# ANEST narrative–affect dataset (ANAD v1): A large-scale derived feature resource for quantifying narrative–affective discrepancy

## TL;DR

This paper introduces a dataset that links how people tell stories about their relationships with their emotional expressions, using a new measure called the Narrative–Affect Discrepancy Index.

## Contribution

The paper introduces a novel dataset and a new index to quantify the discrepancy between narrative complexity and affective polarity in personal stories.

## Key findings

- The dataset includes 351,734 English narratives with structural, complexity, and affective annotations.
- The Narrative–Affect Discrepancy Index (NADI) is proposed as a computational measure of the gap between narrative structure and emotional expression.
- The dataset is publicly available through Zenodo with derived features only.

## Abstract

Narrative psychology, affective science, and computational social science all assume that how people tell stories about their lives is deeply intertwined with how they feel. Yet most empirical work either treats narrative structure and affective valence as separate constructs, or focuses on one dimension while ignoring the other. Here I present the ANEST Narrative–Affect Dataset (ANAD v1), a large-scale derived feature resource based on N = 351,734 human-written, English-language narratives drawn from public online discussions of romantic and relational life. Each observation is represented by three layers of derived annotation: (i) basic structural descriptors, (ii) a language-based index of narrative complexity (Level of Complexity; LoC), and (iii) a normalized affective polarity score derived from a rule-based sentiment model. From these components, I define the Narrative–Affect Discrepancy Index (NADI), which quantifies the gap between narrative complexity and expressed affect on a common 0–10 scale. NADI is offered as a computable, operational indicator rather than a validated psychological construct. The dataset is openly available via Zenodo (https://doi.org/10.5281/zenodo.18680687). No verbatim text is redistributed; the released files contain only derived, non-identifiable feature representations. This data article outlines the collection pipeline, scoring procedures, core distributions, and recommended use cases.