paper outlines three primary contributions:

LLM-Augmented Narrative Representation: A deterministic framework that leverages LLMs strictly for offline data augmentation (synthetic extraction of narrative structures) and structural decomposition, while relying on efficient, fine-tuned sentence transformers for the final similarity inference.
Multi-View Structural Modeling: A contrastive learning approach that explicitly models the three core aspects of narrative similarity (theme, course of action, outcome) via view-specific projection heads and self-supervised alignment, decoupling complex signals into orthogonal spaces.
Robust Contrastive Training: The implementation of triplet-margin optimization combined with smart layer freezing to improve generalization while preserving the nuance of narrative structures.

Results

Track A: The system achieved a private score of 0.6925, ranking 14th out of 42 teams. The use of pseudonymized data yielded a consistent improvement over raw data (0.6650 to 0.6925) for the all-mpnet-base-v2 model. The authors attribute this to the removal of entity-specific noise, forcing the model to focus on semantic structure.
Track B: The system ranked 6th out of 25 participants with a private score of 0.6875. Ablation studies on the development set indicated that a weighted fusion strategy (emphasizing the full text at 0.5, with smaller weights for theme, plot, and outcome) outperformed equal weighting or single-view approaches. The combination of all-mpnet-base-v2 and the o3-mini LLM extractor produced the best results.

Significance and Limitations
The authors claim that their approach offers a practical alternative to LLM-based generative inference for large-scale narrative retrieval. By shifting the computational load to an offline data generation phase and utilizing efficient embedding models at inference, the system achieves competitive performance with significantly lower latency.

However, the paper modestly acknowledges several limitations:

Generalization: The models rely on synthetic training data and may not fully generalize to diverse narrative styles or domains without further adaptation.
Extraction Noise: The multi-view pipeline depends on automatically extracted narrative elements; without a dedicated manual quality audit of these extractions, residual errors in theme or plot coherence may remain unquantified.
Hyperparameter Tuning: The fixed fusion weights and temperature settings were tuned on the development set and may not be optimal for unseen domains.

The work concludes that while the approach avoids the prohibitive costs of generative reasoning, the reliance on synthetic data and automated extraction introduces constraints on the robustness of the learned representations.