Mendel292 at SemEval-2026 Task 4: Disentangled Narrative Embeddings for Story Similarity

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

Summary

Mendel292 is a system developed for SemEval-2026 Task 4, focusing on Narrative Story Similarity. It features a narrative encoder that disentangles story representations into explicit subspaces for abstract theme, course of action, and outcome. This architecture builds upon a pre-trained sentence embedding model and incorporates a trainable BiLSTM projection layer, optimized with a triplet margin loss. The training set is augmented using backtranslation and includes weakly supervised multi-task objectives derived from unsupervised narrative clustering. Designed for few-shot settings with limited data, Mendel292 was, however, outperformed by an unsupervised pooling approach on the classification task and did not achieve top leaderboard scores. Despite this, the system facilitates systematic study into subspace factorization, weak labels, and data augmentation for narrative similarity modeling.

Key takeaway

For NLP Engineers developing narrative similarity models with limited data, consider exploring disentangled narrative embeddings. While Mendel292's specific implementation was outperformed by simpler pooling, its approach of separating theme, action, and outcome subspaces, combined with backtranslation and weak supervision, offers valuable insights. You should systematically evaluate subspace factorization and data augmentation strategies, even if initial leaderboard performance isn't top-tier, to understand their impact on narrative representation.

Key insights

A narrative encoder can disentangle story representations into explicit subspaces for theme, action, and outcome.

Principles

• Subspace factorization enhances narrative modeling.
• Weak labels and backtranslation augment training data.
• Few-shot design addresses limited training data.

Method

Mendel292 uses a pre-trained sentence embedding model with a BiLSTM projection layer and triplet margin loss, augmented by backtranslation and multi-task objectives from unsupervised clustering.

In practice

• Decompose story embeddings into theme, action, outcome.
• Augment narrative datasets via backtranslation.
• Use unsupervised clustering for weak supervision.

Topics

• Narrative Story Similarity
• Disentangled Embeddings
• SemEval-2026 Task 4
• BiLSTM
• Triplet Margin Loss
• Data Augmentation
• Few-shot 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.
• Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models — CEDAR disentangles vision-language model embeddings post-hoc, revealing compositional structure without increasing dimensionality, improving interpretability.
• 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.
• FLiP: Towards understanding and interpreting multimodal multilingual sentence embeddings — FLiP models effectively interpret sentence embeddings by linearly extracting lexical content, revealing underlying biases.
• The Cosine Similarity Trap: Why Embeddings Can’t Distinguish “War” from “Union” — Language models often cannot distinguish complementarity from conflict due to how embeddings encode relatedness, not relationship.
• Characterizing Narrative Content in Web-scale LLM Pretraining Data — Narrative content in LLM pretraining data is measurable and unevenly distributed, impacting model capabilities.

Open in AIssential →

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