File size: 2,995 Bytes
d84c451 4b9b015 82364fd 4b9b015 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 |
---
license: mit
---
# LitVISTA: A Benchmark for Narrative Orchestration in Literary Text
**[Paper (arXiv:2601.06445)](https://arxiv.org/abs/2601.06445)**
## 📖 Overview
**LitVISTA** is a structurally annotated benchmark designed to evaluate the **narrative orchestration** capabilities of Large Language Models (LLMs). Unlike traditional benchmarks that focus on causal coherence or local semantics, LitVISTA maps stories into a high-dimensional representational framework called **VISTA Space**.
It focuses on how narrative events (Verbs+) serve distinct structural roles—**Impulses**, **Resonances**, and **Pauses**—to create the pacing, tension, and rhythm inherent in human-authored literature.
## 🗂️ Dataset Structure
The dataset is organized into `train`, `val`, and `test` splits (8:1:1 ratio). Each entry consists of two files:
* `.txt` file: The raw literary text (sourced from LitBank/Project Gutenberg).
* `.ann` file: The structural annotations in a tab-separated format.
### Annotation Format
Each line in the `.ann` file follows this structure:
`Index` | `Anchor Topology` | `Offset` | `Verb+` | `Narrative Dependency`
:--- | :--- | :--- | :--- | :---
**0** | **Resonance** | 376,381 | Smoke | 49
1. **Index**: The unique identifier for the narrative anchor.
2. **Anchor Topology**: The functional role within VISTA Space:
* **Impulse (V_I)**: Plot drivers that form the narrative backbone and advance the plot to a new stage (τ → τ+1).
* **Resonance (V_R)**: Descriptive expansions that provide texture and context without advancing the plot stage (τ → τ+δ).
* **Pause (V_P)**: Vertical intensity markers that freeze narrative time to dive into psychology or micro-details (τ → τ).
3. **Offset**: Character-level start and end positions of the anchor in the source text.
4. **Verb+**: The textual anchor (canonical verbs or event-denoting nominals like "marriage").
5. **Narrative Dependency**: The ID of the parent node, defining the global structural backbone or recursive attachments.
## 📊 Key Statistics
| Metric | Train | Val | Test |
| --- | --- | --- | --- |
| Avg. Length (Tokens) | 10.2k | 9.9k | 10.7k |
| Predominant Role | Resonance | Resonance | Resonance |
| Complexity | High cross-dependency and long-range narrative relations. | | |
## 🚀 Task: Narrative Structure Reconstruction
The core task is to reconstruct the narrative topology. In the **Oracle Setting**, models are provided with the text and the `Verb+` anchors and must predict the `Anchor Topology` and `Narrative Dependency` (Head) for each.
## 📝 Citation
If you use this dataset, please cite our work:
```bibtex
@article{lu2026litvista,
title={LitVISTA: A Benchmark for Narrative Orchestration in Literary Text},
author={Lu, Mingzhe and Wang, Yiwen and Liu, Yanbing and You, Qi and Liu, Chong and others},
journal={arXiv preprint arXiv:2601.06445},
year={2026}
}
``` |